(Gordon Belot) Geometric Possibility

Geometric Possibility
This page intentionally left blank

Geometric Possibility
Gordon Belot
1
3
Great Clarendon Street, Oxford ox2 6dp
Oxford University Press is a department of the University of Oxford.
It furthers the University's objective of excellence in research, scholarship,
and education by publishing worldwide in
Oxford New York
Auckland Cape Town Dar es Salaam Hong Kong Karachi
Kuala Lumpur Madrid Melbourne Mexico City Nairobi
New Delhi Shanghai Taipei Toronto
With offices in
Argentina Austria Brazil Chile Czech Republic France Greece
Guatemala Hungary Italy Japan Poland Portugal Singapore
South Korea Switzerland Thailand Turkey Ukraine Vietnam
Oxford is a registered trade mark of Oxford University Press
in the UK and in certain other countries
Published in the United States
by Oxford University Press Inc., New York
q Gordon Belot 2011
The moral rights of the author have been asserted
Database right Oxford University Press (maker)
First published 2011
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
without the prior permission in writing of Oxford University Press,
or as expressly permitted by law, or under terms agreed with the appropriate
reprographics rights organization. Enquiries concerning reproduction
outside the scope of the above should be sent to the Rights Department,
Oxford University Press, at the address above
You must not circulate this book in any other binding or cover
and you must impose the same condition on any acquirer
British Library Cataloguing in Publication Data
Data available
Library of Congress Cataloging in Publication Data
Data available
Typeset by SPI Publisher Services, Pondicherry, India
Printed in Great Britain
on acid-free paper by
MPG Books Group, Bodmin and King’s Lynn
ISBN: 978–0–19–959532–7
1 3 5 7 9 10 8 6 4 2
To my parents, with love
Preface
This book is about one aspect of the substantival-relational debate con-

cerning the nature of space: the role that the notion of geometric possi-
bility plays in allowing relationalists to make sense of claims about the
structure of space. Appeal is often made to some such modal notion in
the literature on relationalism—but very seldom is very much said about
the nature of the modality involved. My aim here is to examine the
accounts of nomic possibility provided by several philosophical accounts
of laws of nature to see whether they can be adapted to the geometric
context. I hope that there will be something in it not only for those
interested in relationalism, but also for those interested in laws of nature,
in varieties of modality, or in ontology more generally.
A large part of the pleasure that I took in writing this strange little book
derived from the geometry that I learned while working on it. Some of the
weirder things that I learned have found their way into the finished
product and I hope that readers will also take some pleasure in them. At
times, no doubt, I allowed myself to get carried away and crammed in
more details, or examples, or stuff than I ought to have. Tip: some readers
will want to skip many of the Remarks set off from the text; many readers
will want to skip some of them.
Appendix C began life about eight years ago as part of another (far too
ambitious) project. Work on it was supported by the National Science
Foundation under Grant No. SES-0I35445 and by research leave from
NYU. The rest of the book was written mostly in 2008, in part during a
research leave from the University of Michigan.
Let me begin at the end. I would like to thank everyone at OUP for
their help: Rowena Anketell, Daniel Bourner, Eleanor Collins, Elmandi
du Toit, Francis Eaves, Laura O’Brien, Sarah Parker, and, especially, the
inimitable Peter Momtchiloff.
Thanks also to all those who suffered through presentations of this
material in Ann Arbor, Boulder, Buffalo, Montreal, and Pittsburgh. For
helpful comments and discussion, I owe special thanks to: Chloe Armstrong,
Dave Baker, Cian Dorr, Hilary Greaves, Nick Huggett, Lina Jansson, Peter
Petersen, Dan Peterson, Oliver Pooley, Adam Rigoni, Michael Tooley,
viii preface
Mark Wilson, and Bas van Fraassen. Very special thanks to the following,
who voluntarily read great chunks of the manuscript and provided many
invaluable suggestions: Anonymous, Caro Brighouse, Jeremy Butterfield,
Craig Callender, Adam Caulton, Laura Ruetsche, Brad Skow, and Ed
Slowik. Most special thanks of all for my fellow inhabitants of Le Vieux
Shack: Grisbi, Stargell, and Honus for their forceful and enigmatic leadership;
and Laura for many, many things.
Contents
Introduction 1
I Possible Structures of Space 8

1. Introduction 8
2. Distance 10
3. Candidates for Spatial Structure 14
4. Plenitude of Possibilities 30
5. Distance Relations? 31
II Spatial Structure for Relationalists 35

1. Substantivalism and the Structure of Space 35
2. Conservative Relationalism 37
3. Modal Relationalism 49
III Best-System Approaches 54

1. Introduction 54
2. Lewis on Physical Possibility 55
3. Huggett on Geometric Possibility 57
4. Worries About Best-System Approaches 60
5. Summation 76
IV Primitivist Approaches 78
1. Introduction 78
2. A First Primitivist Approach 83
3. Geometric Facts for Substantivalists 86
4. Another Primitivist Strategy 90
5. Beyond Metricity 95
6. Summation 100
x contents
V Necessitarian Approaches 102

1. Introduction 102
2. Necessitarianism about Laws 103
3. Compatibility Properties 114
4. Necessitarianism about Geometry 117
5. Super-Grounding and Necessitarianism 121
6. Summation 132
Conclusion 134
Appendix A. Simplicity and Ontology 139
Appendix B. Limits of Sequences of Metric Spaces 150
Appendix C. Some Background to the Absolute-Relational Debate 157
Appendix D. Leibniz and Modal Relationalism 173
Appendix E. More on Congruence and Superposability 186
References 199
Index 213
Introduction
[S]pace is something; but like time . . . [it] is a general order of things.

Space is the order of co-existents and time is the order of successive
existents. They are true things, but ideal, like numbers.
Leibniz
It is helpful to distinguish between two sorts of philosophical debates

concerning space and geometry: on the one hand we have debates
between realists and others; on the other hand debates between substan-
tivalists and relationalists.
Realists about space attribute to reality a determinate spatial structure,
anti-realists refuse to do so. Setting aside views motivated by modern
physics, we can divide anti-realist views concerning space into two
types. There are sceptical positions, according to which one should with-
hold judgement on the realist thesis.1 And there are views (often, broadly
speaking, of an idealist flavour) under which reality is fundamentally non-
spatial. Advocates of the latter sort of anti-realism typically concede that
there is also a sense in which it is non-misleading to speak of material
objects and their spatial distribution. So such authors in effect hold that
realism is false strictly and fundamentally speaking but allow that it is true
(enough) in some other sense.
Our concern will be with the opposition between substantivalism and
relationalism. This is an in-house dispute among realists (or those who
allow that realism is as good as true in some pertinent context). Substantiv-
alists maintain that space consists of parts and that the geometric relations
1
Bayle describes a sceptical argument against the existence of space in Remark G of the
article on Leucippus in the Historical and Critical Dictionary. The arguments against material
extension in the first part of Remark G of the article on Zeno of Elea would appear to apply
equally against any sort of extension.
2 introduction
between bodies are derivative on the relations between the parts of space
that they occupy. Relationalists are realists about spatial structure who deny
this. Among archetypical relationalists, we might number Aristotle,
Descartes, and Leibniz (although there is some question in some people’s
minds about Descartes).
The above characterization of relationalism is a little involved. Why not
simply take the characteristic difference between substantivalists and rela-
tionalists to be that the former but not the latter include space in their
ontology (or in their fundamental ontology)? Because that would threaten
to obscure the distinction between relationalists and anti-realists about
space. Realists about space can be either relationalists or substantivalists.2
And anti-realists who recognize a sense in which reality can be said to be
spatially organized can adopt either a relationalist or a substantivalist stance
about that sense.3 Consider the case of Leibniz, who is usually treated as
first among equals among relationalists. Leibniz was of course an anti-
realist as well a relationalist. Speaking as an anti-realist, Leibniz denied that,
fundamentally speaking, anything bears a spatial relation to anything
else—and in this mode he would presumably deny sentences asserting
that space is three-dimensional or that it is infinite. But Leibniz was also
willing to speak about the phenomenal world. And in that mode he spoke
as a realist and a relationalist—and was happy to speak of space as a sort of
thing (see the epigraph to this chapter) and would have been willing to
affirm that space is three-dimensional and infinite (see Appendix D
below). So it seems unnatural to say that Leibniz qua relationalist excluded
space from his ontology—he spoke of it as a thing and was willing to
attribute features to it. And while Leibniz did exclude space from his
fundamental ontology, this was because he was an anti-realist, not because
he was a relationalist.
Relationalists and substantivalists can agree (as Leibniz and Newton
did) that space is a thing of some sort and that it has some given geometric
structure. Their disagreement concerns the nature of the existence
of space. This disagreement is reflected in a disagreement over the
truth conditions appropriate for claims about the geometric structure of
space.
2
Aristotle was a realist and a relationalist, Newton a realist and a substantivalist.
3
See e.g. the discussion on pp. 513 f. of Foster, “In Defence of Phenomenalistic Idealism.”
introduction 3
Substantivalists take space to consist of parts standing in geometric

relations to one another. So on their view the truth conditions for a
claim about the structure of space will be stateable in terms of the
geometric relations obtaining between parts of space. For example, they
will hold that space is finite if and only if there is some number N such that
no two parts of space are separated by more than N units of distance.
A very similar approach is available for some relationalists. Let us speak
of the pattern of geometric relations instantiated by the parts of matter at a
world at a time as the material geometry of that world. Suppose that one
holds that matter necessarily forms a plenum—perhaps because one holds
that there is some sort of conceptual incoherence in the notion of void
space.4 Then one can identify the geometry of space with material geom-
etry, and give truth conditions for claims about spatial structure that differ
from those of substantivalists only in quantifying over material points
rather than over points of space: ‘space is finite’ is true if and only if
there is an upper bound on the distances between material points, and
so on.
But for contemporary relationalists the identification of spatial geometry
with material geometry is a non-starter—most of us believe that there are
possible worlds at which space is Euclidean but at which matter consists of
point-particles or of an expanding sphere.5 How can relationalists who
countenance such possibilities provide analogues for the substantivalist
truth conditions for claims about the structure of space?
Leibniz was one historical relationalist who faced this difficulty. He took
matter to form a three-dimensional Euclidean plenum at our world. But
he allowed that void space was possible and took space to have the same
three-dimensional Euclidean structure at every world. So he could not
simply adapt the substantivalist truth conditions for claims about the
structure of space by quantifying over material points rather than points
of space—this would have forced him to say that space was finite at some
worlds, infinite at others, rather than that space had the same Euclidean
structure at all worlds. Leibniz sometimes seems to suggest that the
solution to this problem lies in appeal to a distinctive geometric notion
4
See e.g. bk. iv of Aristotle’s Physics or §§ii.16 ff. of Descartes’s Principles of Philosophy.
5
The exceptions being those philosophers who hold that the only metaphysically possible
worlds are the physically possible ones while also denying that worlds of the sort described are
physically possible. They are invited to substitute examples that they like better, if they take
substantivalism and relationalism to be worth discussing at all.
4 introduction
of possibility: we should retool the substantivalist truth conditions so

that they demand not that certain patterns of geometric relations be
(un)instantiated by actual material points, but rather that the instantiation
of such patterns be geometrically (im)possible. Thus, space is finite if there
is some number N such that it is impossible for material points to be located
more than N units away from one another, and so on.
The historical claims of the preceding paragraph are contentious.6 But
the invocation of a distinctive brand of geometric modality is one way that
relationalists might respond to the challenge of providing truth conditions
for claims about spatial structure. And it has seemed to many to be the
most promising available avenue. My project here will be to investigate
accounts we might give of this brand of modality.
Overview. Chapter I below addresses a preliminary question—What

structures should we take to represent possible spatial geometries? It is
essentially expository in character, aiming primarily to put in place a
framework of concepts and examples for the chapters to follow. It is
argued that for present purposes it is safe to focus on metric spaces, math-
ematical structures that generalize the structure of distance relations be-
tween points of Euclidean space. Much of the chapter is given over to a
quick tour of the realm of metric spaces, aiming at familiarizing the reader
with some of the examples that feature in later discussion. There is also a
quick discussion of some rather surprising consequences for our problem
of one natural way of making precise the principle that there should be no
gaps in logical space. The chapter concludes with a brief discussion of
questions concerning the role that distance relations play for relationalists.
What sort of ontology do they involve one in? Need distance relations
themselves be regarded as fundamental?
Chapter II returns to the territory we have just run over, considering
more carefully the options available to substantivalists and relationalists
to make sense of assertions that space has one or another particular
structure at a given world. The chapter has two main goals. (i) It considers
and rejects some more conservative strategies that relationalists might
adopt in place of the modal strategy. (ii) It lays out an argument that
shows that three individually prima facie desirable conditions on accounts
of geometric possibility are jointly incompatible (in the presence of
6
For further discussion, see Appendix D below.
introduction 5
a reasonable auxiliary assumption). Roughly speaking, the three condi-

tions are:
(a) For every substantivalist world w with material configuration C,
there is a relationalist world w* whose material configuration C*
mirrors the geometry of C such that the facts about geometric
possibility at w* mirror those at w.
(b) If two worlds are geometric duplicates (i.e., agree about all
intrinsic geometric facts) then they agree about all facts about
geometric possibility.
(c) One specifies the intrinsic geometric facts that obtain at a
relationalist world by specifying the distance relations between
the material points at that world.
Once these are made a bit more precise, it is not hard to see that there is
essentially no hope of satisfying all three conditions simultaneously. For,
presumably, substantivalists think that there could be a world with
Euclidean geometry and a world with hyperbolic geometry, each contain-
ing only a single particle. These worlds differ as to what is geometrically
possible, but they agree so far as matter is concerned as to the facts about
distance (each world contains just one particle, at zero distance from itself ).
So relationalists who want to mirror substantivalist commitments will have
to allow either that the intrinsic geometric facts are not determined by the
facts about distance or that the facts about geometric possibility do not
supervene on the intrinsic geometric facts about configurations of matter.
Chapters III–V consider possible responses to this observation.
The first considers a variety of modal relationalism that gives up on
(a) while holding on to (b) and (c); the second, a variety that gives up
on (b) while holding on to (a) and (c); the third, a variety that gives up on
(c) while holding on to (a) and (b). In each case, the strategy considered can
be seen as being more or less closely related to a strategy familiar from the
literature on nomic possibility and necessity. The strategy discussed in
Chapter III (due to Nick Huggett) is self-consciously modelled on the
best-system approach to nomic modality developed by David Lewis and
others. The strategy of Chapter IV, which involves, roughly speaking, an
invocation of a primitive geometric possibility operator, is reminiscent of
the primitivism about laws of nature defended by John Carroll, Tim
Maudlin, and others. The strategy of Chapter V, which turns on encoding
facts about geometric possibility in the properties instantiated at a world, is
6 introduction
in many ways analogous to the necessitarianism about laws developed by

Sydney Shoemaker and Brian Ellis, amongst others.
In each of these chapters the relation between the nomic and geometric
cases plays an important role since in each instance the nomic and the
geometric versions of a given strategy share some characteristic strengths
and weaknesses. As will become clear, I regard the best-system approach
as the least promising of the options considered (because in both its
nomic form and its geometric form it involves implausible empirical
conjectures). In their geometric forms, the primitivist and the necessitarian
options are rather closely related—but the latter seems to me more inter-
esting because it leads to some interesting questions about the range of
geometric properties.
In a concluding chapter, I briefly survey the territory covered, and
sketch how the sort of relationalism about ontology that I favour is related
to the question of the relativity of motion and how it can be naturally
generalized to yield an account of spacetime as well as of space.
There are also five appendices.
Many readers will begin with a conviction, that when laid out in any
detail, relationalism is metaphysically baroque in comparison to substan-
tivalism—and nothing in what follows is likely to undermine this convic-
tion. On the orthodox view of the methodology of ontological inquiry,
this will suggest that relationalism is of little interest: because substantival-
ism is notably simpler than relationalism, orthodoxy tells us that we
have pretty decisive reason to think that if it is one or the other, then
the spatial ontology of our world is substantivalist rather than relationalist.
In Appendix A I give my reasons for rejecting the usual view of the
evidential role of simplicity considerations in discussions of ontology.
The resulting view of the reach and strength of metaphysical argument
occasionally plays a role in the main argument of the book.
Appendix B gives a brief introduction to an important technical
notion—the limit of a sequence of metric spaces. This notion bears on
the content of the principle that there should be no gaps in logical space
(discussed in Section 4 of Chapter I) and also on a question that arises in the
discussion of conservative relationalism in Chapter II.
Two of the remaining appendices are historical in nature. Appendix C is
essentially a bibliographical essay, consisting of a brief overview of the
substantival-relational debate before the time of Newton and Leibniz.
introduction 7
Appendix D is more detailed, and argues in favour of the standard ( but not
universally accepted) view that Leibniz was a modal relationalist.
In the course of the discussion of Chapter IV, a certain technical
question arises: In what sort of metric spaces does knowing the distance
relations between a set of points determine the embedding of those points
in the space up to a symmetry of the space? This question is briefly
discussed in Appendix E.
I
Possible Structures of Space
Our approach to general metric spaces bears the undeniable imprint

of early exposure to Euclidean geometry. We just love spaces sharing
a common feature with Rn.
Gromov
1 Introduction
The goals of this preliminary chapter are primarily of an expository nature:
to get on the table for later use some mathematical notions and facts that
provide a framework for talking about the structure of space at various
possible worlds.
In order to avoid complete tedium, the discussion is structured around
the question: Which mathematical structures should we think of as repre-
senting the spatial structure of metaphysically possible worlds? For short:
What are the possible structures of space?
The question will strike many readers as a silly one. Some will think it
has a determinate answer—but that we have little hope of finding it.
Others will doubt that the question is anywhere near sharp enough to
admit of a determinate answer.
The question is a little silly. But I don’t think it is completely pointless.
Most of us have some intuitions about this question—so our concept of
spatial geometry is not a completely formless one. Here I will in effect be
developing one way of assigning content to this concept. I proceed partly
by exploring our ordinary concept, partly via stipulation.1
I take it to be obvious that Euclidean three-space corresponds to a
possible structure of space. And I think it reasonable to assume the class
1
So often ‘argue,’ ‘argument,’ etc. have invisible scare quotes in this and the next two
sections.
possible structures of space 9
of mathematical structures that represent possible spatial geometries cor-

respond to some suitable natural generalization of Euclidean geometry.2
And I think we have (hazy and infirm) views about what counts as natural
and suitable here.
Platitudinous though these commitments are, some think that they
should be rejected out of hand. For some philosophical views imply that
only physics can tell us which structures correspond to possible spatial
geometries. This happens if metaphysical possibility and physical possibility
are taken to coincide.3 Or if it is denied that there is a coherent notion of
metaphysical possibility.4 Or if it is accepted that metaphysical possibility
outstrips physical possibility, but held that only considerations arising from
the sciences can give us the relevant modal knowledge.5 But each of these
views involves a radical and (to my mind) unmotivated revision of our
ordinary conception of possibility and I set them aside here.
Other philosophers will think that the methodology sketched above is
okay in so far as it goes, but that it needs to be supplemented by further
constraints. For instance, David Lewis at one time held that possible worlds
form a set rather than a proper class—so he was careful to stipulate that
there must be some sort of limitation of size on structures representing
possible spatial geometries.6 In fact, Lewis’s concerns were somewhat
idiosyncratic and he seems to have backed away from them in later
work.7 In any case, the discussion below is such that constraints of this
kind can easily be plugged in.
Others still will be puzzled by my willingness to talk about the spatial
geometry of a world in abstraction from its physics, since they take claims
about geometric structure to have meaning only in the context of a
2
Bricker argues for a view like this in “Plenitude of Possible Structures.”
3
See Shoemaker, “Causal and Metaphysical Necessity” and Bird, Nature’s Metaphysics, chs. 3
and 8. For critical discussion, see Fine, “The Varieties of Necessity” and Sidelle, “On the
Metaphysical Contingency of Laws of Nature.”
4
This seems to be the position of Leeds in “Physical and Metaphysical Necessity” and of
Maudlin in the Epilogue to The Metaphysics within Physics.
5
In “Plenitude,” Bricker interprets Adams as expressing a view like this in §III of his
“Presumption and the Necessary Existence of God.” There is room to question this inter-
pretation—see Adams, op. cit., pp. 30 f. and Leibniz, p. 213.
6
On the Plurality of Worlds, §§1.8 and 2.2. One gets “too many” possible spatial geometries
if one can associate with each cardinal number a distinct structure representing a possible
spatial geometry—e.g., if there is such a structure of cardinality Œ for each cardinal number Œ.
7
For critical discussion see Nolan, “Recombination Unbound.” For Lewis’s reaction see
“Tensing the Copula,” p. 8.
10 possible structures of space
physical theory.8 What does it mean to speak of the distance relations at a

world if nothing has been said about measurement or principles of coord-
ination? My lack of engagement with this worry below should be read not
as a dismissal, but as founded on the hope that for any possible spatial
geometry to be considered, we could if pressed cook up a (perhaps very
artificial) story about what sort of physics would go along with that spatial
geometry in a satisfying nomic-geometric package. I aim to remain neutral
here on the relation between geometry and physics.
This is how things are going to go. In the next section, I narrow down
the range of options we have to consider, arguing that it is safe to consider
generalizations of the distance structure of Euclidean space and sketching
some of the mathematical notions that arise naturally in such generaliza-
tions. There follows a survey of a number of such generalizations, in which
it is argued that we will probably want to recognize at least some spaces
from each of the classes considered as representing possible spatial struc-
tures. The following section considers one way of making sense of the
principle of the plenitude of possibilities and discusses some implications
that this approach has for our question. In a final section, I turn to a
question that will probably have bothered some ( but only some) readers all
along—How are we to think of the distance relations?
2 Distance
Let us turn, then, to the project of considering generalizations of Euclidean
geometry. Of course, Euclidean geometry can be characterized in many
ways. And each variant characterization will support a plethora of gener-
alizations.
Where to begin? Consider the most famous sort of axiomatization of
Euclidean geometry, the quantifiers of which range over points and the
non-logical vocabulary of which consists of non-quantitative relations like
betweenness and congruence.9 We can generalize this approach by con-
sidering the result of tinkering with the primitives or the axioms in various
ways—and in this way we will easily reach generalizations that include the
8
For recent influential defences of this view of the relation between geometry and physics,
see Friedman, Dynamics of Reason and Brown, Physical Relativity.
9
A variety of approaches of this kind are discussed in pt. i of Henkin et al. (eds.), The
Axiomatic Method.
classical non-Euclidean geometries (elliptic, hyperbolic, and spherical

geometry). And if we give up our demand for nice axiom systems, we
can also capture more general spaces (e.g., Riemannian manifolds) as
spaces whose structure can be characterized in terms of betweenness and
congruence.10
All of this is to the good. But we also have to ask whether the
generalizations we consider permit mathematical structures with no
claim to represent possible spatial geometries. Among the natural gener-
alizations of the sort of characterization we are considering is the class of all
structures in which a set of objects are organized by finitely many relations.
Surely we won’t want to count each such structure as representing a
possible structure for space—but which ones should we so count? An
intuitively plausible sufficient condition is that the structure in question
supports a natural notion of distance between points.11 Is this condition
also necessary? The suggestion has some attractions. But there are a few
cases, such as that of projective geometry, in which one may feel some
temptation to count a structure as corresponding to a possible spatial
geometry even though it doesn’t support a notion of distance.
We could reach the same sort of point by starting with other charac-
terizations of Euclidean geometry. Suppose that we think of Euclidean
space as a manifold carrying a certain sort of tensor. What generalizations
of this characterization count as permitting only possible spatial geom-
etries? We go too far if we allow manifolds equipped with arbitrary
tensors. Again, it would seem that if the structure contemplated comes
equipped with a reasonable notion of distance, this provides something
like sufficient grounds for allowing that it represents a possible spatial
geometry. Is coming equipped with a notion of distance also a necessary
condition? This suggestion has some attractions, although, there may again
be some cases (such as that of symplectic spaces) that give one pause.
This pattern suggests that under one reasonable (albeit perhaps some-
what restrictive) sense of ‘spatial geometry,’ our task of singling those
mathematical structures that correspond to possible spatial geometries
can be recast as the task of characterizing those equipped with notions that
can reasonably be considered notions of distance. That is a wishy-washy
conclusion. But it is good enough for present purposes. So let us proceed
10
See Mundy, “Space-Time and Isomorphism.”
11
Or perhaps we should prefer a notion of ratios of distance between pairs of points. For
expository convenience, I usually ignore this subtlety. See §5 below for discussion.
by considering a characterization of Euclidean space in terms of the

distance relations between its points, then look for suitable generalizations
of this characterization.
Euclidean Geometry
For convenience, let us consider the Euclidean plane. We work with a set
X of points and and function d that assigns non-negative real numbers to
pairs of points. The space (X, d) has the usual structure of the Euclidean
plane (supplemented by a choice of unit) if and only if it satisfies the
following eight axioms.12
Axiom I. d(x, y) ¼ 0 if and only if x ¼ y.
Axiom II. d(x, y) ¼ d(y, x).
Axiom III. The triangle inequality: d(x, z) # d(x, y) þ d(y, z).
We say that y is between x and z if d(x, z) ¼ d(x, y) þ d(y, z) and y 6¼ x, z.
Axiom IV. If x 6¼ z then there is a y between x and z.
Axiom V. If x 6¼ y then there is a z such that y is between x and z.
Axiom VI. There are distinct points x1 , x2 , x3 2 X such that no one
of them lies between the other two.
We say that a sequence of points {xn } converges to x if for every > 0
there is an N > 0 such that d(x, xn ) < for all n > N. We call a
sequence of points {xn } a Cauchy sequence if for every > 0 there is an
N > 0 such that d(xm , xn ) < for all m, n > N.
Axiom VII. Every Cauchy sequence of points in X converges to
some x 2 X.
Given four points x1 , x2 , x3 , x4 2 X, we write aij for d(xi , xj ). The
following is the Cayley–Menger matrix for the xi .
0 1 1 1 1
1 0 a212 a213 a214
1 a212 0 a223 a224
1 a213 a223 0 a234
1 a214 a224 a234 0
Axiom VIII. For any x1 , x2 , x3 , x4 2 X the determinant of the

Cayley–Menger matrix vanishes.
12
See Blumenthal, A Modern View of Geometry, ch. vii.
Metric Spaces
Let us now begin the task of considering generalizations of the distance
structure of Euclidean geometry.
Once we have made a decision to encode facts about distance in a
real-valued function on pairs of points, Axioms I–III would appear to be
unavoidable: it is difficult to imagine calling anything a notion of distance
which did not satisfy them.13
The remaining axioms do not have this status: it is not difficult to think
of examples in which they fail, but which are nonetheless plausible
examples of spatial geometries. Axiom IV is violated, e.g., if the points
of space are discrete atoms. Axiom V can be violated, e.g., when space is
finite in extent. Axiom VI is violated if space is one-dimensional. Axiom
VII is violated if space has gaps of a certain kind—e.g., if space is all ordered
pairs of rational numbers (with the usual Euclidean notion of distance).
Axiom VIII is violated, e.g., in the classical non-Euclidean plane geometries.
So the class of structures satisfying just Axioms I–III for Euclidean
geometry provides an interesting upper bound for the class of structures
we are looking for. Let us codify this with a definition: a metric space (X, d )
is a set X together with a function d that assigns non-negative real numbers
to pairs of points of X in such a way that for any x, y, z 2 X the following
conditions are satisfied: (i) d(x, y) ¼ 0 if and only if x ¼ y; (ii) d(x, y) ¼ d(y, x);
(iii) d(x, z) # d(x, y) þ d(y, z). We call d a distance function on X.
The notion of a metric space abstracts just a few of the most essential-
looking features of Euclidean space. The resulting family of metric spaces is
very capacious. But the notion of distance built into the definition is strong
enough to support a number of familiar and important mathematical
notions. We will rely on a few of these in considering various generaliza-
tions of Euclidean geometry in the next section.
A map f : X ! Y between metric spaces (X, d ) and (X * , d * ) is distance-
preserving if d(x1 , x2 ) ¼ d* ( f (x1 ), f (x2 ) ) for all x1 , x2 2 X. A distance-
preserving map is an isometry if it is also bijection (i.e., every y 2 Y is the
13
Later we will consider alternatives to the strategy of taking distances to be encoded via
an assignment of real numbers to pairs of points (see pp. 28 ff.). But if this strategy is pursued,
then Axioms I and II seem unassailable. One could conceivably entertain (but not sustain)
scepticism about Axiom III, worrying that it represents an artificial constraint on the notion of
distance; for something along these lines, see Blumenthal, “Distance Geometries,” §4. But
most philosophers seem happy to take Axiom III to be somehow constitutive of our concept
of distance. See e.g. Bricker, “The Fabric of Space,” p. 382; Forrest, “Is Space-Time Discrete
or Continuous?,” p. 329; Maudlin, Metaphysics within Physics, §3.2.
image under f of some x 2 X and f (x1 ) ¼ f (x2 ) implies x1 ¼ x2 ). We say that

two metric spaces share a metric geometry if they are related by an isometry.
Given a point x in a metric space X and a real number r > 0 we can
define some interesting subsets of X. (i) The open ball of radius r centred at x
is Br (x): ¼ {y 2 X: d(x, y) < r}. (ii) The closed ball of radius r centred at x is
Br(x) : ¼ {y 2 X : d(x, y) # r}.14 (iii) The sphere of radius r centred at x is
Sr (x): ¼ {y 2 X: d(x, y) ¼ r}.
As in the Euclidean case, we can then define a notion of convergence
and related notions. We say that a sequence of points {xn } converges to x if
for every > 0 there is an N such that d(x, xn ) < for all n > N (i.e., the
sequence {xn } is eventually restricted to B (x)). We then say that a
function from one metric space to another is continuous if it maps
convergent sequences to convergent sequences (this generalizes the
usual notion of continuity for functions from the real line to itself ).
We again call a sequence of points {xn } a Cauchy sequence if for every >
0 there is an N such that d(xm , xn ) < for all m, n > N. A Cauchy
sequence is a sequence that looks like it should converge, given
the pattern of relative distances between elements of the sequence.
We call a metric space complete if every Cauchy sequence converges. The
real numbers form a complete metric space. Examples of incomplete
metric spaces include the rational numbers and the positive real numbers.15
3 Candidates for Spatial Structure

In this section we consider a sequence of generalizations of the distance
structure of Euclidean space (with one exception, each of these is wider
than its predecessor). Each corresponds to a family of metric spaces. In
most cases it is impossible to do more than hint at the full range of
structures encompassed. In each case I mention examples of structures
permitted that are not permitted by the more narrow generalizations
14
Warning: despite what the notation may suggest, in general B r(x) need not be the
closure of Br(x) (e.g., let r be the distance between the points of a two-point metric space).
15
Note that completeness is not a topological notion: under their standard metrics, the
real numbers form a complete space while the space of positive real numbers is incomplete—
but these two spaces have the same topology. Note further that completeness and incom-
pleteness are intrinsic notions: we do not have to think of the space of rational numbers as a
subset of the space of real numbers in order to determine that it is incomplete; and the space of
real numbers is complete whether or not we think of it as a subspace of some larger space
(such as the space of complex numbers).
considered—and in some cases I mention examples of structures permitted

that may well seem too odd to count as possible structures of space.
Generalization 0: Euclidean Geometry. Our first generalization

encompasses the Euclidean spaces of all (finite) dimensions, considered as
metric spaces.16 At one time, of course, philosophers trafficked in argu-
ments purporting to show that space was necessarily three-dimensional.17
But few today are likely to deny that space could have had two, or four, or
twenty-six dimensions. And if these are allowed, it seems parochial to
exclude Euclidean spaces of any finite dimension.
Generalization 1a: The Classical Geometries. This class comprises

Euclidean geometry and the three geometries that were first recognized as
serious alternatives to it: hyperbolic geometry, spherical geometry, and
elliptic geometry. Each admits an elegant characterization in terms of the
distance relations between points.18 A version of each exists in any finite
dimension; I focus here on the two-dimensional case.
Hyperbolic geometry is the geometry associated with Gauss, Bolyai, and

Lobachevski. It arises when one replaces the classical parallel postulate
of Euclidean geometry by an axiom that says that given a line and a
point not on that line, there exist at least two lines through the point
parallel to the given line.
Spherical geometry arises when we consider the points lying on the
surface of a sphere in Euclidean space, with the distance between two
points given by the shortest path between them that is itself restricted
to the surface of the sphere. In spherical geometry, lines are given by
great circles (the intersections of the sphere with planes through the
origin in the ambient space). In spherical geometry distinct lines
always intersect in exactly two points, so there are no parallels.
Spherical geometry has some awkward features: in Euclidean and
hyperbolic geometry, two lines that intersect determine a single point
and any two points determine a unique line, but neither of these
assertions holds in the spherical case.
16
For characterizations of each of these spaces and of the class of all of these spaces, see
Blumenthal, Theory and Applications of Distance Geometry, §§48 and 50.
17
Callender, “Answers in Search of a Question” surveys the fortunes of a line of argument
initiated by Kant. We will encounter an argument due to Leibniz in Appendix D below.
18
For the two-dimensional case, see Blumenthal, Modern View, chs. vii and viii.
Elliptic geometry is locally like spherical geometry (i.e., for every point,
there is a ball around that point isometric to a ball in spherical geom-
etry). But the elliptic plane has a global structure different from that of
spherical geometry, and in this way it manages to avoid the awkward
features of spherical geometry noted above. In the standard case, we
begin with a sphere of unit radius and in the resulting elliptic space lines
are closed curves of length . One can think of the elliptic plane as
constructed out of spherical geometry by identifying diametrically
opposite points. It can also be thought of as follows: points in
the elliptic plane are represented by lines through the origin in R3 ;
the distance between two points in the elliptic plane is given by the
smaller of the angles formed by the corresponding lines in R3 ; a line in
the elliptic plane is represented by all of the lines through the origin
in R3 that lie on some plane passing through the origin.
A traditional topic in philosophy of space and time concerns status and
implications of the axiom of free mobility (this says, roughly speaking, that
bodies may be transported without altering their shape). One can show
that within a very large family of metric spaces, the only ones satisfying a
certain strong version of the axiom of free mobility are the Euclidean,
hyperbolic, and spherical spaces of various dimensions.19 For this reason
and others, it is helpful to have a label that applies to the Euclidean,
hyperbolic, and spherical spaces (but not to the elliptic spaces): I will
speak of these three as the elementary geometries. Elliptic geometry will
prove to be an important source of examples in later chapters.
Generalization 1b: Flat Spaces. Let us now generalize Euclidean

geometry in a quite different direction. A path in a metric space is a
continuous map ª : [a, b] ! X from a closed interval of real numbers to
X.20 A metric space X is a flat space if it satisfies the following three
conditions. (i) X has the local structure of Euclidean space—for every x
2 X there is an open ball centred at x that is isometric to an open ball in a
Euclidean space. (ii) Any two points in the space can be connected by a
path. (iii) The distance between points is given by the greatest lower
bound of the lengths of the paths connecting them.21
19
See Appendix E below.
20
For real numbers a and b such that a < b, the closed interval [a, b] is {t 2 R : a # t # b}
while the open interval (a, b) is {t 2 R : a < t < b}.
21
In the context of (i), the content of conditions (ii) and (iii) should be intuitively clear. These
conditions will be explained more thoroughly in the discussion of path metric spaces below.
Euclidean geometry is the only one of the classical geometries that

qualifies as a flat space. While flat spaces have the local structure of
Euclidean space, they can have quite different global structures. One
way to make a flat space is to begin with an n-dimensional Euclidean
space and then to excise some region A of the right form (closed, with
connected complement).22 In the new space X: ¼ {x 2 Rn : x 2 = A} the
distance between points x and y is given by the greatest lower bound on
the lengths of paths in X that connect x and y. For some A, the distance
relations for points in X will be identical to the relations between those
points in Rn (e.g., if A consists of a single point, or of the region external to
some open ball). But in general, although the distance relations between
nearby points will be unchanged, some alterations will occur in the relations
between faraway points (consider the case where A is a closed ball).
Flat spaces constructed according to the recipe of the preceding para-
graph are never complete (consider a sequence of points in X whose limit
in the original Euclidean space lies in A). There also exist complete flat
spaces, which are called flat space forms. Flat space forms are much harder to
come by than are incomplete flat spaces. For instance, in two dimensions
the possible global structures for a complete flat space are: the plane, the
cylinder, the twisted cylinder, the torus, and the Klein bottle.23 Roughly
speaking: to make a cylinder, we cut an infinite strip out of the Euclidean
plane and glue the edges together; to make a twisted cylinder, we again
begin with a strip, but now give it a twist before gluing the edges together;
to make a torus, we cut a square out of the Euclidean plane and glue
opposite edges together; to make a Klein bottle, we again begin with a
square but now give it a twist before gluing the edges.
The flat space forms are familiar characters in the philosophy of space
and time, figuring in just about every contemporary discussion of
Kant’s incongruent counterparts argument. Presumably this indicates a
22
A subset U X of a metric space is open if for every x 2 U, there is some open ball
around x in X that is contained in U. A set A X is closed if is the complement of some open
set. In the present context, a set is connected if and only if every pair of points in it can be
connected by a curve in the set.
23
These are the possible topologies for complete connected flat Riemannian two-mani-
folds (each topology admits a small family of isometry classes of flat Riemannian metrics); see
Wolf, Spaces of Constant Curvature, §2.5. Our flat spaces are path metric spaces because of
condition (iii) in their definition; and it follows that they are Riemannian manifolds because
by assumption they are locally isometric to a Riemannian manifold; see Burago et al., A Course
in Metric Geometry, pp. 38 and 143.
widespread willingness to consider them as representing possible structures

for space.
What about incomplete flat spaces? These are a bit stranger: they feature
inextendible straight lines that are of finite length not because they form
closed curves but because they “run out of space.” This is easily pictured in
the simplest examples such as the space that results from deleting a single
point from Euclidean three-space and the space that results from deleting
everything outside of some open ball. The latter corresponds to Aristotle’s
cosmos—extension forms a solid sphere, with nothing, not even empty
space, external to it. There is an ancient objection to this sort of cosmos:
suppose that we approach the edge of the cosmos and attempt to throw a
spear outwards beyond the edge; if we are unsuccessful, then something
beyond the cosmos prevents us; if we are successful, then there is somewhere
beyond the cosmos.24 The argument appears to have no contemporary
adherents—in part, I suppose, because general relativity has schooled us to
think that the matter can simply disappear when it reaches the edge of
geometric extension. In any case, since space(time)s that arise from nice
spaces via deletions are sometimes taken to pose problems for relationalists,
any honourable relationalist is under pressure to recognize such possibilities.25
Generalization 2: Spaces of Constant Curvature. This class arises

when we consider well-behaved spaces which have the local structure of
one of the elementary geometries (Euclidean, hyperbolic, spherical).26
That is, a metric space X is a space of constant curvature if it satisfies the
following three conditions.27 (i) X has the local structure of one of the
elementary geometries—for every x 2 X there is an open ball centred at x
that is isometric to an open ball in the elementary geometry in question.
(ii) Any two points in X can be connected by a path. (iii) The distance
between points is given by the greatest lower bound on the lengths of the
paths connecting them.
We call spaces of constant curvature with the local structure of spherical
geometry spaces of positive constant curvature, those modelled on Euclidean
24
Lucretius, De Rerum Natura, i.968–83. The provenance and influence of this argument
are discussed in Sorabji, Matter, Space and Motion and Grant, Much Ado About Nothing.
25
For the challenge that such possibilities pose to relationalism, see Earman, World Enough
and Space-Time, §8.4.
26
The reasons for being interested in flat spaces carry over to this case.
27
Again, I rely on the reader’s intuitive sense of the content of (ii) and (iii) in the context
of (i). See the discussion of path metric spaces below for a detailed account.
geometry spaces of vanishing curvature (this is a variant on ‘flat space’),

and those modelled on hyperbolic geometry spaces of constant negative
curvature.28
Space forms are complete spaces of constant curvature. In two dimen-
sions, the space forms of positive constant curvature are just given by
spherical geometry and elliptic geometry.29 The story for space forms of
negative constant curvature is much more interesting—to start with, even
in the two-dimensional case there are infinitely many topologies that such
spaces may have.30 There are of course also many incomplete spaces of
constant curvature—such as those that arise via excision of suitable subsets
from any of the classical geometries.
Generalization 3: Riemannian Manifolds. The notion of a Riemann-

ian manifold is a relatively technical one. Here I keep things as simple as
possible by following an ugly short cut.31
Let X be a subset of Rn such that every pair of points x, y 2 X can be
joined by a path that lies in X.32 We make X into a metric space by using
the usual Euclidean metric on Rn to calculate the length of X-paths, and
taking the distance between points in X to be given by the greatest lower
bound on the lengths of the X-paths joining them. We now forget about
how we constructed (X, d ) and think of it intrinsically as a metric space in
its own right.
28
As in the flat case, all of our spaces of constant curvature are Riemannian manifolds. A
Riemannian manifold has constant curvature k if at every point the sectional curvature of every
tangent plane at that point is k. The sectional curvature of a tangent plane is defined as follows.
Let x be a point in a Riemannian manifold M and let — be a tangent plane at x (i.e., a two-
dimensional linear subspace of the tangent space at x). Let Br,—(x) be the subset of Br(x)
consisting of those points that can reached by geodesics from x of length r or less and whose
tangent vectors at x lie in —. For sufficiently small r > 0, Br,— (x) is a well-behaved surface
sitting inside M. For small > 0, let L() be the circumference of the circle in this surface with
centre x and radius . The quantity K(—): ¼ lim!0 33 (2r L() ) is the sectional curvature of
the tangent plane — at x.
29
In fact, this is true in any even dimension. The story is more complicated in odd
dimensions. See pt. iii of Wolf, Spaces of Constant Curvature.
30
For the compact, oriented case, see e.g. ch. B of Benedetti and Petronio, Lectures on
Hyperbolic Geometry.
31
For complaints about the ugliness of this route, see Gromov, Metric Structures for
Riemannian and Non-Riemannian Spaces, p. xvi.
32
i.e., there is a path ª : [a, b] ! Rn with ª(a) ¼ x and ª(b) ¼ y such ª(t) 2 X for each
t 2 [a, b].
The resulting metric space is called a Riemannian manifold when X satisfies

a further technical condition whose point is to ensure that the space provides
a hospitable environment for the notions of the differential calculus. The
condition in question is that for each point x 2 X one is able to find an > 0
and a set {x1 , . . . , xn } of (possibly curvilinear) coordinate functions on Rn
defined for points in B (x) Rn such that for some m # n the set X \ B (x)
is given by the condition that x1 , . . . ,xm vanish.33 This ensures, amongst
other things, that X not include its own edges, that its dimension does not
vary from point to point, that it does not have any self-intersections, that it
does not have any cusps or other singularities, and that it does not have any
regions that approach one another arbitrarily closely in Rn while being
separated by a finite distance according to the induced metric on X.
Every space of constant curvature is a Riemannian manifold. But the
notion of a Riemannian manifold is far more general than the notion of a
space of constant curvature—the local geometry can vary from point to
point in a Riemannian manifold and at near no point need it look like that
of one of the classical geometries.34 Consider, for example, the surface of a
doughnut. This has the global structure of the flat torus, discussed above—
but it inherits from its standard embedding in R3 a non-flat metric
structure.35 And of course things just get more complicated if we allow
the surface of the doughnut to involve lots of indentations and bulges.
In any general relativistic world in which it makes sense to speak of the
geometry of space at an instant of time, that geometry always has the
structure of a Riemannian manifold. For this reason, Riemannian mani-
folds are often the most general spaces considered in discussions of rela-
tionalism and substantivalism, and it is often tacitly or explicitly assumed
that every Riemannian manifold corresponds to possible spatial geometry.
Part of the point of the discussion below is to call into question the
motivation for restricting attention to the Riemannian case.
Remark 3.1. Readers familiar with the standard approach to Riemannian
geometry, as involving smooth manifolds equipped with metric tensors,
may wonder whether anything is lost in treating a Riemannian manifold as
33
i.e., X is an embedded Riemannian submanifold of Rn. That all Riemannian manifolds
arise in this way follows from the Nash embedding theorem.
34
In a general Riemannian manifold, one finds variation in the sectional curvature of
tangent planes even if one restricts attention to tangent planes at a single point.
35
In a flat space, if we look at a small scale, all directions look the same. This is not true for
a torus embedded in R3. The flat structure on the torus arises from an embedding in R4.
a metric space. Clearly, any Riemannian metric tensor on a (connected)

manifold induces a metric space structure on the point set of the manifold
( just take the distance between two points to be given by the greatest lower
bound on lengths of piece-wise smooth curves connecting the points). But
does the metric space structure of a Riemannian manifold determine its
smooth structure and Riemannian metric tensor? Yes. Any metric space
isometric (as a metric space) to a Riemannian manifold is a Riemannian
manifold; and any metric space isometry between Riemannian manifolds
preserves their smooth structures and metric tensors.36 Further: there exists a
direct characterization of Riemannian manifolds in the vocabulary of metric
spaces: see Remark B.3 in Appendix B below. &
Generalization 4: Generalizations of Riemannian Manifolds. There

are diverse modest generalizations of the notion of a Riemannian mani-
fold.37 For instance, one can loosen the requirements so as to allow spaces
that fall short of being Riemannian manifolds only because they include their
own boundaries or feature occasional mild singularities of various sorts.
Of more interest here is that fact that it is also possible to allow infinite-
dimensional analogues of Riemannian manifolds.
If we had followed the standard route (rather than the ugly short cut) in
characterizing Riemannian manifolds, we would have first introduced the
notion of a manifold as a space that has the local topological and differen-
tial structure of Rn for some n 2 N, then shown how such spaces can be
made into metric spaces by being equipped with Riemannian metrics
(gadgets that assign lengths to curves by assigning a length to the tangent
vector at each point along a curve).
The standard approach can be generalized to allow infinite-dimensional
versions of Riemannian manifolds. Recall that a Hilbert space is a vector
space equipped with well-behaved notions of the length of a vector and of
the angle between vectors.38 For each natural number n, Rn equipped
with its Euclidean structure is a Hilbert space. But there are also infinite-
dimensional Hilbert spaces. Now, in analogy with the standard definition
36
For these and other results, see Burago et al., Course, §5.1; and Petersen, Riemannian
Geometry, ch. 5, esp. §§5.10.1 and 5.10.3. Grove, Riemannian Geometry offers a treatment of
Riemannian manifolds as metric spaces.
37
For a compact overview, see Berger, Riemannian Geometry During the Second Half of the
Twentieth Century, §TOP.9.
38
For details, see e.g. §5.5 of Folland, Real Analysis.
of a Riemannian manifold, we take a Hilbert manifold to be a space with the

local topological and differential structure of a Hilbert space and define a
Riemannian Hilbert manifold to be a Hilbert manifold with a distance
function that arises from a Riemannian metric.39
There is a sense in which the transition from Riemannian manifolds to
Riemannian Hilbert manifolds involves a massive generalization. There
are a lot of Riemannian manifolds.40 But there are many, many, more
Riemannian Hilbert manifolds—so many, that they form a proper class
rather than a set.41
For my own part, I would be inclined to allow that at least the tamer of
the Riemannian Hilbert manifolds correspond to possible spatial geom-
etries. But since some readers will doubtless stick at this, such monsters play
no essential role in what follows.
Generalization 5: Path Metric Spaces. In all of the spaces we have
considered so far, the distance between two points is given by the greatest
lower bound on the lengths of paths joining those points. Let us define a
path metric space to be a metric space in which distances are determined by
the lengths of curves in this way. The following remark shows that the
length of a path is a well-defined notion in any metric space.
Remark 3.2. Recall that for a, b 2 R with a < b, the closed interval is [a, b] : ¼ {x
2 R : a # x # b} and that a path in a metric space X is a continuous function of
the form ª : [a, b] ! X. A partition of [a, b] is a finite set Y ¼ {y0 , . . . , yn } with
yi < yiþ1 , y0 ¼ a, and yn ¼ b. Define for any partition Y:
P P
i¼n
(Y ) ¼ d((yi1 ), (yi ) ):
i¼1
Then we take L(ª), the length of ª, to be the least upper bound on (Y ) as
Y varies over all partitions of [a, b]. &
39
For details, see Lang, Fundamentals of Differential Geometry or Abraham et al., Manifolds,
Tensor Analysis, and Applications.
40
For each n $ 2 there are at least continuum-many diffeomorphism classes of differenti-
able manifolds; Eichhorn, Global Analysis on Open Manifolds, proposition 5.0.3. Any differ-
entiable manifold admits continuum-many isometry classes of Riemannian metrics.
41
Every Hilbert space is an infinite-dimensional Riemannian manifold; see Abraham et al.,
Manifolds, p. 353. And for every cardinal number Œ, there is a Hilbert space that has Œ dimensions,
in the sense that its Hilbert space bases have cardinality Œ; see e.g. Folland, Real Analysis, p. 169.
As we will see below, not every metric space is a path metric space.
Every Riemannian (Hilbert) manifold is a path metric space—but there
are many non-Riemannian path metric spaces.
Example 3.1 (A Non-standard Metric on the Plane.). The taxi-cab metric on
R2 is given by:
d( (x1 , y1 ), (x2 , y2 ) ) ¼ jx1 x2 j þ jy1 y2 j:
The standard Euclidean metric and the taxi-cab metric disagree systemat-
ically about distances. This disagreement is reflected, e.g., in a divergence
over the shape of balls and spheres: these are round in the Euclidean case
and diamond-shaped in the taxi-cab case. But this makes no difference to
the notions of convergence for these spaces (or, therefore, of continuity): a
sequence of points in R2 that converges according to one of these notions
of distance if and only if it converges according to the other. It is not hard
to see that the plane equipped with the taxi-cab metric is not a Riemann-
ian manifold.42 &
Example 3.2 (Axes of Evil). Begin with the plane equipped with the taxi-
cab metric. Delete all points except those lying on the x- and y-axes, but
leave the distance relations between the remaining points as in the original
space. Using an obvious convention according to which the name of a
point tells us which axis it lies on and the value of the relevant coordinate
on that axis, the metric on this space is given by the conditions:
d(x1 , x2 ) ¼ jx1 x2 j, d(y1 , y2 ) ¼ jy1 y2 j, and d(x, y) ¼ jxj þ jyj. This
space is not a Riemannian manifold (it doesn’t even have the local
structure of Rn at the point where the two axes meet). &
Example 3.3 (Metric Bouquets). Let {XÆ } be a set of metric spaces. Choose
a distinguished point xÆ in each XÆ . The bouquet of the XÆ is the result of
gluing them together by identifying each of the points xÆ .43 Even if the XÆ
are Riemannian manifolds of the same dimension, the resulting bouquet
will typically not be Riemannian—as in the previous example. &
Example 3.4 (Graphs). An edge is a metric space isometric to a closed finite
interval of the real line. Each edge has two endpoints. A metric graph is a
42
Under the taxi-cab metric, there are length-minimizing routes from (1, 0) to (1, 0)
and to (0, 1) that initially travel along the x-axis. But Riemannian geodesics never coincide
and then diverge in this way; see e.g. Petersen, Riemannian Geometry, §5.2, lemma 7.
43
For details, see §4.2 of Burago at al., Course.
path metric space constructed by gluing together a collection of edges,

with gluing allowed only at endpoints. A graph is finite if it assembled out
of finitely many edges. &
There are some pretty weird Riemannian manifolds (e.g., specimens
with zillions of dimensions or strangely convoluted geometries). To rec-
ognize such structures as representing possible spatial geometries while
denying that status to the plane with the taxi-cab metric or to the space
made out of the x- and y-axes seems entirely unmotivated. And if one lets
in the axes it will seem arbitrary to forbid finite graphs and metric bouquets
of finitely many Riemannian manifolds.
Generalization 6: Metric Spaces. Of course, many metric spaces are

not path metric spaces—either because there are not enough paths to
determine the distances between points or because although there are
plenty of paths the distances between points are not determined by the
lengths of paths connecting them.
A very helpful technique for constructing metric spaces is restriction: if
(X, d ) is a metric space and Y is a subset of X, then Y inherits from X a
distance function under which it becomes a metric space—just take the
distance between two points in Y to coincide with the distance that they
have in X.
A metric space (X, d) is discrete if for every x 2 X there is an > 0 such
that x is the only point in the ball B (x). Any convergent sequence {xn } in
a discrete space must eventually be constant (i.e., there is an N such that
xm ¼ xn for m, n > N ). It follows that distinct points in a discrete metric
space X cannot be joined by a path.44 So no discrete metric space with
more than one point is a path metric space.
Discrete metric spaces are easy to come by.
Example 3.5 (Finite Metric Spaces). Any metric space whose set of points is
finite is discrete. (Such sets arise, e.g., by restricting the metric in an
arbitrary metric space to some finite subset of points.) &
Example 3.6 (The Discrete Metric). Let X be any set. We define a distance
function d for X by letting d(x, y) ¼ 1 if x 6¼ y (as always, d(x, x) ¼ 0). d is
called the discrete metric on X. &
44
Paths are continuous. So they map convergent sequences of real numbers to convergent
sequences of points in X—and so cannot “jump the gaps” between points in discrete spaces.
Example 3.7 (Integer Lattices). Let Rn be equipped with its Euclidean

metric and consider the integer lattice, Zn , the set of ordered n-tuples of
integers. As a subset of Rn , Zn , inherits a notion of distance, under which it
is a metric space in its own right.45 &
Examples of spaces in which there are plenty of paths but in which
distances are not determined by the lengths of paths are also easy to come by.
Example 3.8 (Euclidean Metric on the Sphere). Begin with R3 with its
standard Euclidean metric and consider the set of points one unit from the
origin. Equip this set with the metric d it inherits from Euclidean space.
The result is not a path metric space—if we measure distances by lengths of
paths on the surface of the sphere, then the South Pole is twice as far from
the North Pole as the equator is, whereas relative to the Euclidean distance,
the ratio of these distances is given by the square root of two.46 &
Are there possible worlds whose spatial structure is that of a metric space
that fails to be a path metric space? Most philosophers will allow that some
discrete metric spaces have this status. Beyond that there is room for
dispute.
Consider first the question whether distant points of space stand in
direct geometric relations or whether these relations are mediated by the
relations between points lying between them.47 For example: are distance
relations between pairs of points direct or do they depend in some sense on
facts about how much space lies between the two points? The question
can be given intuitive content by considering whether one could change
the geometric relations between two given points by annihilating inter-
vening parts of space: Would the distance between (2, 0) and (2, 0) in R2
be altered if we annihilated the ball of radius one centred at the origin?48
Some authors are happy to take spatial relations between distant points
to be primitive.49 Others see some advantage in taking geometric relations
between distant points to be mediated. (i) This picture corresponds closely
45
For other notions of distance for Zn see Forrest, “Discrete or Continuous?”
46
For more examples in the same spirit, see Bricker, “Fabric.”
47
Warning: the fact that we are representing spatial geometries via metric spaces is not
meant to prejudge this question. For discussion of the relation between representor and
represented in this context, see §5 below.
48
A more careful treatment would cash the question out in terms of pairs of worlds sharing
some but not all of their geometric features. See Bricker, “Fabric.”
49
See e.g. Lewis Plurality, p. 62.
to the way distances are defined in differential geometry.50 (ii) Adopting

this picture allows one to give a sort of explanation of the triangle
inequality: the length of the shortest path between x and z cannot be
longer than the sum of the lengths of the shortest paths from x to y and
from y to z, because one of the paths joining x and z is the result of first
travelling from x to y along the shortest path, then travelling from y to z
along the shortest path.51
Some who see advantages in taking distance relations to be mediated
nonetheless allow that there are possible worlds in which distance relations
are unmediated.52 But some take the advantages of mediation to be
decisive, denying that there are possible worlds with unmediated distance
relations.53 For authors of this latter rabid tendency the requirement that
geometric relations be mediated involves constraints on what metric spaces
can represent possible spatial structures. At least some path metric spaces
will be allowed. And so will those discrete spaces which can be understood
as equipped with a fundamental notion of adjacency between spatial
atoms, with the distance between two given atoms given by counting
the minimum length of a chain of adjacent atoms that connects the given
atoms.54 But generic discrete spaces—such as the integer lattice Z2
equipped with the metric it inherits from R2 and most finite metric
spaces—will be ineligible, as will spaces of the sort discussed above in
which pairs of points can always be connected by continuous curves, but
50
See Bricker, “Fabric,” §VI; Butterfield, “Against Pointillisme about Geometry,” §3.3.3;
and Nerlich, “Space-Time Substantivalism,” §1.3.
51
See Maudlin, “Buckets of Water and Waves of Space,” p. 196 and Metaphysics
within Physics, §3.2; also p. 329 of Forrest, “Discrete or Continuous?”
52
See Bricker, “Fabric.”
53
See Maudlin, Metaphysics within Physics, §3.2. Nerlich may well be committed to the
same view: at one time he espoused the view that distance relations are mediated as deeply
intuitive and suggested that we cannot understand the thesis that there are direct geometric
relations between distant parts of space (The Shape of Space, §1.3); more recently he has
asserted that the nature of geometric relations does not vary from world to world (“Space-
Time Substantivalism,” §§3.3 and 3.4).
54
Such discrete spaces are discussed in Forrest, “Discrete or Continuous?” The possibility
of such spaces is endorsed by Maudlin (Metaphysics within Physics, p. 89 fn. 6). Nerlich
embraces a related view, on which the size of spatial atoms is allowed to vary and the distance
between two given atoms is given by minimizing the lengths of chains of adjacent atoms
connecting them (Shape of Space, §§9.2–9.6). For an interesting framework that supports a
unified treatment of discrete spaces with a relation of adjacency and of path metric spaces, see
Maudlin, “Time, Topology, and Physical Geometry.”
in which the distance between points is not determined by the lengths of

such curves.
I am not myself moved by arguments of the kind under discussion. We
begin, it seems to me, by regarding both mediated distance relations and
unmediated distance relations as prima facie metaphysical possibilities. Then
we are given reason to think that the latter involve us in a more complex
metaphysics than do the former. This is taken to be reason to think not
only that distance relations are mediated at our world, but that they are
mediated at every possible world—in other words, that unmediated dis-
tance relations are metaphysical impossibilities. But how could simplicity
considerations do that sort of work? Compare with scientific cases. Sup-
pose that we are interested in two prima facie physically possible mechan-
isms known to make identical predictions for a certain class of
observations. We might view simplicity considerations as giving us reason
to believe that one rather than the other of these mechanisms was more
likely to be actual—but I don’t think we would feel any temptation to
conclude that the more complex of the two was physically impossible, just
because it was more complex.
So I am happy to recognize any finite metric space as representing
a possible spatial geometry. And having thus given up on the notion
that distances between points must depend on the lengths of paths
joining them, I am happy to consider allowing as well spaces like that of
Example 3.8 above ( but nothing in what follows will depend on the status
of spaces of this latter kind).
It is less clear to me that we should recognize arbitrary metric spaces as
representing possible spatial geometries. Let Œ be some ginormous cardinal
number. Is there really a possible world whose spatial structure is represented
by a space consisting of Œ points equipped with the discrete metric, or by
space that results from gluing together Œ spaces in some crazy way, and then
by deleting Œ holes? It is a little hard to say. But then nothing in what follows
hangs on whether or not such spaces are allowed.
Where do we stand?
I think that it is clear that we should recognize as representing possible
spatial geometries at least some of the tamer spaces from each of the
categories we have considered (although I sympathize with readers with
qualms about infinite-dimensional Riemannian Hilbert manifolds). I am
in fact happy to count any metric space as representing a possible way space
could be—but in the following chapters I will be as conservative as

possible, relying only on very tame examples in most cases.
But have we gone far enough? Our expedition through the realm of
metric spaces was launched by the observation that if we want to represent
the distance structure of a space by a real-valued function on pairs of
points, then it is hard to resist the conclusion that the structure of space is
represented by a metric space. But there are of course other strategies
available. While I think we can safely proceed below by considering just
metric spaces, I mention a couple of interesting alternative approaches in
the following remarks.
Remark 3.3 (Pointless Metric Spaces). The notion of a pointless metric
space generalizes the notion of a metric space by taking distance relat-
ions to hold between regions rather than points.55 The primitive no-
tions required are: a set R of regions; a partial order # on R
(corresponding to inclusion); a function j j on R taking values [0, 1]
(corresponding to the diameter of a region); and a distance function that
assigns values [0, 1) to pairs of regions. A structure (R,#,j j,) is a pointless
metric space if the following conditions are satisfied for any x, y, z 2 R.
A1 If x $ y, then jxj $ jyj.
A2 If x $ y, then (y, z) $ (z, x).
A3 (x, x) ¼ 0.
A4 If jxj $ jyj and for all z, (z, x) # (y, z), then x $ y.
A5 Generalized triangle inequality: (x, z) # (x, y) þ (y, z) þ jyj.
One way to get a pointless metric space is to begin with a metric space
(X, d ) and then to take the set of regions R to be some set of subsets of X,
partially ordered by inclusion and with the obvious notions of diameter
and distance determined by d.
We call a sequence of {xn } of regions in a pointless metric space a
Cauchy sequence if limn !1 jxn j ¼ 0 and for every > 0 there is an N so that
(xm , xn ) < for m, n > N. We call Cauchy sequences {xn } and {yn } of
regions equivalent if limn! 1 (xn , yn ) ¼ 0. A point in a pointless metric space
is an equivalence class of Cauchy sequences of regions.
Some pointless metric spaces have plenty of points—if we begin with a
metric space, (X, d ) and take R to be the power set of X, then the resulting
55
See Gerla, “Pointless Metric Spaces.”
pointless metric space will have points corresponding to each of the

ordinary points of X.
But other pointless metric spaces have few or no points. For example if
we take R to be the set {( t, t): t 2 Rþ } of open intervals, with the
obvious order and notions of distance and diameter, then the resulting
pointless metric space has only a single point (corresponding to sequences
of regions that shrink to the origin) even though it has uncountably many
regions.
One can construct pointless metric spaces with no points at all. Let R be
the set of closed intervals of the real numbers of the form [m, n] with m, n 2
Z and m # n. Let $ be inclusion and define the diameter and distance in
the obvious way. Then every region has diameter at least one, so there are
no points in this space. We can think of this pointless metric space as being
the sort of discrete space that Nerlich et al. approve of: space consists of
atoms of finite size and the distance between two points is given by
summing the size of the atoms one would have to pass through in
travelling from one point to the other. &
Remark 3.4 (Distances from Discrete Fields). In any treatment of the
distance structure of the Euclidean plane, the real numbers play two
roles. (1) On the one hand, the distance function takes values in R. (2)
On the other, we want the plane to be a continuum, in the sense that if we
introduce coordinates for the plane, these coordinates will take values in
the real numbers and there will be exactly one point in the plane for each
(x, y) 2 R2 .
In the discussion above, we have cheerfully allowed possible spatial
structures to violate (2) but we always held fast to (1). What happens if
we instead generalize by imposing counterparts of both (1) and (2) in
which we allow some other structure to play the role played by the real
numbers in the Euclidean case?
A natural starting point is to consider finite plane geometries whose
points are coordinatized by ordered pairs of elements from a finite field,
and to require distances to take values in the same finite field.56 Of course,
if one proceeds in this way, one will not expect the triangle inequality to
56
A field is a set with well-behaved notions of addition and multiplication. The simplest
finite fields are sets of the form {0,1, . . . , p1} for p prime with addition and multiplication
modulo p.
take its usual form—it will have to be generalized in some way. Some
interesting suggestions along these lines have been made.57 &
4 Plenitude of Possibilities
Anyone who employs a possible worlds framework to make sense of
modal and counterfactual discourse needs to build into this framework
some sort of principle of plenitude: there have to be enough possible
worlds to do the job. David Lewis points out that what we want, intui-
tively, is a principle like the following:
(A) “absolutely every way a world could possibly be is a way that
some world is, and every way that a part of a world could
possibly be is a way that some part of some world is.”58
As Lewis notes, that looks at first glance like it should do the job.
(B) “It seems to mean that the worlds are abundant and logical space
is somehow complete. There are no gaps in logical space, no
vacancies where a world might have been but isn’t. It seems to be
a principle of plenitude.”
But, as Lewis also notes, (A) is not in fact up to the job: the ways a world
could be just are the possible worlds of the framework (and likewise,
mutatis mutandis for parts), so (A) as it stands is trivial. Indeed, it is a matter
of controversy how the notion of plenitude of possibilities should be
understood in general.59
But one can give a partial interpretation of the principle by taking
seriously the metaphor behind the characterization of plenitude offered
in ( B) above. For in some contexts there are perfectly clear mathematical
criteria for the gappiness of a space, and one can understand the principle
57
See Archbold, “A Metric Geometry for Plane Affine Geometry over GF(2n).” For a
related approach, see Kustaanheimo, “On the Relation of Congruence in Finite Geometries”
and “A Note on a Finite Approximation of the Euclidean Plane Geometry.” For further
discussion and references, see Fisher, “Geometry According to Euclid.”
58
Plurality, p. 86; see also p. 2.
59
Lewis introduces his principle of recombination in order to give content to the principle
of plenitude; Plurality, §1.8. For an argument that this strategy is inadequate to Lewis’s needs,
see Divers and Melia, “The Analytic Limit of Genuine Modal Realism.” For other attempts
to give content to the principle of plenitude, see Bricker, “Plenitude” and Hawthorne,
“Causal Structuralism.”
of plenitude as ruling out (at least) the sort of gappiness in question.

Contrast the real numbers with the rational numbers, considered as metric
spaces. The real numbers are gap-free, the rational numbers absolutely full
of gaps. Why? Because the real numbers are complete as a metric space, the
rational numbers incomplete. Every sequence of real numbers that looks
like it should converge converges to a real number. But the fact that a
sequence of rational numbers is a Cauchy sequence is no guarantee that it
converges to a rational number—one can approximate any irrational
number to any desired degree of accuracy by a sequence of rational
numbers. Conversely, one can approximate any rational number arbitrar-
ily accurately by a sequence of irrational numbers. So the rational num-
bers, considered as a metric space, are very gappy (the same is true of the
irrational numbers, of course).
What bearing does this have on our problem? It turns out that relative to
one extremely natural way of making such things precise, the space of
Riemannian manifolds is extremely gappy in much the same way that the
rational numbers are. So if we accept the principle of plenitude and take
seriously the idea that it forbids gaps in logical space, then we cannot take the
Riemannian manifolds to exhaust the possible spatial geometries—we have
to allow, e.g., arbitrary finite graphs to represent possible spatial structures.
For the same sort of reason, a person who was tempted to recognize only
Riemannian manifolds smaller than some given dimension as representing
possible spatial structures would find themselves driven to recognize as well
(at least some) Riemannian manifolds of every finite dimension. The
relevant technical notions and results are surveyed in Appendix B below.
5 Distance Relations?
Suppose that one agrees that all possible spatial geometries are represented
by metric spaces. Does that commit one to any particular view about what
geometric properties and relations are instantiated at various worlds?60
Not in itself. It does not follow from the fact that the spatial geometry of
a world is represented by a metric space that the basic or fundamental
60
Here and throughout, I speak in the idiom of the ontological realism, according to
which there is a fact of the matter concerning a world’s most basic constituents (objects,
properties, relations). On alternative deflationary or anti-realist views, questions like the one
just posed will threaten to collapse—which would tend to benefit the varieties of relation-
alism that I find most attractive.
geometric relations at that world are distance relations. As observed above,

Euclidean space admits a number of axiomatizations in terms of non-
quantitative primitives. One can employ a three-place relation of between-
ness and a four-place relation of congruence.61 Or one can employ a single
three-place relation of equidistance (‘x is as far from y as it is from z’).62
Presumably, corresponding to each axiomatization of Euclidean geometry,
there are worlds at which the primitive notions of that axiomatization
correspond to the fundamental geometric properties and relations. Of
course, at any world with Euclidean structure, the choice of a unit of
measurement induces a notion of distance that satisfies the quantitative
axioms for Euclidean geometry discussed in Section 2 above.
In general, to say that the spatial geometry of a world is represented by a
metric space is to say something along these lines: the geometric properties
of and relations between the points of space at that world determine facts
about distance that are consistent either with a single metric space structure
or with a one-parameter family of such structures differing only as to scale.
On most accounts, a possible world consists of (or represents) some
objects instantiating some properties and relations. What sort of relations?
Some philosophers think that we ought to be able to get by with pure
relations that relate only objects within a world to one another.63 Others
are happy to allow impure relations, countenancing, for instance, relations
between objects in a world and abstract objects such as numbers or
functions.
Those who recognize purity as a virtue will always prefer to take the
fundamental geometric relations at a world to be non-quantitative. This
should be possible so long as one restricts oneself to worlds whose geom-
etries are represented by Riemannian manifolds.64 But one expects things
to become more difficult for wilder sorts of spaces.
Here is a relatively versatile strategy. Given a metric space (X, d ), we
introduce an eight-place relation R on X as follows:
61
Tarski, “What is Elementary Geometry?”
62
See Blumenthal, “Distance Geometries,” §i.1 and Robinson, “Binary Relations as
Primitive Notions in Elementary Geometry,” §2.
63
The locus classicus is §5 of Field, “Can We Dispense with Space-Time?” For further
discussion see Melia, “Weaseling Away the Indispensability Argument,” §7; Mundy, “On
Quantitative Relationist Theories;” and Skow, “Are Shapes Intrinsic?”
64
See the main result in Mundy, “Space-Time and Isomorphism.”
R(w1 , w2 , x1 , x2 , y1 , y2 , z1 , z2 )()jd(w1 , w2 ) d(x1 , x2 )j$jd(y1 , y2 )

d(z1 , z2 )j:
We call (X, d ) replete if whenever R(w1 , w2 , x1 , x2 , y1 , y2 , z1 , z2 ) there

exist u1 , u2 2 X such that:
jd(w1 , w2 ) d(u1 , u2 )j ¼ jd(y1 , y2 ) d(z1 , z2 )j:
It turns out that if (X, d ) is replete then R fixes d up to a scale factor.65 So

within the class of worlds whose geometry can be represented by a family
of replete metric spaces differing only as to a choice of scale, purists can
make do with a single primitive eight-place geometric relation.
But this approach won’t work for all sorts of worlds one is likely to want
to allow—it is easy, for instance, to construct finite metric spaces that are
not replete in the present sense.66
Any purist who wants to recognize such spaces will be driven to
desperate measures, such as postulating a one-parameter family, {Rs },
with each Rs doing the work of the impure notion ‘the distance between
x and y is s.’ I will leave it to purists to decide: (i) whether this sort of
purity-in-name-only is good enough; and (ii) whether having had to
stoop this low at some worlds, they still think it worthwhile to work
hard to retain their purity at others.67
Finally, consider the problem of scale. If I tell you that a given metric
space represents the spatial geometry of two possible worlds, have I told
you that the worlds share exactly the same geometry, or merely that their
geometry is the same up to a choice of size factor? Are there worlds whose
spatial geometry is represented by a two-point metric space or by the
intrinsic geometry of a spherical surface in Euclidean space, but which
differ as to how far apart it is possible for objects to be?
65
i.e., given R arising in this way, one can find a distance function d on X such that R(w1,
w2, x1, x2, y1, y2, z1, z2) if and only if jd(w1, w2) d(x1, x2)j $ jd(y1, y2) d(z1, z2)j—and any
two such distance functions differ only by a scale factor. See Krantz et al., Foundations of
Measurement, vol. i, ch. 4, theorem 6.
66
Consider a metric space with just three points, x, y, and z. Label the distances a ¼ d(x, y),
b ¼ d(y, z) and c ¼ d(x, z). Assume that a $ b $ c and take c ¼ 1. There are only four ways of
fixing the values of a and b so that the resulting metric space is replete: (i) a ¼ b ¼ 1; (ii) a ¼ b ¼
2; (iii) a ¼ 2 and b ¼ 1; (iv) a ¼ 3 and b ¼ 2.
67
On (i), cf. Field, “Can We?,” §7; Mundy, “Quantitative Relationist Theories,” §3; and
Melia, “Weaseling,” §7.
For my own part, I am tempted to say that there are worlds that differ in
this way. If we suppose that we live in a world at which space is expanding
and spherical in geometry, doesn’t it then make sense to consider two
worlds in which space is spherical but of constant volume—one in which
space is the size our space was in olden times, one in which space is the size
it is now?
But this issue is not an important one for present purposes, and it is one
on which philosophers tend to divide, since it is one for which overarching
views about the metaphysics and epistemology of modality can have
consequences. I will typically leave it to the reader to fill in qualifications
such as ‘up to a choice of scale, if you believe that makes any sense.’
II
Spatial Structure for
Relationalists
The more universal a concept is the more frequently it enters into

our thinking; and the more indirect its relation to sense-experience,
the more difficult it is for us to comprehend its meaning; this is
particularly the case with pre-scientific concepts that we have been
accustomed to use since childhood. Consider the concepts referred to
in the words “where,” “when,” “why,” “being,” to the elucidation of
which innumerable volumes of philosophy have been devoted. We
fare no better in our speculations than a fish which should strive to
become clear as to what is water.
Einstein
1 Substantivalism and the Structure of Space

Consider sentences attributing geometric features to space—‘space is finite,’
‘space is three-dimensional,’ ‘space is continuous,’ etc. Our problem in this
chapter is to consider how relationalists can make sense of such sentences.
The situation is straightforward for substantivalists. They take space to
consist of parts standing in geometric relations to one another. So it is
natural for them to take the truth conditions for the sort of sentences we
are interested in to be conditions on the pattern of instantiation of
geometric relations by the parts of space.
It will be helpful to have some examples in play. To keep things simple,
let us focus on substantivalist worlds where the structure of space does not
change over time. Choosing a unit of measurement if necessary, we can
take the points of space at such a world to form a metric space (X,d ). Then
the following criteria provide something like canonical truth conditions
for various claims about the size of space.
36 spatial structure for relationalists
Finitude. Space is finite in extent if and only if there is an upper

bound on the distance between its points (i.e., there is a real number a
such that for any points x, y 2 X, d(x, y) # a).
Indefiniteness. Space is indefinite in extent if and only if there is no
upper bound on the distance between its points (i.e., for every real
number a, there is a pair of points x, y 2 X such that d(x, y) $ a).
Infinitude. Space is infinite in extent if and only if there exists a
set of collinear points of space separated by arbitrarily large distances
(i.e., there is a set of points Y X such that: (i) any three points
from Y can be assigned labels y1 , y2 , and y3 in such a way that
d(y1 , y2 ) þ d(y2 , y3 ) ¼ d(y1 , y3 ); and (ii) for every real number a,
there are points y1 , y2 2 Y such that d(y1 , y2 )$a).1
Indefiniteness is incompatible with finitude. Infinitude implies indefinite-
ness but not vice versa.2
The same sort of thing can be done for dimension, but in this case the
criterion is a bit more involved.3 For a > 0 we call a set of points equilateral
with side-length a if any two distinct points in the set are separated by exactly
a units of distance. By way of motivation, note that in two-dimensional
Euclidean space, one can find equilateral sets containing three points (the
vertices of an equilateral triangle), but no four-point equilateral sets; more
generally, in a Euclidean space of dimension n, the largest equilateral sets
that one can find contain n þ 1 points.
Dimension. Space is n-dimensional if and only if: for any point of x
of the space, there exists a > 0 such that for any < , x is a member
of an (n þ 1)-member equilateral set of side-length , but is not a
member of any (n þ 2)-member equilateral set of side-length .4
1
Clause (i) is not a good definition of collinearity in general (think of a line in spherical
geometry). But it is adequate when combined with (ii).
2
Let U be the subset of the positive quadrant of the x-y plane bounded by but not
including: the x-axis, the y-axis, and the curve y ¼ 1x. The extent of U indefinite but not
infinite.
3
The criterion below is adequate for worlds whose spatial geometry is represented by a
Riemannian manifold; see §3 of Berger, “Une caractérisation purement métrique des variétés
riemanniennes à courbure constante.”
4
Why so complicated? Because although in an n-dimensional Euclidean or hyperbolic
space the largest equilateral set of any side-length has n þ 1 members, it is possible to find an
(n þ 2)-member equilateral set in an n-dimensional spherical space for one special side-length
(think of the regular tetrahedron inscribed in a two-dimensional sphere). The situation is
similar ( but weirder) in the elliptic case; see Example 4.2 of Ch. III below. But if we restrict
attention to sufficiently small side-lengths, the unwanted equilateral sets are excluded.
spatial structure for relationalists 37
Remark 1.1 (Truth Conditions). I take the notion of a truth condition to be

stronger than the notion of a necessary and sufficient condition: let C be a
necessary and sufficient condition for a sentence S to be true; C is a truth
condition for S to the extent that it plays a role in constituting the meaning
of S. This notion is not perfectly clear—but something like it is in
common use and I rely on it here.
The standard example used to drive home the need for some such
distinction is as follows: let S be ‘Firenze è una bella città,’ C1 be the
condition that Florence is a beautiful city, and C2 be the condition that
Florence is a beautiful city and that first-order Peano arithmetic is incom-
plete; anyone who takes there to be a tight connection between under-
standing and grasp of truth conditions will want to count C1 but not C2 as
a truth condition for S.5
For present purposes, the point of invoking this distinction is as follows:
‘space is finite in extent’ is true if and only either of the following two
conditions obtains: (i) there is an upper bound on distance between points;
(ii) space is not Euclidean, nor hyperbolic, nor . . . 6 I take (i) to be in the
neighbourhood of a truth condition for the claim that space is finite; and
I take it that relationalists are faced with the problem of finding a reason-
able surrogate for (i). Finding a surrogate for (ii) would hardly be satisfac-
tory. Indeed, here is one very unsatisfying way to achieve a relationalist
surrogate for (ii): take as primitive the notion of being a Euclidean point,
the notion of being a hyperbolic point, etc.; then take space to be finite at a
relationalist world if and only if the material points at that world are not
Euclidean, nor hyperbolic, nor . . . &
2 Conservative Relationalism
The challenge facing relationalists is to provide surrogates for the canonical
substantivalist truth conditions for attributions of geometric structure to
space. Broadly speaking, relationalist approaches to this problem fall under
two main headings. Modal approaches appeal to possible but non-actual
configurations of matter in cashing out claims about spatial structure.
5
For discussion and references, see e.g. Higginbotham, “Truth and Understanding” and
Soames, “Truth, Meaning, and Understanding.”
6
i.e., this second condition rules out one by one all of the spatial structures that are not
finite in extent.
Conservative approaches make do with actual material configurations. In

essence: on modal approaches, the full structure of space is (in general)
revealed only if one considers possible as well as actual configurations of
matter, while on conservative approaches, the structure of space is fully
revealed by the actual geometric relations between material bodies.7
Our main interest will be in modal approaches. In order to motivate this
interest and to introduce some ideas that will feature in subsequent
discussion, it is worthwhile considering the prospects and resources of a
few representative conservative approaches.
Conservative relationalism is at its most straightforward and appealing in
application to worlds at which the pattern of geometric relations instan-
tiated by matter—the material geometry—does not change over time. For in
this case, the conservative principle that the structure of space is fully
revealed by actual geometric relations dictates that we should identify
the unchanging material geometry of the world with its spatial geom-
etry—i.e., we take space to be finite if and only if there is an upper bound
on the distance between material points, and so on.
This was the context that Aristotle and Descartes worked in and both
were, I believe, conservative relationalists of this type—at any rate, neither
seems to have felt any need for, or shown any interest in, a notion of
geometric possibility.8 In an Aristotelian world, matter always fills a sphere
of unvarying size—so space can be taken to have the structure of a solid
ball in Euclidean space. By appeal to versions of the substantivalist criteria
discussed above, adapted so that material points rather than points of space
are quantified over, we are entitled to say that in such a world space is
three-dimensional and finite in extent. In a Cartesian world, the material
geometry at each instant is that of three-dimensional Euclidean space—so
the conservative relationalist is entitled to say that space is three-
dimensional, infinite in extent, and so on.9
7
Admittedly, the connection between the official characterization of conservative ap-
proaches and this informal gloss is somewhat loose—but this should not undermine the
discussion below, the point of which is to get clear about the shortcomings of several non-
modal approaches.
8
In addition to taking the material geometry to be time-independent, both Aristotle and
Descartes took the postulation of void or empty space to involve one in serious conceptual
difficulties. See e.g. bk. iv of Aristotle’s Physics and §§ii.16 ff. of Descartes’s Principles of
Philosophy.
9
Descartes himself was famously cagey about the extent of material extension; see e.g.
Principles of Philosophy, §§I.26 f. But it is hard to see what else he could have had in mind.
However, the situation becomes more puzzling when one consider

worlds in which the material geometry does change over time. It will be
helpful to have a family of examples in mind. Let us call a world quasi-
Aristotelian if at each time t matter forms a ball of radius r(t). We call a
quasi-Aristotelian world expanding if: (i) r (t) is defined for 0 < t < 1; (ii)
limt ! 0r (t) ¼ 0; and (iii) r (t) is a strictly increasing function of t (so that
t1 < t2 implies r(t1 ) < r(t2 ). We say that the expansion of such a world is
bounded if limt ! 1r(t) ¼ R < 1; otherwise we say that it is boundless.
What should conservative relationalists say about the extent of space in a
boundlessly expanding quasi-Aristotelian world? Three possibilities come
readily to mind: one could say that space is finite in extent at such a world,
that it is indefinite but not infinite in extent, or that it is infinite in extent.
Below I consider some representative species of conservative relationalism
corresponding to these three possibilities, each of which turns out to have
very unattractive features.
First Gambit
Intuitively, to say that space is finite it to say that there is some upper
bound on distances. In a boundlessly expanding quasi-Aristotelian world
there are material points separated by arbitrarily large distances (at sufficiently
late times).
So how could conservative relationalists maintain that space is finite at
such worlds? Two strategies come to mind, neither very attractive. Each
turns on the observation that in a quasi-Aristotelian world, for each time
there is an upper bound on the distance between material points at that
time.
On the one hand, we might resort to a kind of supervaluation, counting
a claim about the spatial geometry of a world as true just if it is true (under
its ordinary reading) of each of the instantaneous material geometries of
that world. For instance, ‘there is an upper bound on distances at quasi-
Aristotelian world w’ would be true in virtue of the fact that for each time
at w, there is an upper bound on the distance between points at that time.
But this is hardly satisfactory, since it will of course make a hash of the usual
logical relations between geometric claims—e.g., in the case of a bound-
lessly expanding quasi-Aristotelian world w, we will have to say that for
every real number a it is indeterminate whether a is an upper bound on
distance at w, even while maintaining that there exists such an upper
bound. It is highly dubious that the set of claims that come out as true at
a such world under this approach can be coherently viewed as characterizing

the geometric structure of a world.
On the other hand, we might relativize claims about spatial geometry to
moments of time, viewing the structure of space as changing whenever the
pattern of geometric relations instantiated by material bodies is altered. So
we would say that the geometry of space is constantly changing (but is
always finite in extent) in any expanding quasi-Aristotelian world. This is a
perfectly coherent view. But it is completely lacking in motivation.
Conservative relationalists are in some sense committed to thinking of
the structure of space as being fully revealed by the distance relations
between parts of matter. But one would normally shy away from assuming
that this structure is fully revealed at each instant. It seems more natural to
think that when matter is sparse, the structure of space may be revealed
only in the course of time via the motions of matter. Consider, for
instance, a world at which two material particles initially overlap, then
become separated by larger and larger distances. Without exception,
contemporary relationalists seem to accept without question that at such
a world space has the structure of a continuum (and that this structure is
time-independent). But on the view under consideration, at such a world
space always has exactly two points and is constantly changing its geometry.
Second Gambit
Suppose, then, that we require that the structure of space be considered a
time-independent feature at simple worlds. Fix for consideration a bound-
lessly expanding quasi-Aristotelian world w. What should we say about the
extent of space at w?
Clearly space is not finite at w. For if space were finite at w, there would
be an upper bound on how far apart things can be—but any such bound
would be violated by material points in w at sufficiently late times.
One intuitive view has it that space is at least indefinite in extent at w
(since for any real number a, we can find a time t and a pair of material
points of w at t whose distance from one another is greater than a) but that
it would be excessive to say that space is infinite at w. For in order for space
to be infinite, we must be able to find an unbounded set of collinear
points. But we cannot find such a set in w: distance relations between
points obtain only for points drawn from a single time; and in w there is for
each time an upper bound on distances between points at that time.
These considerations motivate the following scheme for adapting sub-

stantivalist criteria to conservative needs: if the substantivalist criterion for a
given geometric feature requires that a certain sort of set of distance
relations be instantiated, we will count space at w as having that feature
if and only if we can find material points at w satisfying a suitable set of
distance relations between material points—without requiring that there
be any one time at which all of the required relations obtain. The most
obvious way to implement this strategy is to translate substantivalist criteria
by replacing existential quantifiers over points of space with combinations
of existential quantifiers over times and existential quantifiers over material
points (and likewise for universal quantifiers).
Under this approach, the criteria for the extent of space become:
Finitude. There is an a 2 R such that at every time, every pair of
material points at that time is separated by a distance less than a.10
Indefiniteness. For every a 2 R there exists a time and a pair of
material points at that time separated by at least a units of distance.
Infinitude. There exists a time and a set of collinear material points at
that time such that for every a 2 R there are points in the set whose
separation is at least a.
As desired, the second of these translated criteria holds in our boundlessly
expanding quasi-Aristotelian world, while the first and the third do not.
Further, the approach has the advantage that it is entirely true to the
conservative spirit—the distance relations instantiated by matter fully
reveal the structure of space—while allowing that there may not be any
instant at which this structure is fully displayed.
But the proffered translation scheme has some very awkward features.
Most scandalously: it dramatically distorts the logical relations between the
items translated (and hence has rather weak credentials to be considered a
scheme of translation in the first place). In order to see this, consider a one-
dimensional boundlessly expanding quasi-Aristotelian world w (so at each
instant the material geometry of w is corresponds to the geometry of a
10
Note that it would be a disaster to instead demand that there be a real number a such
that for any time and for any material point at that time and for any other time and for any
material point at that second time, a is greater than the distance between the two points. For
in general cross-time comparisons of distance may be impossible (so that the condition under
consideration would fail)—and this could happen even at Aristotelian worlds that conserva-
tive relationalists will certainly want to count as being finite in extent.
bounded interval of the real line). Call a metric space (X, d ) uni-dimensional
if any three points in X can be labelled y1 , y2 , and y3 with
d(y1 , y3 ) ¼ d(y1 , y2 ) þ d(y2 , y3 ). There is an obvious substantivalist notion
of uni-dimensionality applicable to spatial geometries—and for substan-
tivalists, if the spatial geometry of a world is both uni-dimensional and
indefinite in extent, then it is also infinite in extent. But, sadly, our
translation scheme commits us to saying that the geometry of space at w
is uni-dimensional. And this leads to disaster when combined with our
earlier conclusion that this geometry is indefinite but not infinite in extent.
Third Gambit
Consider again a one-dimensional expanding quasi-Aristotelian world.
The instantaneous configurations of matter can be identified with intervals
of the real line, with length approaching zero as t ! 0 and approaching 0
< R # 1 as t ! 1. The expansion is bounded if R < 1, boundless if R
¼ 1. In either case, it is intuitively plausible that space should have the
metric structure of an interval of the real line of length R—so that space is
infinite at such a world if and only if expansion is boundless.
The gambits just considered were unable to do justice to this intuition—
either they provided no coherent account of the structure of space or they
required us to view that structure as time-dependent. So let us go in search
of a strategy that: (i) assigns to space a determinate time-independent metric
structure that is explored by the dynamical evolution of the instantaneous
material geometries; and (ii) stays as true as possible to the conservative spirit.
We are set the following task.
Selection Problem. Given the set {Xt } of material geometries that
occur at world w, select the metric geometry X that encodes the
geometry of space at w.
Any relationalist will recognize the following as a constraint.
Embeddability. The abstract space X representing the structure of
space at a world must contain, for each instant of time t, a region
isometric to Xt .
On its own, this latter condition is of course far too weak to single out a
solution to the Selection Problem. For note that any one-point metric
space is embeddable in any (non-empty) metric space whatsoever—so
Embeddability alone puts no substantive constraints on solutions to the
Selection Problem for one-particle relationalist worlds. Less dramatic ver-

sions of this problem arise even for worlds with rich and constantly changing
instantaneous material geometries: for example, there exist non-isometric
spaces that each contain a copy of every Riemannian manifold.11
But we are aiming for a conservative account. It is mandatory for
conservatives to maintain that the metric space that we choose to represent
the structure of space at a world should always be the smallest metric space
compatible with the instantaneous material geometries instantiated at that
world. For if this condition fails, then either (a) we are employing a metric
space larger than we need to; or (b) there are multiple otherwise acceptable
candidate metric spaces, none smaller than the the others. In either case,
there is a clear sense in which the structure of space is not being fully
revealed by actual geometric relations. In case (a), a part of space remains
unexplored by matter. In case (b), there will be a geometric feature that
one candidate possesses but which another lacks—and so the geometric
relations actually instantiated by matter will fail to determine whether or
not this feature is possessed by the world’s spatial geometry. So let us
impose the following desideratum.
Minimality. X is minimal among spaces satisfying Embeddability, in
the sense that if X * also satisfies Embeddability, then X * must have a
subspace isometric to X.
Unfortunately, except in certain trivial cases, Minimality is incompatible
with our other desiderata. Consider a boundlessly expanding quasi-Aris-
totelian world w. The real numbers, R, and the positive real numbers, Rþ ,
both satisfy Embeddability. But since Rþ can be viewed as a subspace of R
but not vice versa, Minimality eliminates the real numbers from the
running (thereby disappointing fans of symmetry). Do the positive reals
satisfy Minimality? No. Take each of the metric spaces corresponding to
each of the instantaneous material configurations of w, and glue them
together by their central points to create a horrible metric space Y. The
positive real numbers are not embeddable in Y, since Y does not contain
any uni-dimensional subset with no upper bound between its elements. So
the positive real numbers do not satisfy Minimality for w. Nor, of course, is
Y minimal: it cannot be embedded isometrically in the real numbers since
it is indefinite but not infinite in extent and no subset of the real numbers
11
See Remark 2.1 below.
has this feature. Nor, indeed, can any space satisfy both Embeddability and
Minimality for w. For let Z be a space that, for each length, contains an
isometric copy of a finite interval of real numbers of that length. If Z were
minimal, then it would be embeddable in both R and in Y. But if Z is
embeddable in R and contains copies of intervals of arbitrary length, Z
must be uni-dimensional and unbounded—but as noted above, this is
incompatible with being embeddable in Y.
This is the end of the line for the our third gambit, conceived of as
implementing the strict conservative impulse. But it is natural (and instruct-
ive) to wonder what happens if we stray a little bit from the conservative fold.
In the special case of our one-dimensional expanding quasi-Aristotelian
world, we feel sure we know what space looks like: like a real interval of
length R, where R is the finite or infinite limit of the lengths attained by
the material configurations as t ! 1. More generally, let us call a world
cumulatively expanding if Xt is isometrically embeddable in Xt0 whenever
t < t 0 . Intuitively, at a cumulatively expanding world, there is a natural
candidate to represent the structure of space: the Gromov–Hausdorff limit
of the sequence {X0 , X1 , . . . }.12
Cumulative Limit. If w is cumulatively expanding, then X should be
a Gromov–Hausdorff limit of {Xn }, when one exists.
This delivers the desired solution to the Selection Problem at expanding
quasi-Aristotelian worlds.
Now, clearly, this doesn’t get us all that far: the class of cumulatively
expanding worlds is a very special one and it is not clear how to generalize
the strategy behind Cumulative Limit to more general classes of worlds.13
12
See Appendix B below for the relevant technical notions.
13
(1) Note that outside of the class of cumulatively expanding worlds, one can find that a
sequence of metric space converges to a limit that doesn’t contain isometric copies of any of
the metric spaces in the sequence; e.g., as discussed in Appendix B below, a sequence of two-
dimensional Riemannian spaces can converge to a one-dimensional object (a finite graph). (2)
The notion of an ultralimit of a set of metric spaces generalizes the notion for the Gromov–
Hausdorff limit of a sequence (and so shares many of its flaws). An attractive feature of the
ultralimit construction is that it returns a metric space X whenever fed an arbitrary collection
{Xi}i 2 I of metric spaces (so we can feed it the full set of material geometries that occur at a
world). But don’t expect anything for free: the construction requires auxiliary input (the
choice of a non-principal ultrafilter on the index set I) upon which the output will in general
depend. For the ultralimit construction, see Gromov, Metric Structures for Riemannian and Non-
Riemannian Spaces, §3.29; or Kapovich, Hyperbolic Manifolds and Discrete Groups, ch. 9. For
further discussion and examples, see Sormani, “How Riemannian Manifolds Converge,” §3.4.
Worse, problems arise even for the simple case of one-dimensional

expanding quasi-Aristotelian worlds. Consider first the case of a world
w with bounded expansion. Our apparatus tells us to take space to have the
structure of an interval of real numbers of length R. But which one?
Should we take space to have the structure of the closed interval [0, R]
that includes both its endpoints, of the open interval (0, R) that includes
neither endpoint, or a half-open interval that includes one endpoint but
not the other? Our apparatus doesn’t tell us: it is a peculiarity of the
Gromov–Hausdorff limit that it is blind to these sorts of minute differ-
ences.14 And because our world is permanently expanding, each of the
material geometries is embeddable into each of these three options (so
Embeddability does no work here). At this point we could of course
invoke further considerations. Presumably, we will be tempted to opt
for the open interval on the grounds of parsimony. But note that other
considerations may pull us in other directions: the open interval and the
closed interval manifest a symmetry absent from the half-open interval.
This sort of symmetry might or might not be a symmetry of the material
configurations at the world under consideration. Consider two cases.
(i) Each of the instantaneous material configurations has the structure of
a closed interval. Then considerations of symmetry suggest that space
should either be an open interval or a closed one—presumably we will
choose the former option, since it is more parsimonious. (ii) Each of the
instantaneous material geometries has the structure of a half-open interval.
Since the data are asymmetric, this seems to provide a reason to take space
to be asymmetric—even though this means taking space to be larger than
it strictly speaking has to be.
Consider next what happens in the boundlessly expanding case. Here
the Gromov–Hausdorff apparatus requires as input the choice of a distin-
guished point in each interval—and depending on how we choose these
distinguished points, the limit can be either the real line or the half-line
(which again may or may not contain its initial point). Here again we see a
clash between considerations of size and considerations of symmetry: on
the one hand, we want to posit the minimal possible space—this militates
in favour of taking space to be the (open) half-line; on the other hand, we
14
The Gromov–Hausdorff distance between our three options is zero, so each may be
viewed as the limit of any sequence that converges to any of them. The Gromov–Hausdorff
notion of convergence guarantees only that there will be a unique complete limit of any
convergent sequence.
are loath to take space to be asymmetric when the data are symmetric—so
if the instantaneous material geometries have the structure of open or
closed intervals, considerations of symmetry militate in favour of taking
space to have the structure of the full line.
The upshot is this. In order to construct a selection procedure that
applies to classes of worlds beyond the very simplest cases, one has to begin
balancing parsimony against other desirable features such as symmetry.15
That takes us very far from the conservative impulse that is our present
concern. We will see in the next chapter where exactly it leads.
Remark 2.1 (Universal Metric Spaces) Let k be a class of metric spaces.
We say that a metric space X is universal for k if it contains, for each metric
space Y 2 k, a subspace isometric to Y.
Of course, there can be no metric space universal for the class of all
metric spaces, since there is no upper bound on the cardinality of metric
spaces.16 But there are universal metric spaces corresponding to many
interesting classes of metric spaces.
Consider ‘1 , the set of all bounded sequences of real numbers equipped
with the metric:
d({xn }, {yn }) ¼ sup jxk yk j:

k2N
(the right hand side is the least upper bound on the numbers jxk yk j as k
ranges over the natural numbers).17 Let (Z, dZ ) be a metric space with
countably many points, z0 , z1 , . . . Then the map ç: Z ! ‘1 defined by
z7!{dZ (z, z0 ), dZ (z, z1 ) dZ (z1 , z0 ), . . . , dZ (z, zn ) dZ (zn , z0 ), . . . }:
is an isometry. So ‘1 is a universal metric space for the class of countable

metric spaces.18
A metric space X is said to be separable if it contains a countable subspace
Y that is dense, in the sense that any open metric ball around any point of X
includes at least one point of Y. Many uncountable metric spaces are
15
This appears to be a general feature of the sort of problem that we are considering here.
See Remark 2.2 below.
16
For any cardinal number Œ, one can put the discrete metric on a set of cardinality Œ.
17
A sequence {xn} of real numbers is called bounded if there is an N 2 R such that xk < N
for all k.
18
See e.g. §1.22 of Heinonen, Geometric Embeddings of Metric Spaces.
separable—for example, every Riemannian manifold is separable.19 Any

complete metric space that is universal for the class of countable metric
spaces is also universal for the much larger class of separable metric spaces.20
The space ‘1 is complete, and so universal for separable metric spaces.
While ‘1 is not itself separable, there also exist separable metric spaces that
are universal for the class of separable metric spaces. One such space is the
set C[0,1] of all continuous real-valued functions on the closed unit
interval equipped with the metric d( f , g): ¼ sup0#x#1 j f (x) g(x)j.21
Another example is provided by the mind-boggling Urysohn space. This is
the unique separable metric space U such that: (a) U is universal for
separable metric spaces; and (b) U has the feature that any isometry between
finite subsets of U can be extended to an isometry from U to itself.22
There are a number of ways of characterizing or constructing U.23 One
particularly beguiling characterization is as follows. Note that if we have a
countably infinite set of points x1 , x2 , . . . , then we can encode distance
relations between them via a countably infinite square matrix—with entry
dij corresponding to d(xi , xj ). The Urysohn space is the random metric space,
in the sense that generic such matrices encode the distance relations among
the points of a countable dense subset of the Urysohn space.24 &
Remark 2.2 (Embedding Problems). The Selection Problem considered
above takes the following form: among all metric spaces containing copies
of some given set of metric spaces, select an especially nice one—that is
then to be viewed as representing the structure of space at a world where
the given metric spaces encode the instantaneous material geometries.
19
More generally, a Riemannian Hilbert manifold is separable if and only if it is modelled
on a Hilbert space of countably many dimensions.
20
Given a separable metric space X, find an embedding of one of its countable dense
subsets into the universal space, then extend this embedding to all of X by continuity.
21
See e.g. Heinonen, Geometric Embeddings, theorem 3.6; or Bessaga and Pełczy nski,
Selected Topics in Infinite-Dimensional Topology, §II.1.
22
More explicitly, condition (b) says: for any finite subsets {x1, . . . , xn}, {y1, . . . , yn} U
such that d(xi, xj) ¼ d(yi, yj), 1 # i, j # n, there is an isometry ç : U ! U such that yi ¼ ç(xi),
for 1 # i # n.
23
For discussion and references, see Hušek, “Urysohn Universal Space, its Development
and Hausdorff ’s Approach” and Hubička and Nešetřil, “A Finite Presentation of the Rational
Urysohn Space.”
24
See Vershik, “Random Metric Spaces and Universality.” Note that despite its many
strange properties, the Urysohn space is well behaved topologically; see Uspenskij, “The
Urysohn Universal Metric Space is Homeomorphic to a Hilbert Space.”
Above we saw that even in relatively simple cases considerations of

parsimony and considerations of symmetry can pull in different directions.
This sort of conflict is in fact a general feature of problems of this
sort.25
Suppose that one is given the family f of all finite linear orders and
asked to find a countable structure S whose finite substructures are given
by f. This tells us that S is a countable linearly ordered set. There are
many such structures. For example, S could have either the order structure
of the natural numbers or the order structure of the rational numbers.
There is a sense in which the former is ‘smaller’ than the latter (the latter
contains a copy of the former, but not vice versa), while the latter is more
highly symmetric than the former (while there are no non-trivial sym-
metries of the order structure of the natural numbers, there are many
symmetries of the order structure of the rational numbers). In fact, one can
show that the order structure of the rational numbers is maximal among
the possible structures for S (any other candidate is isomorphic to a subset
of it) and that it is maximally symmetric (it is the only structure for S with
the feature that any isomorphism between finite subsets can be extended
to an isomorphism of the entire structure).
More generally, suppose that we have a countable set r of relations. An
r-structure is a set of objects structured by (all or some) of the relations in
r. A set f of finite r-structures is called an age if it coincides (up to
isomorphism) with the set of finite substructures of some countable
r-structure S.26 We then say that f is the age of S. The same family f
may be the age of many different r-structures (as in the example of the
preceding paragraph). When f has an additional property (the amalgam-
ation property), then there is (up to isomorphism) a distinguished countable
r-structure S, the Fraïssé limit of f, such that: (i) f is the age of S; and (ii)
any isomorphism between finite subsets of S can be extended to an
isomorphism from S to itself.27 Further, if S* is a countable r-structure
and f is the age of S * , then S * is isomorphic to a substructure of the
25
For what follows, see e.g. Hodges, A Shorter Model Theory, §6.1.
26
In order for f to be an age, it is necessary and sufficient that it be closed under taking
substructures and that for any two members of f, there is a third that contains isomorphic
copies of each of them.
27
Intuitively speaking, f has the amalgamation property if: whenever A, B, and C are
structures in f such that B and C each include a copy of A, then f also includes the structure
that results if one glues B and C together by identifying points corresponding to the same
point in A.
Fraïssé limit of f. So the most symmetric structure whose finite substruc-

tures are given by f is also the largest such structure. Here again, we find a
conflict between parsimony and symmetry.
Finally, note that the Urysohn space discussed in the preceding remark
can be thought of as the Fraïssé limit of the age consisting of all finite metric
spaces. More precisely, the picture is as follows.28 Let us take our set of
relations r to be indexed by the rational numbers. Then any metric space
can be viewed as an r-structure: we encode facts about distances via the
rule Rq(x, y) () d(x, y) < q. Consider the class f of all finite metric spaces
with rational distances. f is an age, and the corresponding Fraïssé limit can
be identified with a countable dense subset of the Urysohn space.29 &
3 Modal Relationalism
Substantivalists are able to ground their attributions of geometric structure to
space in criteria that concern the patterns of distance relations instantiated by
the points of space. Conservative relationalists work hard to provide substi-
tutes for these criteria directly in terms of the distance relations instantiated
by parts of matter. But it seems that even relatively simple worlds in which
the geometry instantiated by matter changes over time provide serious
obstacles for the most straightforward varieties of conservative relationalism.
The obvious alternative is to go modal—to take over versions of the
substantivalist criteria, understood now as asserting not that a certain
pattern of relations is instantiated but that the instantiation of such a
pattern is possible relative to the world in question.30 Thus, where sub-
stantivalists take space to be infinite in extent at a world if and only if there
exists a linear, unbounded set of points of space of that world, modal
relationalists take space to be infinite in extent at a world if and only if it is
possible relative to that world for there to exist a linear, unbounded set of
material points.
28
See Kechris et al., “Fraïssé Limits, Ramsey Theory, and Topological Dynamics of
Automorphism Groups,” §§2 and 6(E).
29
Aside: in general, the set of finite subspaces of a metric space fails to determine the
metric space (the real numbers and the positive real numbers have the same finite subspaces);
but some metric spaces are characterized up to isomorphism by their finite subsets. See
Bogatyi, “Metrically Homogeneous Spaces,” propositions 1.11 and remark 1.12.
30
Traditionally, this modal path is motivated by snippets from Leibniz. The question
whether Leibniz was in fact a modal relationalist will be left for Appendix D below. (He was.)
One question that immediately arises is what sort of possibility is

involved in such claims. It would appear that it cannot be logical or
metaphysical possibility, since these notions are not world-relative in any
non-trivial sense. Even the most conservative of relationalists should be
happy to concede that a linear, unbounded set of material points is
logically and metaphysically possible tout court, and hence also that they
are logically or metaphysically possible relative to, say, an Aristotelian
world.
Some will be tempted to take the relevant modality to be physical
possibility. But this is a temptation that should be resisted. (i) Consider
the physical theory that says that spacetime is flat, that the electromagnetic
field obeys Maxwell’s equations, and that charged bodies obey the Lorentz
force law. This theory doesn’t say what the global structure of spacetime is:
in particular, it allows space to be either Euclidean (and hence to be
infinite in extent) or to have the topology of a three-torus (and hence to
be finite in extent). At any world allowed by this theory, it is physically
possible that there should be an infinite array of bodies (because such arrays
exist, e.g., at some of the Euclidean worlds allowed by the theory). But
modal relationalists will want to deny that such arrays are geometrically
possible at those worlds with toroidal spatial geometry that are permitted
by our theory. So in order to make sense of theories that allow distinct
spatial background geometries, relationalists will need to rely on a notion
of geometric possibility distinct from physical possibility. (ii) Some sub-
stantivalists and some relationalists think that it would make sense to say
space was infinite, even if we thought it physically impossible for matter to
spread beyond a certain finite domain, or that space was infinite, even if
the presence of infinitely many particles would cause it to roll up into a
finite space.31 And, presumably, in order to make sense of such claims,
relationalists must employ a notion of geometric possibility distinct from
physical possibility.
Relationalists who adopt a modal account of geometric structure will be
(and should be) pressed to say more about this new brand of modality. In
particular, relationalists invoking a notion of geometric possibility will be
confronted with the following choice: take this notion as primitive or put
31
For discussion and references, see Brighouse, “Incongruent Counterparts and Modal
Relationalism,” §3; see also Butterfield, “Relationism and Possible Worlds,” §4. Note that
the view in question is a contentious one; see fn. 8 of Ch. I above.
forward some account of how facts about geometric possibility at a world

are grounded in the geometric properties and relations instantiated at that
world.
Of course, anyone invoking a distinctive brand of modality faces a
similar choice. Other things being equal, one would prefer to provide a
grounding if possible (if for no other reason, then out of a taste for
elegance). Nonetheless, primitivism about logical, physical, and metaphys-
ical possibility each have their defenders. In the case of geometric possi-
bility, though, it seems to be generally supposed that relationalists are
entitled to this notion only if they can ground it in more basic geometric
facts.32 Perhaps this is in part because substantivalists have ready to hand a
means of grounding claims about geometric possibility in facts about the
geometric relations between points of space: a configuration of material
points is geometrically possible relative to a world if and only if it instan-
tiates a pattern of distance relations actually instantiated by some points of
space of that world.33
Let it be agreed, then, that a means of grounding geometric possibility in
occurrent geometric properties and relations should appear on the wish list
of any modal relationalist just starting out in the world. It will be helpful to
distinguish two possible targets one might aim at here. Let us for the time
being take for granted the notion of an intrinsic property—a qualitative
property that an object has in virtue of how it is in itself, independently
of how the rest of its world is arranged.34 And let us as usual say that
two objects are duplicates if they share all of their intrinsic properties—and,
by extension, that they are geometric duplicates if they share all of their
intrinsic geometric properties. We call a version of modal relationalism
grounded if according to it, any two worlds whose material configurations
are duplicates agree about all (qualitative) questions concerning geometric
32
This supposition drives the classic discussion of geometric possibility, Field, “Can We
Dispense with Spacetime?” See also e.g. Earman, World Enough and Space-Time, §6.12 and
Tooley, “A Defense of Absolute Simultaneity,” p. 230.
33
Why would substantivalists require such a notion? In considering theories in which
there is a nomic dependence of geometry on the distribution of matter, one often has recourse
to test bodies—particles that obey the laws of motion of the theory but which make no
contribution to the shaping of geometry by matter. In introducing test bodies, one has a
choice: either offer a promissory note in place of results showing that solutions featuring test
bodies emerge in certain limits of honest solutions of the equations of the theory; or excuse
oneself from this hard work via recourse to a notion of geometric possibility.
34
For further discussion and references, see §3 of Ch. V below.
possibility. We call an account super-grounded if it is grounded and, further,

it implies that intrinsic geometric properties and relations are themselves
non-modal. So according to a grounded relationalist the facts about geo-
metric possibility at a world supervene on the intrinsic geometric properties
at that world; super-grounded relationalists add that the supervenience base
is not dispositional or otherwise tainted by modality.35
There are, of course, other prima facie attractive goals that modal rela-
tionalists may set themselves.
On the one hand, many relationalists will want to take the basic
geometric facts at a world to be facts about the distances between bits of
matter at that world (and whatever facts follow from these)—such rela-
tionalists take all of the intrinsic geometric facts about a world to have been
specified once the distance relations between material points have been
specified. I call such relationalists metric.
On the other hand, many relationalists will want to recognize as many
qualitatively distinct possibilities as substantivalists can—if there is some
substantival world with matter of a given configuration occupying a space
of a given geometry, then such relationalists will want to recognize the
existence of a world in which the configuration of matter is the same as
that in the substantivalist world and in which exactly the same facts about
geometric possibility of material configurations obtain.36 Let us call rela-
tionalists with this aspiration ambitious.
Now, one expects that substantivalists will recognize the existence of a
large class of worlds at which matter consists only of a single particle, and
within this class many distinct spatial geometries will occur. If this expect-
ation is fulfilled, then it becomes impossible for relationalists to combine
groundedness, metricity, and ambition. For suppose, as is reasonable, that
for each n, substantivalists countenance a one-particle world with the
structure of Euclidean n-space and a one-particle world with the structure
35
Conservative modal relationalists also take all geometric facts at a world to supervene on
the pattern of geometric relations instantiated at the world. But there are important doctrinal
differences between conservative relationalists and grounded modal relationalists. For ex-
ample: if w is a world at which the pattern of geometric relations instantiated by matter does
not change from instant to instant, then any conservative relationalist will take the spatial
geometry of w to be given by that pattern but (as we will see in the next chapter) grounded
modal relationalists need not do so.
36
Roughly speaking, the idea is that for any substantivalist world w with material
configuration C there is a relationalist world w * in which the matter configuration C* is a
geometric duplicate of C and such that w and w * agree about which material configurations
are possible and about which are possible extensions of the given configuration.
of hyperbolic n-space. If two such worlds differ as to the dimension of

space or as to whether it is Euclidean or hyperbolic, then they differ
radically about which configurations are geometrically possible. But
every such space will feature the same set of facts about the actual distance
relations between material points (namely: there is one point, zero units of
distance from itself ). So relationalists who are both ambitious and metric
will be be required to say that there exist relationalist worlds whose
material configurations are duplicates of one another, but which differ
about what is geometrically possible. Ambition, groundedness, and metri-
city are not cotenable under these circumstances.
In the following three chapters, we will examine three possible re-
sponses to this observation: one could attempt to hold on to metricity and
groundedness while giving up on ambition; one could attempt to hold on
to ambition and metricity while giving up on groundedness; or one could
attempt to hold on to ambition and groundedness while giving up on
metricity. Each of these strategies is more or less closely analogous to an
established strategy for handling the notion of physical possibility: the first
is very closely related to the best-system approach associated with David
Lewis; the second is similar in many ways to the primitivist approach of
John Carroll and Tim Maudlin; the third bears some similarity to the
necessitarian approach developed by Sydney Shoemaker, Brian Ellis, and
others.
III
Best-System Approaches
The thoughtful opinions written by The Chief Justice and by

Justice Ginsburg have persuaded me that current decisions by
state legislatures, by the Congress of the United States, and by this
Court to retain the death penalty as a part of our law are the product
of habit and inattention rather than an acceptable deliberative process
that weighs the costs and risks of administering that penalty against its
identifiable benefits, and rest in part on a faulty assumption about the
retributive force of the death penalty.
Justice Stevens
1 Introduction
It seems perfectly natural that relationalists seeking to provide an account
of the notion of geometric possibility should look to well-known accounts
of nomic possibility. Nick Huggett appears to have been the first to have
made a substantive suggestion along these lines, when he proposed a
means of adapting the best-system approach to laws of nature to the
geometric context.1
This chapter is focused on Huggett’s approach, which I take to be
representative of approaches to understanding geometric modality that
are grounded and metric but unambitious.2 I first give a brief overview of
1
“The Regularity Account of Relational Spacetime.” See also ch. 7 of Huggett, True
Motion.
2
Recall that an approach to this problem is grounded if it implies that worlds whose
material configurations are geometric duplicates of one another agree concerning geometric
possibility; that an approach is metric if it implies that two material configurations are
geometric duplicates of one another if and only if they instantiate the same pattern of distance
relations; and that (roughly speaking) an approach is ambitious if it implies that for every
substantivalist world w with material configuration C, there is a relationalist world whose
material configuration is a geometric duplicate of C, and at which the facts about geometric
possibility mirror those of w.
best-system approaches 55
the best-system approach to laws of nature, then discuss Huggett’s adap-

tation of this approach to the geometric context, then raise some worries
about both approaches. These worries are not new ones—they centre on
the difficulty of securing the conceptual resources posited by best-system
approaches and on some well-known unwelcome consequences of such
approaches. These worries seem pretty serious to me in the nomic context
and more serious still in the geometric context.
Note: for ease of exposition, in this chapter and in the following two, the
focus will be on static worlds (i.e., worlds in which matter determines the
same material geometry at every instant).
2 Lewis on Physical Possibility

The best-system approach to laws of nature has been developed in detail
by David Lewis.3 For Lewis, specifying a possible world involves specify-
ing the pattern of instantiation of certain privileged properties and rela-
tions, the fundamental properties and relations.4 Let us call the totality of
such facts about a world the basic facts about that world.
Associated with any possible world w is a language Lw whose non-
logico-mathematical vocabulary provides names for the fundamental
properties and relations that are instantiated at w.5 A theory for w is a set
of sentences of Lw closed under logical consequence, each member of
which is true at w.
There will be many theories for w, varying from the strongest (which
includes every truth about w stateable in Lw ) to the weakest (which
3
See esp. Counterfactuals, §3.3; Philosophical Papers, vol. ii, pp. x–xii; “A Subjectivist’s
Guide to Objective Chance,” pp. 121–4; “New Work for a Theory of Universals,”
pp. 41–3; and “Humean Supervenience Debugged,” pp. 231–3. Lewis’s approach was
prefigured in remarks of Mill and Ramsey.
4
“Fundamental properties are those properties I have elsewhere called ‘perfectly
natural.’ . . . They are not at all disjunctive, or determinable, or negative. They render their
instances perfectly similar in some respect. They are intrinsic; and all other intrinsic properties
supervene on them” (“Ramseyan Humility,” p. 204; see also “New Work” and pp. 60–9 of
On the Plurality of Worlds).
5
Note: the logico-mathematical resources of Lw need not be restricted to those of first-
order logic or one of its close relatives. Of course, this in effect gives Lewis’s analysis another
free parameter—it is implausible that each reasonable candidate for Lw will lead to the same set
of laws at each world (consider by way of illustration the difference in strength between
theories of arithmetic that are finitely axiomatizable in first-order languages and those that are
finitely axiomatizable in second-order languages).
56 best-system approaches
includes only logical and mathematical truths). Here we employ a standard

logical notion of strength: theory T is stronger than theory T * if the set of
models of T is a proper subset of the models of T * . Of course, in general if
we are given two theories for w, we do not expect either to be stronger
than the other.
Two theories for a given world can be compared in other respects as
well. Since both are formulated in Lw , it makes sense to compare them (or
axiomatizations of them) with regard to simplicity in various syntactic
respects. As Lewis emphasizes, such comparisons are plausibly viewed as
being fully objective (i.e., invariant not just across people, but across cog-
nizing agents): “It’s not because of how we happen to think that a linear
function is simpler than a quartic or a step function; it’s not because of
how we happen to think that a shorter alternation of prenex quantifiers
is simpler than a longer one.”6 These criteria are local: they underwrite
judgements concerning simplicity in this or that special respect. More
contentiously, Lewis holds that there is a natural global (i.e., all things
considered) notion of the simplicity of a theory that somehow aggregates
all of the local notions. This notion is assumed to be intersubjective
(i.e., invariant across suitably ideal human agents), although it is conceded
that it is likely to be species-relative.7
In general, one expects there to be a negative correlation between
simplicity and strength. A theory for w is ideal if it achieves an optimal
balance of simplicity and strength. According to the best-system approach,
the laws at w are the generalizations that appear in all of the ideal theories
for w (the hope being that there is a single ideal theory for typical worlds)
and a world w * is physically possible relative to w if the laws of w are true at
w * . As with simplicity, so with balance—Lewis does not expect there to be
a fully objective notion, but hopes that there is a suitable intersubjective
one.
The attraction of the best-system account is evident: specifying the
pattern of fundamental properties and relations instantiated at a world
determines the nomic facts at that world. So nomic facts supervene on
the basic facts and laws of nature and need not be taken to be metaphys-
ically primitive. And if one takes the facts regarding fundamental proper-
ties and relations to be non-modal in nature, then the present approach is
6
“Debugged,” p. 232. But cf. Carroll, Laws of Nature, pp. 50 f.
7
“Debugged,” pp. 231 f.
one under which the facts about nomic possibility and necessity supervene
on non-modal facts.8
Of course the approach also has its problems. Below I discuss several that
also apply to the best-system account of geometry.
3 Huggett on Geometric Possibility

Let us begin by considering what happens if one attempts a simple-minded
adaptation of Lewis’s framework to the geometric case. (This doesn’t work
very well, but a discussion of its shortcomings will help to motivate the
subtler approach advocated by Huggett.)
Suppose that we take as the basic geometric facts at a relationalist world w
the distance relations between material points. Associated with w will be
a language Lw whose vocabulary allows us to talk about such distances.
A geometric theory for w will then be a set of sentences in Lw that is closed
under logical consequence, with each sentence true at w.9 And an ideal
geometric theory for w will then be a theory that best balances simplicity and
strength. We can then say that a material configuration is geometrically
possible relative to w if for each ideal geometric theory of w a pattern of
that form is instantiated in some world permitted by that theory.
This straightforward analysis of geometric possibility is unpromising.
Consider a relationalist world w consisting of a zillion material points. And
suppose that the distance relations instantiated by these points are compli-
cated (no simple and widespread patterns, by human standards) but that
they are consistent with embedding the points into Euclidean three-space.
Best-system relationalists will want to say that a world like w is a world in
which space is Euclidean (since the other geometries consistent with the
material configuration of w are decidedly more complex). In particular,
they will want to say that despite the finitude of the material configuration
of w, the existence of an unbounded, collinear set of material points is a
geometric possibility at w.
But as it stands the present approach will not fulfil this wish. What
would an ideal geometric theory for w look like? Very plausibly such a
8
The question whether Lewis’s approach meets this higher standard is discussed in §5 of
Ch. V below.
9
The four classical geometries admit relatively attractive axiomatizations along these lines.
See Blumenthal, A Modern View of Geometry, chs. vii and viii.
theory would include the following axiom: the pattern of distances be-
tween material points in w is instantiated by some subset of Euclidean
three-space. This axiom is very strong and can be given a relatively simple
form as an algebraic condition on the distances among six-member sets of
material points in w.10 If this axiom alone sufficed to determine an ideal
geometric theory for w, then (setting aside worries about uniqueness) we
would have what we wanted—the patterns of distance relations geomet-
rically possible according to w would be just those consistent with Euclid-
ean geometry. But in order for this theory to be an ideal geometric theory,
it would have to achieve an undominated balance of simplicity and
strength. It does not—we can considerably strengthen our theory without
appreciably diminishing its simplicity by adding as an axiom the assertion
that there are no more than a zillion points or the assertion that there are
finitely many points. By eliminating all infinite models, the addition of
either of these (very simple) axioms would represent a gigantic increase in
strength.11 But then it follows that an ideal geometric theory for w would
permit only finite relationalist worlds—so, in particular, such a theory
would not recognize as geometrically possible an unbounded, collinear
set of material points.
So the most straightforward way of extending the best-system approach
to the geometric setting is a bit of a disaster. Under a Huggett-style
approach one avoids this mess by, in effect, working with a different
notion of a geometric theory. The crucial definitions are as follows.12
10
See Blumenthal, Theory and Application of Distance Geometry, ch. iv.
11
Either addition would cut down the set of isometry classes of models from one with the
cardinality of the power set of the continuum to one with the cardinality of the continuum.
12
The framework sketched here differs from Huggett’s approach in a couple of respects.
(i) For Huggett a geometric theory includes a particular choice of embedding. We can ignore
this wrinkle for present purposes (it engages with complications that we are postponing—see
the discussion of superposability vs. congruence in Ch. IV below). (ii) For Huggett only
metric spaces that arise from Riemannian metrics can play the role of geometric theories. But
as we saw in Ch. I above, this seems unduly restrictive: (a) there are simple metric spaces that
fail to be Riemannian but which nonetheless seem to be every bit as good candidates to
represent a possible spatial geometry as your average zillion-dimensional Riemannian mani-
fold; ( b) the principle of the plenitude of possibilities seems to require us to allow as
corresponding to possible structures of space a very wide range of spaces if we allow even
two-dimensional Riemannian manifolds. Note also that while the main concern of Huggett’s
paper is with relationalist theories of motion, I here focus on the very special case of static
universes.
A metric space (X, d) is a geometry for a relationalist world w if there is

a subset Y X isometric to the material configuration of w.
A geometry (X, d ) for w is ideal if it achieves an undominated balance
of strength and simplicity.
A pattern of distance relations is geometrically possible relative to w if it is
instantiated in each ideal geometry for w.
As in the nomic case, the great advantage of a best-system approach is that
it grounds modal facts in a relatively unproblematic base—facts about
geometric possibility supervene on facts about distances between material
points. Of course, in order to make the account precise, one has to say
something about the notions of simplicity and strength in play here.
Consider simplicity first. In previous discussions we worked with
linguistic entities—sets of sentences drawn from a distinguished language.
So it made sense to consider syntactic criteria for simplicity—and to hope
that those would suffice. In the present setting, our units of analysis are
geometric rather than linguistic entities, and it is natural to look for
language-independent criteria of simplicity. Huggett himself suggests a
list of such criteria and makes a rough proposal for weighing them against
one another.13 Presumably this is meant not as a stipulation but as a
conjecture about how we might systematize our intuitions about the
relative degree of simplicity of various geometries.
Then there is strength. Within Lewis’s framework ‘strength’ meant
logical strength—theory T is stronger than theory T * if the set of models
of T is a proper subset of the set of models of T * . This can’t be carried over
directly to the present context—it doesn’t make any sense to say that one
metric space is logically stronger than another.
One way we might approach this issue would be to note that to specify
a geometry for a world is to specify a metric space (X, d ). And one way
of doing that is to specify a theory whose models are all isomorphic to
(X, d).14 Each theory of this kind is maximally strong—since it determines
its models up to isomorphism there is no consistent theory in the same
13
Restricting attention to the special case of metric spaces arising from Riemannian
manifolds, Huggett suggests that “ ‘simplest’ means in order of decreasing importance, lowest
dimension, most regular, flattest, and smoothest” (“Regularity Account,” p. 54; see also True
Motion, §7.5).
14
Of course, this feat usually requires something more powerful than standard first-order
logic.
language with a smaller set of models. If we follow this line of thought—as

Huggett himself seems to recommend—then we find that strength is no
consideration at all in our present setting, so one achieves the best balance
of strength and simplicity by maximizing simplicity.15
But this way of cashing out strength strikes me as rather unnatural.
Suppose that we are comparing two-dimensional and three-dimensional
Euclidean geometry in respect of strength. Then it seems more natural to
say that the two-dimensional space is a stronger geometry than its three-
dimensional counterpart because every pattern of distance relations be-
tween points allowed by the former is allowed by the latter but not
vice versa, rather than saying that the two geometries are equally strong
because they can both be characterized up to isomorphism by theories in
some suitably powerful logical framework. So I propose that we think of
strength in the following terms: if (X, d ) and (X * , d * ) are geometries for a
relationalist world w, then (X, d) is stronger than (X * , d * ) if (X, d ) is
isometric to a proper subspace of (X * , d * ), but not vice versa. Understood
this way, strength is a non-trivial feature and balancing strength and
simplicity is, as it should be, a difficult task. (Very little in what follows
will hang on whether this notion of strength is used in place of Huggett’s.)
4 Worries About Best-System Approaches

Lewis’s best-system approach to nomic possibility faces a number of well-
known objections.16 Here I will discuss four sorts of worries that carry over
directly to best-system treatments of geometry: worries regarding the notion
of simplicity; worries regarding the notion of balance of simplicity and
strength; worries regarding uniqueness; and worries regarding supervenience.
While such worries play a role in some authors’ rejection of the best-
system approach to laws and give many of the rest of us serious pause,
advocates of the approach are often satisfied to set them aside more or less
cheerfully.17 Defenders of the best-system approach to laws sometimes
15
See Huggett, “Regularity Account,” p. 54.
16
For overviews of problems and responses see Carroll, “Nailed to Hume’s Cross?” and
Laws of Nature, chs. 2 and 3; Loewer, “Humean Supervenience;” and van Fraassen, Laws and
Symmetry, ch. 3.
17
See e.g. Loewer, “Humean Supervenience;” Beebee, “The Non-Governing Concep-
tion of Laws of Nature;” Earman, A Primer on Determinism, §v.5; Earman and Roberts,
“Contact with the Nomic. Part I.”
challenge critics to produce examples that demonstrate that these abstract

worries are genuinely pressing.18 A nice feature of the geometric context is
that it proves a setting in which it is relatively easy to construct such
examples. This ought to make it correspondingly more difficult for advo-
cates of best-system modal relationalism to brush aside the standard worries
about best-system approaches.
Simplicity
Lewis requires for his treatment of nomic possibility a comparative notion
of the simplicity of theories. This notion must be global (simplicity all-
things-considered), intersubjective (invariant across suitably ideal people) if
not objective (invariant across cognizing agents), and suitably natural (so
that we are not faced with alternative notions with no grounds for
choosing between them). I will call such a notion with these features a
notion of Simplicity (with an initial capital).
Lewis in effect advances the bold empirical conjecture that a notion of
Simplicity is available.19 But Lewis does very little to make this conjecture
plausible beyond noting that there are available various relatively objective
but local notions of simplicity applicable to his context.
So it is natural to worry whether an analysis of the notion of law can
legitimately presuppose a notion of Simplicity. Sometimes this worry is
expressed by pointing out how far advocates of best-system approaches are
from being able to give a satisfactory account of Simplicity. Thus, van
Fraassen remarks in the course of a discussion of philosophical accounts of
laws of nature, that “the literature contains no account of [simplicity]
which it would be fruitful to discuss . . .”.20
But when the worry about Simplicity is cast in this form, it is easy for
advocates of best-system approaches to brush it aside. No acceptable
analyses of the notions of knowledge or causation are available. But it is
for the most part not considered to be intellectually irresponsible to take
such notions for granted in giving accounts of reference, intentional
18
See e.g. Earman “In Defense of Laws,” p. 418.
19
None of the other standard philosophical accounts of laws of nature feature an empirical
component. To my mind this is one of the most interesting aspects of the best-system
account.
20
Laws and Symmetries, p. 41. For a survey of some formal approaches to the notion of
symmetry, see Pambuccian, “Simplicity.”
action, and so on. We feel entitled to put such notions to work even in the
absence of satisfactory analyses of them.
The worry about simplicity is not so easily dismissed, however. For it is
only to the extent that we are convinced that there is broad, almost
universal, agreement in the application of the notions of cause and know-
ledge to central cases that we are happy to put these notions to work
elsewhere.
Contrast this with a case like the notion of overall quality as applied to
hockey players, thought of as a global notion that aggregates more spe-
cialized comparative notions—for present purposes, let us take the relative
specialized notions to turn on comparisons of skill and ferocity. No one
values ferocity so highly relative to skill as to claim that Dave “The
Hammer” Schultz (or even Dave “Tiger” Williams) was, all things con-
sidered, a better player than Wayne “The Great One” Gretzky. But the
question of how these two factors should be weighed is a much-disputed
one, and there is spirited disagreement among journalists and drunk people
as to whether Gretzky or Maurice “The Rocket” Richard was the better
player. And the problem is not that these players are tied relative to some
standard, intersubjective notion of quality—imagining a slightly more
ferocious Gretzky would not settle the matter. Nor, at least for the
examples chosen, is the problem that there is a lack of intersubjective
agreement about judgements of ferocity and skill. Rather, each of Gretzky
and Richard is better than the other relative to some reasonable ways of
weighing skill against ferocity, but not others. It is implausible that we here
have a case where there is a univocal intersubjective notion of overall
quality that aggregates the special notions. If seriously pressed to name the
best player, one must ask “Best according to which way of weighing skill
and ferocity?” There is no shared default standard that resolves this ques-
tion in interesting cases.
It is natural to worry that Simplicity is in the same boat. I suspect that if
directed to attend to matters mathematical and matters ontological, almost
anyone would agree that Newton’s theory of gravity was simpler than
Einstein’s. The former enjoys such a large advantage with respect to the
simplicity of its formalism that it would be perverse to claim that this is
outweighed by the fact that the Newtonian explanation of the motion of
heavy bodies requires us to advert to both spacetime and force, while in
general relativity spacetime is the whole story.21 But if pressed to say which
is simpler, the Poisson equation-based version of Newtonian gravity or the
Newton-Cartan version, I think the only proper response is to demand
instructions as to how the advantage of the former over the latter in terms
of mathematical simplicity (only the latter requires any differential geom-
etry) is to be weighed against the ontological simplicity of the latter over
the former (again, one theory invokes both spacetime geometry and force,
the other only geometry).22
Discussion of examples of this kind is not likely to change anyone’s
mind. For while it seems to me that the burden of proof lies with those
who posit a notion of Simplicity in service of their philosophical projects,
they are only too likely to disagree. Further, partisans of Simplicity are
likely to question the intuitions appealed to in the example just given, or
to shrug off that example as involving the sort of tie that one expects will
occasionally arise. In any case, the intuitions of we philosophers are in
fact of limited interest here—intersubjectivity in the relevant sense only
requires agreement across suitably ideal people, and in the present context
ideality will presumably involve a quite high degree of mathematical
facility and sophistication.
Productive debate as to the availability of a notion of Simplicity as
applied to physical theories would require an interesting (non-toy) family
of examples about which experts showed a degree of agreement or
disagreement out of step with the expectations of sceptics or advocates.
There is good reason to doubt that one will turn up such a family of
examples in the nomic context. For serious scientific theories are difficult
to concoct and all too often one finds such great increases of mathematical
complexity from one generation of theory to the next that the question
how one ought to weigh such complexity against other factors does not
really arise.
Things looks somewhat different in the geometric case. As noted above,
in this setting we work, in the first instance, with geometric rather
than linguistic objects—but we are of course free to consider a theory
21
Objection: in fact, some experts do judge Einstein’s theory to be simpler on math-
ematical grounds (see Dafermis, “General Relativity and the Einstein Equations,” p. 483).
Reply: to the extent that such examples are taken to heart, the illustration offered is flawed—
but unless this judgement represents the consensus of the relevant experts, to take it seriously
is to undercut the case in favour of Simplicity.
22
For these theories see e.g. Friedman, Foundations of Space-Time Theories, ch. iii.
characterizing a given geometric object. So we expect there to be in play

both syntactic and geometric (local) notions of simplicity. The question
will again be whether there is some distinguished way of aggregating these
local notions into an intersubjective global notion. But it is far more
plausible in the geometric than in the nomic setting that we can find
cases which will help us to adjudicate this issue.
Consider the question of which among the four classical geometries is
simplest, all things considered. This is far from being a toy question.
Indeed, it is one with considerable historical resonance, and so we might
hope that geometers would have recorded their opinions about this
matter. And in examining what they have to say on this question, we
might hope to garner some evidence for or against the claim that there is a
distinguished global notion of simplicity of geometry that is invariant
across suitably idealized human agents.
It seems to me that cases like this provide the setting for a serious
investigation of global notions of simplicity in geometry. In lieu of such
a study, let me mention a few pieces of evidence that it was easy for an
amateur to turn up.23
The first authority that any philosopher will turn to on this question is
Henri Poincaré.24 For considerations of simplicity play a dominant role in
his reasons for thinking that we will always opt to take physical geometry
to be Euclidean.
Euclidean geometry is, and will remain, the most convenient . . . because it is the
simplest, and it is so not only because of our mental habits or because of the kind of
direct intuition we have of Euclidean space; it is the simplest in itself, just as a
polynomial of the first degree is simpler than a polynomial of the second degree;
the formulas of spherical trigonometry are more complicated than those of recti-
linear trigonometry, and they would still appear so to an analyst who was ignorant
of their geometric significance.25
Let it not be said that the reason why we deem the group of Euclid [i.e., the group
of isometries of Euclidean space] the simplest is because it conforms best to some
pre-existing ideal which has already a geometric character; it is simpler because
certain of its displacements are interchangeable with one another, which is not true
23
I looked in a score or so obvious places. The evidence I collected is adduced below.
24
Poincaré’s judgement concerning the simplicity of Euclidean geometry is endorsed by
Schlick (Mulder and van de Velde-Schlick (eds.), Moritz Schlick, p. 168) and (in a sense) by
Reichenbach (The Philosophy of Space and Time, p. 83).
25
Poincaré, La Science et l’Hypothèse, p. 67. The standard English translation omits the
material following the second semicolon; see Greenstreet (trans.), Poincaré, p. 50.
of the corresponding displacements of the group of Lobachevsky [i.e., the group of

isometries of hyperbolic geometry]. Translated into analytical language, this means
that there are fewer terms in the equations, and it is clear that an algebraist who did
not know what space or a straight line was would nevertheless look upon this as a
condition of simplicity.26
But Poincaré’s position has not gone unchallenged. Karl Menger, speaking
of an axiomatization of hyperbolic geometry whose two sorts of variables
range over points and lines and whose only primitive relation is a notion of
incidence, remarks that:
in Hilbert’s terminology, this is a development on the basis of axioms of alignment
alone, without the traditional axioms of order, parallelism, congruence, and
perpendicularity. In euclidean geometry, on the other hand, assumptions about
concepts other than join and meet are indispensable. This situation does not seem
to bear out a remark repeatedly made by Poincaré: that among the geometries that
might be used in describing physical space, euclidean geometry would always be
distinguished by its greater simplicity.27
As a third authority, who appears to be tempted by a third view, consider

the great anachronistic classical geometer H. S. M. Coxeter, who main-
tains that elliptic geometry “is in many ways simpler than either spherical
or Euclidean geometry.”28
What are we to make of such disagreement? The most natural conclu-
sion would be that each of our authorities is speaking of a different global
notion of simplicity, arrived at by attributing quite different weights to
shared intersubjective but local notions of simplicity. This sort of diver-
gence is, I think, the sort of thing that Felix Klein hints is all too common
when, speaking near the end of a series of lectures, he reminds his audience
26
“On the Foundations of Geometry,” p. 145.
27
Blumenthal and Menger, Studies in Geometry, p. vii; see also Menger, “The New
Foundations of Hyperbolic Geometry,” p. 495. Torretti appears to endorse Menger’s view;
Philosophy of Geometry from Riemann to Poincaré, p. 246. For further discussion, see Pambuccian,
“Axiomatizations of Hyperbolic Geometry,” pp. 334 f.
28
Non-Euclidean Geometry, p. 13. On Coxeter in general, see Roberts, King of Infinite
Space. Although in the passage quoted Coxeter stops short of proclaiming elliptic geometry to
be the simplest in a global sense, the fact that the remark quoted is the only comment about
simplicity that occurs in a survey of the classical geometries suggests that he is at least tempted
by the global claim. Consider also in this connection Clifford’s remarks on the aesthetic
contrast between, on the one hand, Euclidean geometry (with its dreary infinities), and, on
the other, spherical geometry and elliptic geometry (far more complete and interesting);
Lectures and Essays, p. 230.
that he has treated the classical geometries both from the group-theoretic
point of view and from the axiomatic point of view, and continues
I should especially like to emphasize again this fact, in the face of intolerant
utterances which one often hears, and which are aimed at championing this or
that pet concept of the author, as absolutely the simplest and, in fact, the only
suitable one to use in the foundations of geometry.29
It is not difficult to read such tendencies into our divergent authorities.

Poincaré speaks as if there were a natural, intersubjective, and global
notion of simplicity in the air—and as though it speaks unequivocally in
favour of Euclidean geometry as against its upstart rivals. But the consid-
erations that Poincaré appeals to are (mathematically speaking) practical
ones, of the sort likely to loom large to a far-ranging mathematician of his
type, whose work in analysis or geometry sometimes carries him into one
or another classical geometric context. Menger, the erstwhile in-house
geometer of the Vienna Circle, focuses his remarks rather narrowly on
foundational considerations: he and his students have developed an
appealing approach to hyperbolic geometry that renders it in an interesting
respect similar to projective geometry and he is impressed that Euclidean
geometry cannot be put into such a form. Coxeter speaks as a geometrical
generalist with limited interest in foundations.30 It is plausible, I believe,
that in the passage quoted above Coxeter intends to allude to the fact that
of all the classical geometries, elliptic geometry has the closest relation to
projective geometry—a geometry that Coxeter considers simpler than
Euclidean geometry.31
Of course, the foregoing amounts to nothing more than a somewhat
tendentious reading of a very small sample of evidence. But it would
require real stubbornness, I think, to see this evidence as actually supporting
the claim that there is a single coherent global notion of geometrical
simplicity implicit in the judgements of experts.
29
Klein, Elementary Mathematics from an Advanced Standpoint: Geometry, p. 207.
30
For Coxeter’s tastes in geometry see Logothetti, “An Interview with H. S. M. Coxeter,
the King of Geometry.”
31
The close relation between elliptic and projective geometry is emphasized by Coxeter
on pp. viii, 95, and 110 of Non-Euclidean Geometry. Elsewhere he characterizes projective
geometry as involving “a beautiful and intricate system of propositions, simpler than Euclid’s
but not too simple to be interesting” (Projective Geometry, p. 2).
Balance
Lewis’s account of nomic possibility requires a distinguished intersubjec-
tive notion of balance between simplicity and strength. Here again there is,
to put it mildly, a paucity of evidence in favour of the existence of such a
notion.32
But the role of balance in best-system accounts also raises conceptual
issues. These have their roots in the uncontroversial observation that it is
wholly implausible that there could be a suitable notion of the balance of
simplicity and strength which was objective in the sense of being invariant
across suitably ideal rational creatures—surely beings whose innate com-
putational capacities were quite different from our own would arrive at
quite different bargains in trading off simplicity against strength. It follows
that (at best) the notion of law has a surprising species-relativity under the
best-system account: in some cases at least, the laws-for-humans at a given
world will be different from the laws-for-Vulcans at the same world.33
Lewis floats the optimistic hope that the laws of our own world might
have the special feature that each of these species-relative notions of law of
nature agree here.
If nature is kind, the best-system will be robustly best—so far ahead of its rivals that it
will come out first under any standards of simplicity and strength. We have no
guarantee that nature is kind in this way, but no evidence that it isn’t. It is a
reasonable hope. Perhaps we presuppose it in our thinking about law. I can admit
that if nature were unkind and if disagreeing rival systems were running neck-and-
neck, then lawhood might be a psychological [i.e., species-relative] matter, and
that would be very peculiar.34
This suggestion strikes me as doubly strange. In the first place, it seems to

rely upon an oddly impoverished conception of the range of cognitive
powers of possible cognizing beings. In the second, I do not see why talk
of nature’s being unkind at worlds in which distinct species-relative
notions of law diverge is appropriate, even metaphorically. Why think
that worlds at which there is disagreement between species-relative
32
In fact, the situation is even worse for balance than for simplicity, since it is rare for
authorities to record their judgements concerning balance.
33
Note that ‘laws-for-humans’ is rigidified—at each world it picks out the generalizations
of the best-system describing that world as measured by our actual human notions of simplicity
and balance.
34
“Debugged,” p. 233.
notions have to be especially strange (so that it would be a reasonable hope

that we do not live at such a world)?
These worries can be illustrated via the sort of example that causes
trouble for the naïve regularity account of laws of nature, which takes
the laws at a world to be the generalizations (stateable in some suitable
language) that are true at that world. Under this account it is a law at our
world that no gold spheres are more than a mile (say) in diameter and a law
that no uranium spheres are more than a mile in diameter. But, intuitively,
it seems that the former generalization holds accidentally while the latter
holds of physical necessity. The naïve regularity account fails to recognize
this crucial distinction. Lewis’s account is designed to do justice to this
intuition, without compromising the minimalist spirit of the naïve account.
For on Lewis’s account, the fundamental principles of particle physics will
presumably be part of any candidate for ideal theory for our world, and
since they imply the regularity about uranium spheres, this regularity will
count as a law under a best-system approach. But it is implausible that the
regularity about gold spheres will be part of the ideal theory of our world.
It is not implied by our currently accepted physical principles and adding it
as an axiom would contribute so little to the strength of any candidate
theory that it would not be worth the rather minuscule decrease in
simplicity that addition would involve. Or, rather, we would not consider
this to be a good bargain—but presumably to a being with much greater
cognitive and computational capacities than our own, the bargain might
well be worth it. And why not think that there are beings with cognitive
and computational capacities much greater than our own—if not in our
world, then in some possible world? Compared to us, dogs are quite
unwilling to trade even a little bit of simplicity for a great deal of strength
in their theories—and why not think that there are beings who stand to us
as we stand to dogs? But if this is granted, it is hard to see how to avoid the
conclusion that any reasonably complex world can be expected to exhibit
regularities various enough so that there will be at least one regularity that
counts as a law relative to some possible species of brainiacs while counting
as a mere regularity relative to some possible species of dullards.
The following two examples show that the the same sort of species-
relativity shows up in the geometric case—the geometry-for-humans of a
given world may differ from its geometry-for-Vulcans.
Example 4.1 (Discrete vs. Euclidean). Consider a relationalist world w con-

sisting of n particles, and suppose that the distances between the particles are
consistent with embedding into Euclidean three-space but otherwise quite
random-looking.35 What would an ideal geometry for this world look like?
For any n, there are two especially salient alternatives: Euclidean three-
space and the discrete metric space consisting of n points with distances as
in w. Of course, for any n, the discrete geometry will be much stronger
than the Euclidean one—since the latter permits every configuration
permitted by the former, but also permits many more besides. For small
n the discrete geometry will also be simpler than the Euclidean one. So if
the number of material points is small, the discrete geometry will be ideal.
But it seems plausible that there will natural numbers N and N * such that:
for n > N Euclidean geometry is simpler than the discrete alternative;
while for n > N * the advantage in simplicity outweighs the advantage the
discrete geometry enjoys in strength, so that the ideal geometry will be
Euclidean. This much holds relative to any reasonably sophisticated pos-
sible species of cognizers, I take it. But the values of N and N * will of
course depend on the cognitive capacities of the agents involved—pre-
sumably N * is larger for Vulcans than for humans, so there will be
relationalist worlds whose geometry-for-humans is Euclidean but whose
geometry-for-Vulcans is discrete. At such worlds, infinite sets of collinear,
equally spaced points are geometrically possible relative to humans but
geometrically impossible relative to Vulcans. &
An essential feature of the preceding example was that discrete metric

spaces were countenanced as ideal geometries for relationalist worlds. This
is contrary to Huggett’s official approach, which requires that ideal geom-
etries be Riemannian (see fn. 12 above) and it could also conceivably be
objected to as involving an excessively liberal notion of spatial geometry.
The next example shows how a similar result can be arrived at even if one
adopts a very conservative view concerning which metric spaces can
represent the structure of physical space.
Example 4.2 (Euclidean vs. Elliptic). For the purposes of this example, let
us assume that the only contenders to represent the structure of space are
the classical geometries of various dimensionalities. It follows that the only
35
Recall that we are restricting attention to static worlds: the pattern of distance relations
instantiated at w does not vary over time.
way for one geometry to be stronger than another is for it to be a lower-

dimensional space of the same family (elliptic, Euclidean, hyperbolic, or
spherical). So the question of balance-of-simplicity-and-strength reduces
to a question of simplicity.
Let us further stipulate that the standards of simplicity in play are similar
to those suggested by Poincaré: if we consider the four classical geometries
of a given dimension, we always consider Euclidean geometry to enjoy an
advantage in simplicity over its rivals. Indeed, we can even suppose that we
consider a Euclidean space of a given dimension to be simpler than space
of a somewhat lower dimension from the other families. But let’s not be
crazy: when the difference in dimension is sufficiently large, we regard
a lower-dimensional non-Euclidean geometry as simpler than a higher-
dimensional Euclidean geometry. Plausibly, the transition point is species-
dependent: because Vulcans have greater powers of computation and
visualization than humans, they consider a lower-dimensional non-
Euclidean geometry to be simpler than a higher-dimensional Euclidean
geometry in some cases where humans consider the difference in dimen-
sion too small to offset the simplicity of taking there to be exactly one line
through any given point parallel to any given line.
Let us call a set of points equilateral if there is some > 0 such that any
two distinct points in the set are separated by units of distance and call a
relationalist world equilateral if its material points form an equilateral set.
What does the ideal geometry for an n-point equilateral relationalist world
w look like?
Let us write E(k) for the cardinality of the largest equilateral set of points
in Euclidean k-space. Let us do likewise in the hyperbolic, spherical, and
elliptic cases, using in the obvious way the notation h(k), s(k), and e(k). For
any k $ 0, E(k) ¼ h(k) ¼ k þ 1. And for k > 0, s(k) ¼ k þ 2.36 So given
36
First, we argue that s(k) $ k þ 2. This is immediate for k ¼ 1. For k ¼ 2, an obvious
thing to try is spacing three points equally on the equator and putting the fourth at the North
Pole. But this won’t do: the distance between any two equatorial points is greater than the
distance between the pole and the equator. If we take three equally-spaced points on a line of
latitude close to the South Pole, we have the opposite problem—now the three equally-
spaced points are much closer to one another than to the North Pole. So there must be a line
of latitude somewhat south of the equator that will give us the desired equality. The same sort
of considerations show that we can always find an equilateral set of k þ 2 points in a spherical
space of dimension k.
Next we argue that we can never find an equilateral set of k þ 3 points in a k-sphere. A metric
space can be isometrically embedded in a k-sphere if and only if every (k þ 3)-tuple of points from
the space can be isometrically embedded in the k-sphere; Blumenthal, Theory and Applications,
our stipulations about simplicity, we find that Euclidean geometry always

provides a simpler-stronger geometry for an equilateral relationalist world
than does hyperbolic or spherical geometry.
What about elliptic geometry? Since elliptic one-space is just the circle,
e(1) ¼ s(1) ¼ 3. But the spherical and elliptic cases part ways in dimension
two: s(2) ¼ 4 while e(2) ¼ 6.37 Further surprises are in store, as one can see
in the following table.38 Here we see behaviour—occasional rapid
growth, occasional stagnation—very different from the slow, regular
growth of E(k), h(k), and s(k).
Table III.1: Equilateral Sets in Elliptic Geometry
e(1) ¼ 3 e(2) ¼ 6 e(3) ¼ 6 e(4) ¼ 10 e(5) ¼ 16
e(6) ¼ 28 ... e(13) ¼ 28 e(14) ¼ 36 e(15) ¼ 40
e(16) ¼ 48 e(17) ¼ 48 e(18) $ 72 e(19) $ 90 e(20) ¼ 126
e(21) ¼ 176 e(22) ¼ 276 ... e(41) ¼ 276 e(42) ¼ 344
Now return to our n-point equilateral world w. If n is small enough,

both humans and Vulcans will take the ideal geometry for w to be
Euclidean rather than elliptic. And for sufficiently large n, both species
will take the ideal geometry to be elliptic (the difference between the
dimension of the best Euclidean geometry for w and the dimension of the
best elliptic geometry for w can be made arbitrarily large and the corre-
sponding ratio can be made arbitrarily small).39 And there will be inter-
mediate values of n for which Vulcans take the ideal geometry for w to be
elliptic while humans take it to be Euclidean. &
How embarrassing are examples of this sort? That depends on one’s

metaphysical tastes. Very often, advocacy of best-system approaches is
theorem 39.2. Consider X, a countably infinite equilateral set. Clearly there is no isometric
embedding of X into the k-sphere. But if the k-sphere contained an equilateral (k þ 3)-tuple, the
theorem just cited would suffice to establish the existence of such an embedding.
37
For any k, e(k) is the largest number of pairwise equiangular lines that one can construct
through the origin in ℝk þ 1. For k ¼ 2, the largest such set is given by the six diagonals of an
icosahedron centred at the origin.
38
See Seidel, “Discrete Non-Euclidean
pffiffiffiffiffiffiffiffiffiffiffiGeometry,” §3.3.
39
For sufficiently large k, (k 1) k 1#e(k)# 12 k(k 1); Seidel, “Discrete,” §3.3.
presented as an expression of the author’s commitment to some sensible

form of empiricism or scientifically-rooted realism. But the traditional
name of the doctrine that the structure of space depends in part on our
constitution and might well be different for beings of a different cognitive
constitution is not ‘hard-headed empiricism’ or ‘straight-shooting realism’
but ‘transcendental idealism.’40
The point of this gibe is as follows. Kant was right to characterize his
view as a form of idealism. Idealism of this sort is a wild and interesting
metaphysical view. But it is not the sort of thing we are led to expect when
we are first told that the regularity view of laws can be fixed up to take care
of well-known problems, without resort to anything so wild as contingent
relations of necessitation between universals or a primitive notion of
physical possibility.
Advocates of best-system approaches tend to respond to this sort of worry
by pointing out that there are some other worries in the neighbourhood that
they can assuage. (1) It would be very bad if the approach rendered the
question whether a certain regularity holds by law a subjective matter. But it
does not. Once we fix that we are interested in laws-for-humans, it is a fully
objective matter whether a proposition is a law or not (and likewise for the
geometric case). (2) It would be bad if the approach committed us to saying
that if our capacities had been different, the laws of nature would have had to
be different. But it does not. For we should use the standards of simplicity and
balance determined by actual human capacities in deciding, for any world w,
what the laws-for-humans are at that world, no matter what the capacities of
the w-humans (and likewise in the geometric case).
This is well and good. But the basic worry is that under best-system
approaches things (laws, geometry) are rendered species-relative in a sense
that is, to most tastes, dismaying and unwelcome. Advocates of the best-
system approach sometimes acknowledge this worry.41 But they do little
to assuage it.42
40
Kant, Critique of Pure Reason, pp. A27 f./B43 f., B72, A230 f./B283, A286 f./B
342 f.
41
Thus, Lewis: “The standards of simplicity, of strength, and of balance between them are
to be those that guide us in assessing the credibility of rival hypotheses as to what the laws are.
In a way, that makes lawhood depend on us—a feature of the approach that I do not at all
welcome!” (“Subjectivist’s Guide,” p. 123). Others insist that they see nothing unwelcome
here—see e.g. §4.3 of Cohen and Callender, “A Better Best System Account of Lawhood.”
42
Unless you are satisfied with simply counting the notion of a law of nature as being
inherently indexical. For this sort of move, see Halpin, “Scientific Law: A Perspectival
Ties
Under any way of making precise the notions of simplicity and of the
balance of simplicity and strength, there are bound to be worlds at which
uniqueness of best theories fails. There are several ways of handling such
cases. One could take the laws to be the propositions common to all of the
best-systems for a given world.43 Or one could take it to be indeterminate
whether a proposition included in some but not all best-systems is a law.44
Or one could just say that no propositions are good candidates to be
considered laws at such worlds.45 Whichever route one takes, it turns
out that there are worlds at which the nomic facts are surprisingly weak.
Is this scenario worrying? Lewis thinks not: “. . . what of it? We haven’t
the slightest reason to think the case really arises.” 46
It is important to note that if we have no reason to think that our world
is a world that supports multiple ideal theories, it is not because such
worlds would have to be strange in themselves but because we know so
little about our world and about the relevant standards of simplicity and
balance.47
In the geometric case it is not hard to imagine cases where ties arise—
cases involving worlds which are very similar indeed to worlds supporting
unique ideal theories.
Example 4.3 (Euclidean vs. Elliptic, again). Let us return to the context of
Example 4.2 above: we suppose that the desired species-relative notions of
simplicity and balance exist and restrict attention to the classical geom-
etries, and ask what the ideal geometry is for an n-point equilateral world.
Presumably, the cognitive capacities of possible species lie on something
like a continuum. So we can take ‘Vulcan’ to name a possible species for
whom Euclidean geometry is ideal for n < 6, elliptic geometry is ideal for
Account” and Roberts, “ ‘Laws of Nature’ as an Indexical Term” and The Law-Governed
Universe, §3.3.7.
43
Lewis, “New Work,” p. 41.
44
Huggett, “Regularity Account,” §3.
45
Lewis, “Debugged,” p. 233.
46
Ibid.
47
Perhaps it will be suggested that since ties will be rare under any reasonable system of
standards, and since we have a default entitlement to assume that our world is typical, it
follows that we are entitled to think that our world does not support multiple ideal theories.
Very good. Let us add to the wish list of advocates of best-system approaches that the
appropriate standards of simplicity and balance are reasonable in this sense.
n > 6, but there is a tie for n ¼ 6—both Euclidean geometry and elliptic
geometry achieve an undominated balance of simplicity and strength. The
result is that there is no fact of the matter as to whether the space of a six-
point equilateral relationalist world is finite or infinite, relative to Vulcan
standards. This is surprising, since there is a determinate fact of this sort for
every other equilateral relationalist world and since (presumably) there is
also a determinate fact about six-point worlds for species of slightly
different cognitive capacities than Vulcans. &
Supervenience
It sounds attractive to have nomic facts supervene on facts about the
distribution of fundamental properties and relations. But it is not difficult
to cook up examples that put some pressure on this judgement.
Example 4.4 (Contra Supervenience). Suppose that our physics involves
three types of particles—the a-particle, the Æ-particle, and the a-particle—
and that our physical theory is complete except for an account of what
happens in the incredibly rare cases in which like particles collide. General
principles tell us that for each type of collision between like particles, the
nature of the interaction depends on the value of a constant of nature
associated with the particle type. For each possible type of collision, theory
tells us the constant is a complex number, either i or i. The consequences
of a collision of like particles is delightful if the value of relevant constant is
i, disastrous if its value is i. Thirsty for knowledge, we build a giant
experiment that causes two a-particles to collide—with delightful results.
Feeling lucky, we build a second device that leads to a collision of two Æ-
particles—with disastrous results. We decide not to press our luck further.
And in fact there never is a collision between two a-particles. But what if
we had built a device and caused such a collision? It is tempting to say that
there is fact of the matter as to whether the result would have been
delightful or disastrous. But this contradicts the supervenience of nomic
facts on the the pattern of basic facts—since it implies the existence of two
possible worlds just like ours in all occurrent facts, one of which shares its
laws with worlds in which collisions of a-particles are delightful, the other
of which shares its laws with worlds in which such collisions are disastrous.
&
Intuitions conflicting with supervenience are widespread—and cases
like the one above are often put forward in the course of criticism of the
best-system approach to laws of nature.48 Even some defenders of the best-

system approach apparently feel the pull of such intuitions.49
Of course, the same sort of intuitions can be dredged up in the
geometric case. Consider a four-point relationalist world. If the distance
relations are consistent with embedding into Euclidean three-space but
not with embedding into Euclidean two-space, then they are also consist-
ent with embedding into spherical two-space.50 Relationalists with even a
bit of ambition in them will think that we here have a failure of super-
venience of facts about geometric possibility on the distance relations
between material points: for it is tempting to think that two four-point
worlds could agree as to the distances actually instantiated while disagree-
ing as to whether if a straight line were filled out with matter it would be
finite or infinite in extent.
This territory can be approached from another direction. Under a best-
system approach, every world occupied by just a single point-particle has
the same geometry—space is zero-dimensional and the only geometrically
possible configurations of matter are the actual one and the empty one.
Now consider some richer world w. The best-system account tells us that
at w it is true that if all but one of the actual material points had never
existed (and no other material points had existed), then space would have
been zero-dimensional.51
Or, again, consider (non-degenerate) triangles ABC in Euclidean two-
space and A* B* C * in hyperbolic two-space, such that the two triangles
have corresponding sides of the same length. The two triangles are
nonetheless not isometric: the distance between A and the midpoint of
side BC is greater than the distance between A* and the midpoint of side
B* C * .52 If we now consider relationalist worlds w and w * whose material
48
See e.g. Carroll, Laws of Nature, §3.1; Fine, “Varieties of Necessity,” §3; Lange, Laws and
Lawmakers, p. 52; Maudlin, The Metaphysics within Physics, §2.3; Woodward, Making Things
Happen, §6.11.
49
See e.g. Huggett, “Regularity Account,” fn. 15; or Earman and Roberts “Contact with
the Nomic. Part I,” §1. Others steadfastly deny having such intuitions—see e.g. Beebee,
“Non-Governing Conception.”
50
Gödel, “On the Isometric Embeddability of Quadruples of Points of R3 in the Surface
of a Sphere” (item (1933b) in Feferman et al. (eds.), Kurt Gödel ).
51
This is the geometric analogue of the observation that under the best-system approach
to laws it does not follow from the fact that w is physically possible relative to w* that both
worlds have the same laws. For an interesting illustration of this phenomenon, see Earman,
Bangs, Crunches, Whimpers, and Shrieks, p. 182.
52
This is a basic difference between Euclidean and hyperbolic geometry. See e.g. Burago
et al., A Course in Metric Geometry, §4.1.
contents instantiate the same patterns of relations we find in triangles ABC

and A* B* C * then it is fairly plausible that the ideal geometry for w is that
of Euclidean two-space while the ideal geometry for w * is that of hyper-
bolic two-space.53 Now consider the counterfactual “If only the vertices
and not the sides of the triangle had existed, the geometry of space would
have been different.” This counterfactual is true at at least one of w and w *
under any best-system approach to geometry.54
General relativity is amazing and interesting in part because it sets up a
physical dependence of geometry on matter. The examples of the pre-
ceding two paragraphs bring out the sense in which the best-system
approach to geometry sets up a rather less interesting conceptual depend-
ence of geometry on matter.
5 Summation
The best-system approaches to laws and to geometry presuppose very
strong empirical claims about the strength and intersubjectivity of various
notions of simplicity and balance. These claims strike me as rather im-
plausible—and they are in fact largely unsupported by argument. Further-
more, the approaches have some rather surprising—and to my mind,
unattractive—consequences: notably that they render law and geometry
species-relative and imply a supervenience thesis that clashes with wide-
spread intuitions.
How much of this carries over to other grounded, metric approaches to
geometric possibility? Any such approach has to involve a solution to what
was called the Selection Problem in §2 of Chapter II above: given the
metric spaces encoding the instantaneous material geometries at a world,
to construct the metric space encoding the structure of space at that world.
When we discussed this problem previously, we saw that it had no
canonical solution—as soon as one considers even the simplest worlds,
one becomes involved in trading off considerations of size versus consid-
erations of symmetry, and so on. Any selection procedure will be subject
to complaints about the supervenience of modal facts on distance relations
53
All that is required for present purposes is that the two worlds differ in their spatial
geometry.
54
If we allow finite metric spaces as ideal geometries, then it is true at both; if we restrict to
‘nice’ metric spaces, then it may be true at one of w* and w but not the other.
that it enforces. The best-system approach was also subject to a number of

complaints regarding the particular resources that it deployed. Other
approaches employing different resources would evade these—but it is
hard to see how one could single out a solution to the Selection Problem
without relying on a bare stipulation or on species-specific considerations.
IV
Primitivist Approaches
[C]haracteristic theoretical identifications like ‘Heat is the motion of

molecules,’ are not contingent truths but necessary truths, and here
of course I don’t mean just physically necessary but necessary in the
highest degree—whatever that means.
Kripke
1 Introduction
Substantivalists formulate truth conditions for claims about the structure of
space at a world in terms of the pattern of geometric relations instantiated
by the points of space of that world. Modal relationalists aim to rely on
facts about which patterns of material configurations are possible in some
distinctive geometric sense in making sense of such claims.
Anyone invoking a notion of possibility faces a choice: take the notion
as primitive or attempt to ground it in other sorts of facts.
Consider the case of nomic possibility. At a minimum, realists about
nomic possibility require the notion of two possible worlds sharing their
laws of nature. It is natural to think of this notion as being encoded in an
equivalence relation N on the space of possible worlds that obtains if and
only if two worlds share their laws. One can then take the laws of nature at
a world w to be those contingent propositions that hold in all of the worlds
N -related to w. Most of the extant philosophical accounts of laws of
nature can be understood as attempts to ground the relation N by
providing necessary and sufficient conditions for N (w1 , w2 ) in terms of
facts about w1 and w2 that are not simply facts about what is nomically
possible at those worlds. Under the best-system analysis, considered in
Chapter III above, the relevant conditions turn on whether certain the-
ories are paragons of theoretical virtue at both w1 and w2 . Under the
necessitarian approach, to be considered in Chapter V below, the relevant
primitivist approaches 79
conditions turn on whether the same fundamental properties are instanti-

ated at w1 and w2 .1
According to best-system analysts, if two worlds agree as to the pattern
of instantiation of the fundamental properties, then they agree as to the
laws of nature. According to necessitarians, if two worlds agree as to which
fundamental properties are instantiated, then they agree as to the laws of
nature. Thus, in the terminology of Chapter II, both approaches are
grounded (the laws supervene on ordinary physical facts)—possibly even
super-grounded (if the relevant ordinary physical facts are taken to be
suitably non-modal).
Consider a world w containing only a single particle, moving inertially,
and let mass be the only fundamental property. What do the laws of w say a
world containing two such particles would be like? To mention only the
tamest options, do the laws of w imply that massive particles move
inertially in multi-particle worlds or that they accelerate towards one
another in such worlds? Presumably ‘everything moves inertially’ is an
axiom in the simplest–strongest system describing w, so under the best-
system approach massive particles always move inertially at worlds gov-
erned by the laws of w. The picture is different under the necessitarian
approach: w features our property of mass and particles with this property
(and no others) accelerate towards one another at any metaphysically
possible world under the necessitarian approach. Grounded approaches
differ from one another as to what the laws of a world like w are—but they
agree that the description of w given above (or one in the same ballpark)
determines the laws of w.
A number of authors object that this claim is intuitively implausible,
since it seems entirely reasonable to think that there are some one-particle
worlds like w governed by the Democritean laws (according to which,
collisions aside, particles always move inertially), others governed by
the Newtonian laws (according to which massive particles attract one
another). 2 Some of these authors also find the other available approaches
1
Under a third approach, which will not be extensively discussed here, two worlds satisfy
the same laws if the properties instantiated at them stand in the same contingent relations of
necessitation to one another. The sources of this approach are Armstrong, What is a Law of
Nature?; Dretske, “Laws of Nature;” and Tooley, “The Nature of Laws.”
2
For this sort of point, see e.g. Carroll, Laws of Nature, §3.1; Fine, “Varieties of Necessity,”
§3; Lange, Laws and Lawmakers, p. 52; Maudlin, The Metaphysics within Physics, §2.3; Wood-
ward, Making Things Happen, §6.11.
80 primitivist approaches
to explicating nomic possibility unacceptable—and find themselves unable

to do without.3 This drives them towards a primitivism about laws of
nature: nomic possibility is taken to be irreducible.4 In other words,
under such approaches N is posited as a primitive equivalence relation
on the space of possible worlds, and it is denied that there is any enlighten-
ing sense in which it can be grounded in ordinary physical facts such as the
pattern of physical properties and relations (the property of being subject
to such and such laws is of course not normally counted as an ordinary
physical fact).
We can think about geometric modality in much the same terms. At a
minimum, realists about geometric possibility will want to recognize an
equivalence relation g on the space of possible worlds that tells us when
two worlds share their spatial geometry (all worlds that do not have spatial
geometry can be lumped into one equivalence class—which we will
henceforth ignore).
Substantivalists have a straightforward means of grounding this relation
in unproblematic geometric facts: g(w1 , w2 ) if and only if the points of
space instantiate the same pattern of geometric relations at both worlds.
But things are more difficult for relationalists.5 Consider two relation-
alist worlds, w1 and w2 , each of which contains just a single point-particle.
Ambition tells us that w1 and w2 need not have the same spatial geometry.
3
The Armstrong–Dretske–Tooley approach, noted above, allows that worlds satisfying
the description of w can differ in their laws, if the universals mass and charge stand in different
relations of necessitation at different worlds. But the notion of a contingent relation of
necessitation has come in for heavy criticism. See e.g. Lewis, “New Work for a Theory of
Universals,” p. 40; and van Fraassen, Laws and Symmetry, §5.3.
4
For primitivist approaches, see e.g. Carroll, Laws of Nature; Fine, “Varieties of Necessity;”
Maudlin, Metaphysics within Physics; and Pargetter, “Laws and Modal Realism.” Lange
advocates a closely related approach under which subjunctive facts rather than modal facts
are taken as primitive; for discussion of the advantages he sees in this see Laws and Lawmakers,
pp. 139–41 and 210 n. 32. Of course, the introduction of N need not be the end of the story
for primitivists—a view counts as primitivist so long as what is added doesn’t amount to an
attempt to ground the nomic in the non-nomic. Maudlin, for instance, adds to the core view
the thesis that laws together with initial states generate histories. See Maudlin, op. cit., esp.
§§1.3 f. and 4.4; for critical discussion, see Loewer, “Time and Law.”
5
Recall that a form of modal relationalism is: grounded if it implies that worlds whose
possibility; metric if it implies that two material configurations are geometric duplicates of one
another if and only if they instantiate the same pattern of distance relations; and (roughly
speaking) ambitious if it implies that for every substantantivalist world w with material
configuration C, there is a relationalist world whose material configuration is a geometric
duplicate of C, and at which the facts about geometric possibility mirror those of w.
Metricity tells us that the material configurations at w1 and w2 are geometric

duplicates of one another (since distance relations are trivial at both
worlds). Groundedness tells us that if the material configurations at two
relationalist worlds are geometric duplicates, then those worlds must agree
about geometric possibility (and hence about spatial geometry).
So if we want the facts about geometric possibility at a world to
supervene on pedestrian geometric facts at that world, we face a choice:
(i) reject ambition; or (ii) reject metricity. As we saw in the preceding
chapter, the most natural way to develop the first option has high costs.
We will see in the next chapter that a natural way of developing the
second option also leads in an unsettling direction (the base on which
geometric possibility supervenes involves properties of an unusual sort).
But, of course, in the geometric context as in the nomic one, it is also
natural to consider abandoning the supervenience thesis—denying that
there is any enlightening story to be told about how the facts about
geometric possibility at a world (including the facts about which worlds
share its spatial geometry) can be grounded in more ordinary geometric
facts. The goal of this chapter is to investigate how the execution of this
manoeuvre allows one to combine metricity and ambition.
Ambitious relationalists accept the following Mirror Principle:
Let w be a substantivalist world with material configuration C. Then
there is a relationalist world w * whose material configuration C *
instantiates the same pattern of geometric relations as C does and
such that the facts about geometric possibility at w * mirror those of w.
Our first task will be to come to grips with precisely what this principle
demands. We can begin with three quick and easy points.
First, note that some claims about geometric possibility at a substantiv-
alist world w will imply the existence of parts of space—e.g., “w has room
for a configuration C 0 that would occupy more points of space than its
actual material configuration C does.” Relationalists are of course only
interested in mirroring claims about material configurations that don’t
imply the existence of parts of space.
Second, note that since we are interested at present in a form of
relationalism which is both ambitious and metric, we should take C and
C * to instantiate the same pattern of geometric relations if and only if they
instantiate the same pattern of distance relations—i.e., if and only if C and
C * correspond to the same metric geometry.
Finally, note that (non-perverse) substantivalists recognize a variety of

non-qualitative geometric properties that relationalists do not recognize.
Suppose that w is Euclidean substantivalist world and that C fills space.
Then there will be many sub-configurations of C that have the structure of
a Euclidean ball of radius one. According to substantivalists, each of these
balls possesses distinct geometric properties: one occupies region R1 of
space, another the adjacent region R2 , etc. But given a relationalist world
in which the material configuration C * has the geometric structure of
Euclidean space, the relationalist will want to say that each unit-ball-
shaped region has the same geometric properties.6 Similarly, if we consider
an empty Euclidean world, substantivalists will say that there are many
ways that we could add one point-particle to this world (corresponding to
different choices of points of space that would be occupied), while rela-
tionalists will insist that (restricting attention to geometric features) there is
just one way to build a one-particle Euclidean world. The point is a
general one: (non-perverse) substantivalists recognize possibilities that are
qualitatively identical (in the sense that the same pattern of basic geometric
relations is instantiated in each) but which are distinct in virtue of the fact
that different regions of space are occupied; relationalists do not do so.7
6
Two unit-ball-shaped regions of matter will of course differ in the identity of the
material points composing them and may also differ in their mass etc., but these will not
constitute a difference in geometric properties.
7
So-called sophisticated substantivalists count possibilities as relationalists do (see e.g. Butter-
field, “The Hole Truth” and Brighouse, “Spacetime and Holes”). We can make this precise as
follows. Consider Euclidean world w containing a single material particle, Fred. And consider
all of the various subjunctive conditionals of the form: “If matter and space were just as they
are except that Fred were located in a different part of space, then . . .” Now evaluate these
conditionals as counterfactuals at w. Clearly, if w is a relationalist world, each of these
counterfactuals is vacuously true—there are no worlds satisfying the shared antecedent of
these conditionals. What if w is a substantivalist world? Non-sophisticated substantivalists will
think that there exist worlds satisfying the antecedent of our counterfactuals—and that when
we use these worlds to evaluate our counterfactuals, we find that some of them are true, some
false. But sophisticated substantivalists dig in their heels at this point. They are counterpart
theorists and do not believe that anything exists in more than one world, including points of
space; there may well be in addition to w other one-particle substantivalist Euclidean
worlds—but at any such world w 0 the only reasonable counterpart for the point of space x
occupied by Fred at w is the unique point of space x 0 that is occupied at w 0 . So when they
evaluate our family of counterfactuals, sophisticated substantivalists find (as relationalists do)
that their common antecedent is false at every world—so that each of the counterfactuals in
this family is vacuously true. This approach has some strange consequences. For if we are this
strict about counterpart relations, we will think that the counterpart of x is occupied at just
about any Euclidean world at which there is matter (exceptions include: world with rotational
So our Mirror Principle assumes the form:

MP Let w be a substantivalist world with material configuration C.
Then there is a relationalist world w * whose material configur-
ation C * is isometric to C and such that every claim concerning
(only) geometric features of material configurations geometric-
ally possible at w is also true at w * .
This is still a bit vague. But it is sharp enough to allow us to proceed. The
next section introduces a very simple primitivist approach, which aims
only to capture facts about which worlds share the same spatial geometry.
In Section 3 below, it is observed that substantivalists in general need to
recognize qualitative facts about geometric possibility that capture infor-
mation beyond the structure of space. Section 4 argues that relationalists
should do likewise, and develops a primitivist strategy that lives up to this
standard. In Section 5 it is argued that primitivists should view metricity as
an unmotivated assumption and some remarks are made concerning how
the general approach of the earlier sections looks when this assumption is
dropped. A final section surveys the terrain traversed. For convenience,
I continue to focus discussion on static relationalist worlds.
2 A First Primitivist Approach

Let us for the time being assume that the only facts about geometric
possibility that substantivalists recognize are facts about which material
configurations are geometrically possible at given world. Under this
assumption, the Mirror Principle becomes:
MP1 For each substantivalist world w with material configuration C
there is a relationalist world w * with material configuration C * ,
such that C and C * instantiate the same pattern of distance
relations and w and w * agree about which patterns of distance
relations among material points are geometrically possible.
symmetry about an unoccupied point, and small perturbations of such worlds). This attributes
to x a strong and interesting modal property lacked by the other points of space at w: it is,
roughly speaking, a necessary truth concerning x that it is occupied if any point of space is.
And this seems at odds with thinking that space was Euclidean (and hence homogenous) at w
in the first place. I am indebted here to §2.2.1 of Skow, Once Upon a Spacetime and to
discussions with Oliver Pooley.
Let us say that a region in a substantivalist world w is a set of points of space

of w equipped with their distance relations. Substantivalists should accept:
(y) A pattern of distance relations corresponds to a geometrically
possible material configuration at w if and only if there is a region
in w instantiating that pattern of distance relations.
We can construct a framework for modal relationalists tailor-made for MP1
and (y) as follows.8 Since we are at present concerned with metric relation-
alism, we take the basic geometric facts at a relationalist world to be given by
a specification of the distance relations between material points. Let us take
as primitive the notion of a full relationalist world—one in which matter fills
all space—and an accessibility relation, g, on the space of relationalist possible
worlds. g is an equivalence relation which is to be interpreted as telling us
when two worlds share the same spatial geometry. Let us call a metric space
admissible if it corresponds to a possible structure for space.
We lay down the following postulates:
A1 For each admissible metric space (X, d), there is a full relationalist
world in which matter instantiates the metric geometry of (X, d).
A2 Full worlds share their geometry if and only if their material
configurations instantiate the same metric geometry.
A3 Each non-full relationalist world shares its geometry with a full
world.
A4 If X represents the geometry of a full relationalist world wf , then
for any subspace X1 X there is a relationalist world w that
shares its geometry with wf and whose material configuration
instantiates the metric geometry of X1 .
Let us further add a couple of natural principles governing the notion of
geometric possibility in this context.
(i) If two relationalist worlds share their geometry, then they agree
about which patterns of distance relations correspond to geo-
metrically possible material configurations.
(ii) At a full relationalist world w, a pattern of distance relations
corresponds to a geometrically possible material configuration if
and only if it is in fact instantiated by a set of material points at w.
8
A related strategy is discussed in in §1.9 of Skow, Once Upon a Spacetime.
This framework delivers what we wanted. Let w be a substantivalist world

with material configuration C. Then there is a full relationalist world wf
whose material points instantiate the same metric geometry instantiated by
the points of space of w. Further, some set of points at wf instantiates the
same pattern of distance relations as C does at w. So there is a relationalist
world w * whose material configuration is isometric to C and which shares its
geometry with wf . And w, wf , and w * agree about which patterns of distance
relations correspond to geometrically possible material configurations.
This is just about the most spare form of primitivism imaginable: we
have taken as primitive the notion of a full world and an accessibility
relation on the space of relationalist worlds that tells us when two worlds
share their geometry; and supplemented these by some assumptions guar-
anteeing that there are enough relationalist worlds to suit ambitious
relationalists and by some reasonable-looking rules for talking about
geometric possibility in this framework. Relationalists adopting this
approach can successfully mirror substantivalist claims about which mater-
ial configurations are possible. In particular, for any pattern of distance
relations, the class of relationalist worlds in which matter instantiates that
pattern includes worlds in which space has every structure consistent with
containing a pattern of this sort. Even more particularly: for every (non-
trivial) admissible spatial geometry, primitivist relationalists of this sort
recognize one-particle worlds with that geometry.
As we will soon see though, this is not yet enough to satisfy any
reasonable ambitious relationalist. In the next section we look a bit more
carefully at what sorts of qualitative geometric facts there are at substan-
tivalist worlds.
Remark 2.1 (Fullness). Was it really necessary to take the notion of a full
world as primitive in the above discussion? Consider a full world whose
material geometry is given by the metric geometry of the set R of real
numbers. Let w1 and w2 be non-full worlds that share their spatial geom-
etry with wf , with the material geometry of w1 being given by the metric
geometry of the integers and that of w2 being given by the metric
geometry of the positive real numbers. Then wf contains sets of material
points that instantiate the material geometry of w1 and of w2 . But neither
of w1 and w2 includes a set of points isometric to the other or to the
material geometry of wf . This sort of example might be taken to suggest
the following approach to defining fullness: take a world w to be full if the
metric geometry of any world sharing its spatial geometry with w is

isometric to a subconfiguration of the material configurations of w. But
this won’t work. Let w 0 be a full world with the spatial geometry of the
positive real numbers and w 00 be a full world with the spatial geometry of
the non-negative real numbers. Let c0 be the set of metric geometries
instantiated by material configurations at worlds sharing their spatial
geometry with w 0 and let c00 stand in the same relation to w 00 . The strategy
under consideration attempts to single out geometry of the positive real
numbers as corresponding to the full worlds with metric geometry in c0
and the geometry of the non-negative real numbers as corresponding to
the full worlds with metric geometry in c00 by looking at relations of
embeddability among these classes of metric geometries. But this cannot
be done. Presumably, c0 includes all and only those metric geometries
corresponding to subsets of the positive real numbers while c00 includes all
and only those metric geometries corresponding to subsets of the non-
negative real numbers. But for every subset of the positive real numbers,
there is a subset of the non-negative real numbers with the same metric
geometry, and vice versa. So c0 ¼ c00 . So just knowing the class of metric
geometries instantiated at worlds sharing their geometry with a given full
world does not suffice to determine the geometry of that full world. &
3 Geometric Facts for Substantivalists

One specifies all of the geometric facts holding at a substantivalist world w
by specifying two sets of facts: (i) facts concerning the geometric relations
that obtain between the points of space at w; and (ii) facts concerning
which of these points are occupied by matter.
Suppose that instead of specifying all of the geometric facts holding at w,
one aims to specify just the qualitative geometric facts about w—the facts,
that is, that concern the pattern of relations instantiated but which are
blind to the identities of the relata. One part of the job is straightforward:
one specifies all of the qualitative geometric facts about space itself by
specifying a metric space that encodes the geometry of space (or, equiva-
lently, by specifying the corresponding metric geometry). But having done
this, it remains to capture the qualitative geometric facts about matter—to
specify all of the qualitative geometric features of the region of space
occupied by matter. This second task turns out to be somewhat tricky.
If this comes as a surprise, it is because our intuitions tend to be

thoroughly Euclidean. One way to specify a region of the Euclidean
plane is to just directly specify the points of the plane that constitute the
region in question. A second method has us proceed in stages, first
specifying the pattern of distance relations instantiated by the parts of the
region; then specifying where and how it is located (say by specifying the
location of an arbitrary point in the region, then specifying the orientation
of the region via reference axes originating from this point). For present
purposes, this second method has the advantage that it factors information
about our region into two parts: intuitively speaking, at the first stage one
specifies the shape of the region, at the second, its location and orientation
in space. And of course in specifying the first sort of information one
specifies precisely the qualitative geometric facts concerning the region: in
Euclidean geometry, regions are qualitatively indistinguishable if and only
if they have the same shape—location and orientation are irrelevant.
But this sort of picture doesn’t hold in general. Consider two equiva-
lence relations on the set of regions of a metric space X.
Regions X1 and X2 are congruent if there is an isometry f : X1 ! X2
(here X1 and X2 are equipped with the metric structures that they
inherit from X).
Regions X1 and X2 are superposable if there is an isometry ç : X ! X
such that X2 ¼ ç(X1 ).
We call the equivalence class of regions superposable with X1 X the
superposability type of X1 and call the equivalence class of regions congruent
to X1 X the congruence type of X1 .9 Congruent regions instantiate
identical patterns of distance relations while superposable regions play
identical roles in the geometry of the ambient space: one specifies the
congruence type of a region by specifying the pattern of distance relations
instantiated by its points; one specifies the superposability type of a region
by specifying all of the qualitative information about how it sits inside the
ambient metric space X.
Clearly, superposable regions are also congruent.10 In the Euclidean
plane the converse also holds: if two regions are congruent, then they are
9
By extension, we also apply these terms to pairs of regions in distinct but isometric
metric spaces.
10
For references and for further discussion of the material through the end of this section,
see Appendix E below.
superposable. The same holds true in a number of spaces, including the

elementary geometries of any dimension; spaces with the metric structure
of the integers or rational numbers; and finite metric spaces with the
discrete metric. In spaces of this kind, in which the congruence of regions
guarantees their superposability, one specifies all of the qualitative features
of a region (such as the region occupied by matter) by specifying the
distance relations that hold between its points.
But this is a very special feature. A generic metric space will feature
congruent regions that are not superposable. In such spaces, specifying the
distance relations between the points of a region will in general fall short of
determining the qualitative geometric features of the region, since it can
fail to distinguish between regions that play distinct geometric roles despite
being congruent.
It is worthwhile to have in mind some of the reasons why congruence
may fail to guarantee superposability in a given space.11
Recall that a metric space X is called homogeneous if for any two points x,
y 2 X, there is an isometry f : X ! X such that y ¼ f (x). In a homogeneous
space, every point plays the same geometric role as every other.
But generic metric spaces are inhomogeneous. In an inhomogeneous
metric space X there exist x, y 2 X that are not related by an isometry of X:
the corresponding one-point regions are qualitatively distinguishable—
they play different roles in the geometry of X. Thus, if we consider a three-
point metric space in which the distances correspond to the side-lengths of
a scalene triangle, then we know that the three points play distinct roles.
Similarly, if we begin with the surface of a sphere, then gently dent or
bulge the surface in several places, but continue to measure distances along
curves lying in the resulting surface, then the metric space we end up with
will be an inhomogeneous Riemannian manifold, in which each point
plays a different role in the geometry of the resulting space.
Of course, there are also inhomogeneous spaces in which some points
are related by isometries of the space as a whole. Consider, for example, a
three-point metric space in which the distances correspond to the side-
lengths of an isosceles triangle: two of the three points are interchangeable
with one another, but neither of these is interchangeable with the third; so
we have three points, but two geometric roles. Or, again, consider the
11
For convenience, the following discussion focuses on finite metric spaces and on
Riemannian spaces.
punctured plane (the Euclidean plane R2 with the origin deleted). In this
space, two points are related by an isometry if and only if they lie at the
same distance from the origin—so we have continuum-many geometric
roles, each corresponding to continuum-many points.
If we are told to select a single point in a homogeneous space there is, as
far as the qualitative facts go, only one way to proceed. But if we are told
to select a single point in an inhomogeneous space, there are, even
qualitatively speaking, multiple choices available. So if the geometry of
space is inhomogeneous at a world, we do not necessarily manage to
specify all of the qualitative facts about that world’s material configuration
by specifying only the distance relations between material points—in
particular, in the case in which we are interested in a one-point configur-
ation, the (trivial) facts about distances between points of the configuration
always fall short of determining the geometric role of this single point, and
must be supplemented by facts about distances to landmarks (distinguished
points or regions or singularities of the space).
The homogeneity of a space tells us that all of its points play the same
geometric role. In the setting of Riemannian manifolds, it also makes sense
to speak about the directions one might proceed in from a given point.
A Riemannian manifold is isotropic if all directions are equivalent at each
point.12 Isotropy too, is a necessary condition for congruence to guarantee
superposability.13 The cylinder, for instance, is a homogeneous but aniso-
tropic space: at any point, there is a big difference between the direction that
points along the axis of the cylinder and the direction perpendicular to this
axis. And if we merely know the distance between two points in the cylinder,
we do not know all of the qualitative geometric facts about the region that
they form—in particular, we do not know whether the line that these two
points determine is a circle (perpendicular to the axis of the cylinder), or
straight (parallel to the axis), or spirals around the cylinder (skew to the axis). So
there are distinct geometric roles open to a two-point set with a given distance
between the points—and the same will be true in any anisotropic space.
Further, every Riemannian manifold in which every pair of congruent
regions is superposable is a space of constant curvature: roughly speaking,
12
i.e., for any point, and for any two unit tangent vectors at that point, there is an isometry
of the space that fixes the point and maps one of the two tangent vectors to the other.
13
Note that in the Riemannian setting, isotropy is equivalent to the property that any
isometry between two-point sets can be extended to an isometry of the entire space. So
isotropy implies homogeneity for Riemannian spaces.
the idea is that if arbitrary congruent three-member sets of points of a

Riemannian manifold are superposable, then one can show that at every
point in the space, every two-dimensional family of directions at that point
is equivalent to every other such family at that point—and this latter
condition is equivalent to being a space of constant curvature.
These necessary conditions are not jointly sufficient, however. The
elliptic plane is a homogenous and isotropic space of constant curvature:
but for any cardinality 3 # n # c, one can find congruent but non-
superposable n-point regions of the elliptic plane (similar results hold for
higher-dimensional elliptic spaces). In the elliptic plane, as in the Euclid-
ean plane, the location and orientation of a region are irrelevant to its
qualitative features. But in the elliptic case, unlike in the Euclidean case,
specifying the distance relations between its parts can fail to specify the
shape of a region—e.g., being told that three points are pair-wise separated
by 3 leaves open whether or not the points are collinear in an elliptic plane
in which all lines have length . So when space is elliptical, knowing the
distance relations between material points can leave us far short of know-
ing all of the qualitative geometric features of the disposition of matter.
But elliptic spaces are the only freaks. For a Riemannian manifold has
the feature that congruence of regions guarantees their superposability if
and only if it is a homogeneous and isotropic space of constant curvature
that is not an elliptic space. The n-dimensional elementary geometries
(Euclidean, hyperbolic, and spherical) are the only Riemannian manifolds
that satisfy these conditions.
4 Another Primitivist Strategy

Let us call two worlds materially congruent if their respective complete material
configurations instantiate the same pattern of distance relations, spatially
congruent if their geometry can be represented by the same metric space.
Substantivalists certainly think that two worlds can be materially but
not spatially congruent, in virtue of disagreeing as to which patterns of
distance relations are instantiated by points of space. The primitivist
strategy of Section 2 above was designed to allow metric but ambitious
modal relationalists to follow suit in recognizing a distinction between
material and spatial congruence—by first rephrasing substantivalist
claims about the structure of space as claims about what sorts of patterns
of distance relations can be materially instantiated, then laying down
postulates to ensure the existence of a rich range of such facts at typical

relationalist worlds.
But the observations of Section 3 show that substantivalists can take two
worlds to be spatially and materially congruent without taking them to
agree even in all of their qualitative geometric facts—because facts about
distance between its points may fail to fix the superposability type of the
region of space occupied by matter. In the cases considered above, the
additional information required to fix all of the qualitative geometric facts
took the form of information about the distances between points in the
occupied region and certain landmarks in the space or information about
the nature of the straight lines connecting occupied points. Presumably
this information might take quite different forms in wilder cases.
As noted above, substantivalists will agree (albeit with puzzlement) that
assertions about which patterns of distance relations are instantiated by sets
of points of space can be replaced by assertions about which patterns of
distance relations correspond to geometrically possible material configur-
ations. They will likewise agree that claims about whether or not the set of
occupied points of space at some world is or is not a subset of a larger set
of points of space with some interesting geometric feature can be replaced
by claims about whether or not it is geometrically possible to extend the
material configuration to a larger configuration that has the interesting
geometric feature in question. Indeed, it sounds quite natural to say that if
we want to know all of the qualitative facts about a world in which space
has the structure of a two-dimensional cylinder and whose material
configuration consists of just two points, we need to know how a line
constructed through those two points would behave—and it does no
harm to think of construction as a physical/material process here. Simi-
larly, if space has the structure of a three-point metric space with distances
between points of one, one million, and one million units, then it sounds
natural to say that in specifying that space contains only a single material
point, one has not yet specified all of the qualitative geometric facts until
one says whether if there were a second material point it would be certain
to be very far from the first one.14
14
One might also reasonably want to handle certain sorts of information that is non-
qualitative in virtue of depending on the identities of the material points involved. For
instance, in three-point space just discussed, the ambitious relationalist could well want to
say that there is only one way in which two particles, a and b, could be located nearby one
another, but that there are two ways that these two particles could be located far away from
In Section 2 above, we assumed for convenience that substantivalists

recognized only one sort of claim about the geometric possibility of
patterns of distances among material points: claims to the effect that a
certain pattern was or was not geometrically possible. But now we see that
there is a second class of claims that is of interest to substantivalists: claims
to the effect that a given material configuration could (or could not) arise
by supplementing the actual material configuration by further material
points. It seems mandatory for ambitious relationalists to take such further
facts into account—surely saying that space has the structure of a cylinder
rather than that of the Euclidean plane brings with it a commitment to say
that there are qualitatively distinct ways to place two points a unit apart in
this space. So the Mirror Principle becomes:
MP2 For each substantivalist world w with material configuration C
there is a relationalist world w * with material configuration C * ,
such that C and C * instantiate the same pattern of distance
relations and w and w * agree about which patterns of distance
relations among material points are geometrically possible and
about what sorts of possible material configurations constitute
geometrically possible ways of extending their respective material
configurations.15
There is a relatively straightforward approach that allows relationalists to
live up to MP2.16 As we did under the approach of Section 2, we take a
relationalist world to consist of a set of material points standing in distance
relations to one another. We again introduce an accessibility relation g on
the space of worlds—but now the intuitive meaning of g(w1 , w2 ) is that w2
arises by adding some matter to the material configuration of w1 , without
changing the geometry of space.17 The relation g is transitive, irreflexive,
and asymmetric. Fullness is now a defined notion: a world w is full if there
is no world w 0 such that g(w, w 0 ). We lay down the following postulates.
one another—the situation in which if a third particle were present it would be near a and far
from b is different from the situation in which if a third particle were present it would be near
b and far from a. Note that traditional substantivalists would want to recognize two possibil-
ities here in the case in which a and b were near to one another as well as in the case in which
they were far from one another.
15
The import of the final clause is: for any material configuration C1 that extends C, we
require there to be a material configuration C1 that extends C* such that there is an isometry
from C1 to C1 that maps C on to C* (and vice versa, with the roles of C and C* swapped).
16
The following way of setting things up was suggested to me by Cian Dorr.
17
This approach relies on a notion of trans-world identity for material points. See Remark
5.2 below for an alternative scheme that avoids this feature.
B1 For every admissible metric space X, there exists a full relationalist

world whose material geometry is represented by X.
B2 If g(w1 , w2 ) then the material points of w1 exist at w2 and instantiate
the same distance relations (among themselves) at both worlds.
B3 For every non-full relationalist world w there is a full relationalist
world wf such that g(w, wf ).
B4 If w has material configuration C and is g-related to two full
worlds, w1 and w2 , then there is an isometry from the set of
material points of w1 to the set of material points of w2 that sends
each point in C to itself.18
B5 For any full world wf and any proper subset C of the material
points at wf , there is a relationalist world w whose material points
are the points of C, with g(w, wf ).
B6 If w1 , w2 and wf are relationalist worlds such that wf is full, w1 and
w2 are g-related to wf , and all the material points of w1 exist at
w2 , then g(w1 , w2 ).
Now we add a couple of natural principles governing the notion of
(i) At any relationalist world w, a pattern of distance relations cor-
responds to a geometrically possible material configuration if and
only if it is instantiated by a set of material points at a full world
identical to or g-related to w.
(ii) At any relationalist world w with material configuration C, the
possible extensions of C are given by the complete material
configurations of the worlds g-related to w.
This gets us what we wanted. Let C be a material configuration in a
substantivalist world w. Then there is a full relationalist world wf whose
material points instantiate the same geometry as do the points of space of w.
And there is a region C * in wf that plays the same qualitative geometric
role there that C plays in w. And there is a relationalist world w * whose
material points are just those of C * (standing in the same distance relations
to one another in w * as they do in wf ). At w * the geometrically possible
18
In particular, the material points of the full worlds instantiate the same metric geometry,
with C corresponding to the same superposability type at each full world.
patterns of distance relations are given by the patterns instantiated in wf —

which are the same as the patterns instantiated in w. So w and w * agree
about which patterns of distance relations are possible. They also agree
about which types of extensions of C and C * are possible. Let C1 be a
material configuration at w that extends C, and let ç be an isometry from
the material points of w to the material points of wf that maps C to C * .
Then C1* : ¼ ç(C1 ) is a region of wf that contains C * . Let w1* be the world
g-related to wf whose material configuration is C1* . Then w * is g-related
to w1* and so C1* is a possible material extension of C * at w * . Further, the
restriction of ç to C1 is an isometry from C1 to C1* that maps C on to C * .
So MP2 is satisfied.
It is tempting to think that for substantivalists all qualitative facts about a
material configuration are captured by facts about ways of extending that
material configuration—in the sense that two regions in a given world are
superposable if and only if they underwrite the same facts about possible
extensions. I suspect that some suitable, sharp version of this claim is true.
Here it is worth noting that in some cases the facts involved are a bit subtle.
Take the example of the punctured plane (the result of deleting the origin
from the Euclidean plane R2 ). In the punctured plane, the one-point
regions C1 ¼ {( 1, 0)} and C2 ¼ {( 2, 0)} are not superposable be-
cause they lie at different distances from the distinguished “missing point.”
The simplest way to cash this out in terms of extensions of configurations
would appear to be: C1 has an extension, C1* ¼ {(x, 0): 1#x < 0}, that
has the structure of a half-closed interval of unit length and that cannot
itself be extended to a configuration isometric to a closed interval of length
one, while C2 has no such extension. So in order to handle even this
simple case, we had to quantify over extensions of extensions.
Thus there is room to wonder whether facts about geometrically
possible extensions of material configurations capture all of the relationa-
listically acceptable information encoded in knowledge of a configuration
at a substantivalist world. If they fail to do so, there is room to ask whether
MP2 yet takes into account all of the facts about geometric possibility that
substantivalists have at their disposal. This looks like a tricky question—but
we needn’t worry about it too much, since it is clear that the approach
developed above equips relationalists with as much material to work with
as substantivalists: in effect, for every substantivalist world w with material
configuration C, the relationalists have a world w * with material config-
uration C * that picks out the same superposability type in the associated
full world wf that C picks out in w. So relationalists of the type under

consideration can match substantivalists move for move, so long as
substantivalist facts about geometric possibility supervene on qualitative
geometric facts (together, if you like, with facts about the identities of
material points).
5 Beyond Metricity
My chief concern about the primitivist proposals discussed so far is that the
combination of ambition and metricity that they embody cannot easily be
motivated.
Primitivism about geometric possibility is motivated by a conviction
that, in general, the facts about geometric possibility at a relationalist world
do not supervene on the base of ordinary geometric facts at that world.
Metricity tells us that all intrinsic geometric facts at a world are fixed by
facts about distances. What kind of justification could a fan of the proposals
discussed above give for metricity?
This question is a live one. In the (small) literature on geometric
possibility, one can find authors who take for granted that the super-
venience base of interest consists of facts about non-quantitative relations
as well as authors who take metricity for granted.19 So it is not as if
advocates of metricity can reasonably claim that the burden of proof
clearly lies with their opponents.
Nor is it reasonable to claim that the mere fact that we are taking
possible spatial structures to be representable by metric spaces establishes
metricity, any more than this fact establishes that the fundamental geo-
metric relations at substantivalist worlds are distance relations (see the
discussion of Section 5 of Chapter I above).
If metricity seems innocuous, this is presumably due to our Euclidean
upbringing. For consider the question which we should take as our super-
venience base at a given world, facts about quantitative distance relations or
facts about non-quantitative relations such as congruence and collinearity. In
the setting of the elementary geometries (Euclidean, hyperbolic, spherical),
congruent regions are always superposable—so specifying the distances
between the points making up a region suffices to specify all facts concerning
19
On this question, cf. Field, “Can We Dispense with Space-Time?” and Huggett, “The
Regularity Account of Relational Spacetime.”
(respectable) non-quantitative relations holding between the parts of the

region. But in general knowing just the facts about non-quantitative rela-
tions comes nowhere close to fixing the distance relations.20 So facts about
distance may seem like the natural starting point for modal relationalists.
Of course, the elementary geometries are very special. But the examples
that come to mind when one tries to think of spaces in which congruence
fails to guarantee superposability are liable to be ones in which distance
relations still call the shots, in the sense that specifying the distances
between the parts of a region suffices to determine its “shape,” while its
“location” is determined by specifying the distance to certain landmarks
(e.g., the singular point in the punctured plane).
But as we have seen above, other situations are possible. Consider again
our example of three points in the unit elliptic plane forming the vertices
of an equilateral triangle of side-length 3. Knowing just the distances
between the three points fails to determine the non-quantitative facts
about them—in particular, it fails to determine whether the three points
are collinear. On the other hand, as in the setting of the elementary
geometries, knowing just the facts about non-quantitative relations typic-
ally fails to determine facts about distance.21
Further, one can concoct examples in which knowing the facts about non-
quantitative relations suffices to determine the facts about distance relations,
but not vice versa. Consider, for example, the set X ¼ {0, 1, 2, 4} with the
metric, d(x, y) ¼ j xy j. Consider the configuration X1 consisting of 0 and 2
and the configuration X2 consisting of 2 and 4. These configurations are
congruent but not superposable in our metric space (since X1 but not X2 has a
midpoint in X ). So distance relations internal to the regions do not allow us to
distinguish between X1 and X2 . But clearly we could characterize the
structure of X in terms of some non-quantitative relations that would allow
us to distinguish between X1 and X2 (say, the eight-place relation R discussed
in Section 5 of Chapter I above and two monadic predicates, understood as
meaning, respectively, that the point has a nearby immediate neighbour and
that the point has a distant immediate neighbour).
20
Euclidean geometry can be axiomatized in terms of betweenness and congruence. But
facts about these relations fail to distinguish between typical three-point sets in Euclidean
space.
21
e.g., facts about congruence and collinearity might tell us that three points form the
vertices of a non-degenerate equilateral triangle, without allowing us to determine the side-
length of this triangle (measured using the natural unit of distance of the elliptic plane).
Reflections of this kind give one reason to be wary of metricity—and

hence give one reason to regard the varieties of primitivism developed
above as unmotivated due to the special role that distance relations play in
them. In effect, the primitivist who accepts metricity takes distance rela-
tions to be the basic geometric relations at every relationalist world, and
introduces a primitive notion of geometric possibility, one of whose jobs is
to entitle one to take non-basic geometric facts to obtain at worlds in
virtue of world-relative facts about geometric possibility (think of the way
in which in our favourite example the invocation of geometric possibility
allows one to distinguish between three-point elliptic worlds in which the
three points are collinear and worlds in which they are non-collinear).
It would be far more natural, it seems to me, to allow the basic facts to
vary from world to world. Let there be some relationalist worlds with the
spatial geometry of the elliptic plane at which distance relations are basic
and other such worlds at which betweenness and congruence are basic
(and, if you like, others at which both quantitative and non-quantitative
relations are basic). At worlds of the first sort, appeal to geometric possibility
secures one facts about non-quantitative geometric relations that are not
implied by facts about the distance relations instantiated; at worlds of the
second sort, appeal to geometric possibility secures one facts about distance
relations that are not implied by the facts about the non-quantitative
relations instantiated.
One way to accomplish this would be via a generalization of the
framework of Section 4 above that allows the basic geometric relations
to vary from world to world.
Remark 5.1 (A Non-Metric Primitivism). We again introduce an accessi-
bility relation g on the space of worlds—but now the intuitive meaning of
g(w1 , w2 ) is that w2 arises by adding some matter to the material config-
uration of w1 , without changing the geometry of space or the family of
basic geometric relations. The relation g is again transitive, irreflexive, and
asymmetric. Fullness is again a defined notion: a world w is full if there is no
world w 0 such that g(w, w 0 ). We lay down the following postulates.
C1 For every admissible metric space X, and every set of relations in
terms of which we can characterize the metric geometry of X,
there exists a full relationalist world whose material geometry has
the given structure and at which the basic geometric relations are
the given relations.
C2 If g(w1 , w2 ) then the set of basic relations is the same at both

worlds and the material points of w1 exist at w2 and instantiate the
same basic relations (among themselves) at both worlds.
C3 For every non-full relationalist world w there is a full relationalist
world wf such that g(w, wf ).
C4 If w has material configuration C and is g-related to two full
worlds, w1 and w2 , then there is an isometry from the set of
material points of w1 to the set of material points of w2 that sends
each point in C to itself.22
C5 For any full world wf and any proper subset C of the material
points at wf , there is a relationalist world w whose material points
are the points of C, with g(w, wf ).
C6 If w1 , w2 and wf are relationalist worlds such that wf is full, w1 and
w2 are g-related to wf , and all the material points of w1 exist at
w2 , then g(w1 , w2 ).
Finally, we add the usual sort of principles governing the notion of
(i) At any relationalist world w, a pattern of basic relations corres-
ponds to a geometrically possible material configuration if and
only if it is instantiated by a set of material points at a full world
identical to or g-related to w.
(ii) At any relationalist world w with material configuration C, the
possible extensions of C are given by the complete material
configurations of the worlds g-related to w. &
Remark 5.2 (An Alternative to Trans-World Identity). The varieties of
primitivism outlined in this and the preceding section rely on a notion of
trans-world identity for material points. Many will view this as an undesir-
able feature of these approaches, either because they are suspicious of the
notion of trans-world identity, or because they think that this notion is out
of place in an account of geometric possibility. It is, however, possible to
22
In particular, the material points of the full worlds instantiate the same metric geometry,
with C corresponding to the same superposability type at each full world. Note that the
material configurations of a full world determine a metric geometry, so it makes sense to speak
of isometries and superposability types here, even if the basic geometric relations at the worlds
under consideration are not distance relations.
construct near-relatives of the approaches developed above that do with-

out trans-world identity.
Note that the suspect notion really performed only one job. The facts
about geometric possibility are as straightforward at full relationalist worlds
as they are at substantivalist worlds. Consider a non-full relationalist world
w with material configuration C. The function of trans-world identity was
to single out a basic relation-preserving bijection from the points of C to
some material configuration C * at a full world w * sharing w’s geometry
so that one could then take the facts about geometric possibility for C at w
to coincide with the facts about geometric possibility for C * at w * .
Here is one way to develop an alternative that foregoes trans-world
identity. As a warmup, consider the notion of congruence. We have been
employing an intra-world four-place relation: C(x1 , x2 ; y1 , y2 ) if and only if
d(x1 , x2 ) ¼ d(y1 , y2 ) with the xi and yi drawn from the same world. There
are three obvious directions in which we could generalize this notion. (1)
We could introduce higher-adicity counterparts of this four-place rela-
tion—so that Cn (x1 , x2 , . . . , xn ; y1 , y2 , . . . , yn ) if and only if the map xi 7!yi
preserves distances. More generally yet, we might collect all of these
relations and their trans-finite analogues into a single variably polyadic
relation c that tells us when there is an isometry from one indexed set of
points to another. (2) We could allow the xi to be drawn from a different
world than the yi .23 (3) We could replace distance by some other set of basic
relations, counting {x1 , . . . } as congruent to {y1 , . . . } if and only if the
map xi 7!yi preserves the appropriate basic geometric relations.
In order to formulate primitivism without trans-world identity, we
work with a trans-world variably polyadic relation s of super-congruence.
Super-congruent configurations always agree with one another as to the
pattern of instantiation of the basic relations. Two configurations at full
worlds are super-congruent if there is a way of indexing their sets of
material points and an isometry from one of these sets to the other that
sends each point to the point in the other world labelled by the same index.
Any region at a relationalist world is super-congruent to a region at a full
world, and any region at a full world is super-congruent to the complete
material configuration at some world. And, of course, super-congruent
configurations agree as to the facts about geometric possibility. &
23
Those who disapprove of trans-world comparisons of distance will want to do every-
thing in a slightly more complicated framework in which what is required is preservation of
ratios of distances.
6 Summation
Should relationalists be satisfied with an approach along the lines of those
sketched above?
For many philosophers, the sticking point will be the reliance on a
primitive modal notion—the invocation of such notions is often taken to
be objectionably obscure or profligate.24 Of course, just how much of a
sticking point this should be depends not only on how objectionable the
proffered primitive notion is, but also on how unobjectionable the alter-
natives are.
Not being myself a man of much metaphysical principle (see Appendix
A below), I have little to say about these questions. But I will note that
primitivism seems to enjoy some benefits over the best-system approach,
in not relying on implausible empirical assumptions about our standards of
simplicity etc., and in allowing for violations of supervenience that accord
with many people’s intuitions.
Note, further, that primitivists about nomic and geometric possibility
may, if they like, build into their approaches a feature that is unavailable to
best-system analysts: they are in a position to countenance the existence of
multiple worlds devoid of matter that differ from one another as to the
facts about physical or geometric possibility.
Primitivists about laws sometimes assert that there could be two worlds
that: (i) were each empty of matter; (ii) were duplicates of one another as
far as their non-material contents go; yet (iii) differed in their laws and so
differ, e.g., as to how massive point particles would move.25 Primitivists
about geometric possibility may be tempted to follow suit. General rela-
tivity encompasses a plethora of worlds empty of matter and energy that
differ from one another as to their spacetime geometry. And this provides
some motivation for modal relationalists about spatial geometry to ac-
knowledge the existence of matter-free relationalist worlds differing as to
the facts about what sorts of material configurations are geometrically
possible.26
24
See e.g. Dorr, “Finding Ordinary Objects in the World of Quantum Mechanics,” §2;
Loewer, “Time and Law;” or van Fraassen, Laws and Symmetry, ch. 4.
25
See e.g. Carroll, Laws of Nature, p. 64 fn. 4; and Fine, “Varieties of Necessity,” §3.
26
For discussion of the force of this last point, see e.g. Butterfield, “Relationism and
Possible Worlds,” §2 and Brighouse, “Incongruent Counterparts and Modal Relationism,”
§3.3.
Of course, primitivists about geometric possibility face a special diffi-

culty here. In the nomic case it is natural enough to say that there are
worlds devoid of matter—since, e.g., the existence of such worlds is a
consequence of substantivalist interpretations of many physical theories.
But does it really make sense to speak of a possible world at which neither
matter nor regions of space(time) exist? And if it does, does it further make
sense to speak of there being multiple such worlds, differing as to modal
facts? What, exactly, is supposed to be the bearer of the properties that
differ between such worlds?
What sorts of responses to these challenges are available to primitivists
about geometric possibility? One possibility would be to say that even at
empty worlds space exists—although, not of course in the sense in which
substantivalists take space to exist. It is not obvious to me whether this is a
viable response.
Alternatively, one could adopt a certain subtle view about mereology.
Above I have spoken loosely, identifying material configurations with sets
of material points. It would have been more orthodox to take the material
configurations at a world to be mereological sums of material points. Note
that under some (rather unpopular) approaches to mereology, there exists
a null individual (analogous to the empty set), which is a part of every
individual at every world, but which itself has no proper parts.27 Modal
relationalists who are willing to posit the existence of the null individual
can recognize worlds at which it is the complete material configuration.28
And primitivists who accept such empty worlds can add that the facts
about geometric possibility need no more be decided by the geometry of
the total material configuration at such worlds than at others.29
27
For discussion and references, see Casati and Varzi, Parts and Places, p. 45; Lewis, Parts of
Classes, pp. 10 f.; and Sorenson, “Nothingness,” §2.
28
Indeed, proceeding along these lines is the most obvious way to avoid commitment to
the claim that it is a necessary truth that something exists. On this question, cf. Lewis, On the
Plurality of Worlds, pp. 73 f. and Bricker, “Island Universes and the Analysis of Modality,”
pp. 47 ff.
29
Relationalists who countenance the null individual and who allow the basic geometric
relations to vary from world to world will have to add another epicycle, so that some empty
worlds are structured by distance relations, others by non-quantitiative relations—even
though no relations are instantiated at any empty world.
V
Necessitarian Approaches
From recent paragraphs it becomes evident not only that the

subjunctive conditional has no place in an austere canonical notation
for science, but also that the ban on it is less restrictive than would at
first appear. We remain free to allow ourselves one by one any
general terms we like, however subjunctive or dispositional their
explanations.
Quine
1 Introduction
Ambitious relationalists maintain that there are many one-particle rela-
tionalist worlds that differ as to the structure of space. To mention just the
tiniest tip of the iceberg, they recognize worlds in which space is Euclid-
ean, and worlds in which it is spherical; worlds in which space is two-
dimensional and worlds in which it is zillion-dimensional.1
Ambitious relationalists who are also grounded hold that when one-
particle worlds differ in their geometries, it is because they differ in the
geometric intrinsic properties instantiated—if two such worlds differ as to
the geometry of space, this can only be because their respective particles
are not geometric duplicates of one another.
The challenge for ambitious, grounded relationalists is to provide an
account of how this can be so. What sort of geometric intrinsic property
1
Recall that, roughly speaking, a form of modal relationalism is ambitious if it implies that
for every substantivalist world w with material configuration C, there is a relationalist world
whose material configuration is a geometric duplicate of C, and at which the facts about
geometric possibility mirror those of w. An approach is grounded if it implies that worlds whose
possibility. An approach is metric if it implies that two material configurations are geometric
duplicates of one another if and only if they instantiate the same pattern of distance relations.
necessitarian approaches 103
does the particle of a one-particle Euclidean world possess that is not

possessed by the particle of a one-particle spherical world?
A response to this challenge is developed below that turns on carrying
over to the geometric context an interesting feature of the so-called
necessitarian account of laws of nature. On that approach, if exactly the
same fundamental properties are instantiated at two worlds, then those
two worlds share their laws of nature.2 It follows that nomic necessitarians
recognize severe restrictions on how entities from distinct worlds can be
combined to form new possibilities. Roughly speaking, these restrictions
encode facts about physical possibility in the identities of the fundamental
physical properties. Translating this feature to the geometric context gives
us the wherewithal to construct an ambitious, grounded account of
geometric possibility. The resulting form of relationalism has the vaguely
Leibnizean feature that a point-particle of a relationalist world has prop-
erties that encode a surprising amount of information about the structure
of the world external to it.
The discussion below takes the following route. Section 2 is devoted to
necessitarianism about the laws of nature. Section 3 isolates the key feature
of nomic necessitarianism that is to be carried over to the geometric case.
Section 4 develops geometric necessitarianism. One of the advantages of
both forms of necessitarianism is that they are grounded. But in both the
nomic and the geometric case it is natural to question whether necessitar-
ianism is super-grounded (i.e., whether it has modal facts supervening on
non-modal facts). This question is taken up in Section 5. For convenience,
I continue to focus attention on static relationalist worlds.
2 Necessitarianism about Laws

Roughly speaking, nomic necessitarians maintain that massive bodies
attract one another at our world not because this world just happens to
be governed by Newton’s law of universal gravitation (or a successor), but
because it is in the very nature of massive bodies to behave this way (and
similarly, of course, for other laws). In this they see an advantage of their
2
In this context, a fundamental property is supposed to be thought of in something like
Lewis’s fashion (see fn. 4 of Ch. III above): as an elite intrinsic property whose instances are
perfectly similar in some metaphyscially special respect.
104 necessitarian approaches
account over those of the best-system analyst, the follower of Armstrong–

Dretske–Tooley, and the primitivist.
It will be helpful to develop this view by way of contrast with a foil, the
views about properties and laws promulgated by David Lewis.
Lewis on Properties and Laws
Recall that Lewis enforces the Humean dictum that there are no necessary
connections between distinct existences via his principle of recombination:
. . . patching together parts of different possible worlds yields another possible
world. Roughly speaking, the principle is that anything can exist with anything
else, at least provided they occupy distinct spatiotemporal positions. Likewise
anything can fail to coexist with anything else.3
Since Lewis denies that an object can exist in two worlds, his official
characterization of the principle of recombination requires that a duplicate
of any object can exist with a duplicate of any other object. It of course
follows from this principle that laws are contingent:
Episodes of bread-eating are possible because actual; as are episodes of starvation.
Juxtapose duplicates of the two, on the grounds that anything can follow anything.
Here is a possible world to violate the law that bread nourishes. So likewise against
the necessity of more serious candidates for fundamental laws of nature—perhaps
with the exception of laws constraining what can coexist at a single position, for
instance the law (if such it be) that nothing is both positive and negative in charge.4
Elsewhere Lewis offers a reformulation of the principle of recombination,

in which he emphasizes that one is allowed to separate and recombine the
roles played by properties:
Possibility is governed by a combinatorial principle . . . We can take apart the
distinct elements of a possibility and rearrange them. We can remove some of
them altogether. We can reduplicate some or all of them. We can replace an
element of one possibility with an element of another. When we do, since there is
no necessary connection between distinct existences, the result will be a possibility.
How much this means depends on what we take the distinct elements to be. Here,
let us take them to include not only spatiotemporal parts, but also abstract parts—
specifically, the fundamental properties.5
3
On the Plurality of Worlds, pp. 87 f.
4
Ibid. 91.
5
“Ramseyan Humility,” pp. 208 f. For Lewis’s notion of fundamental properties, see fn. 4
of Ch. III above.
In this new setting, the point about the contingency of laws takes a new
form: “Let it be a law that every F is a G; combinatorialism generates a law
in which every F is not a G, so that this law is violated.”6 Lewis has in mind
here a scenario of the following sort. Begin with a world w in which three
fundamental physical properties, F, G, and G * , are instantiated and in
which no F is G * but in which every F is G by law. Here G and G * are
both monadic properties. And Lewis’s combinatorialism allows us to swap
the roles of any two fundamental properties or relations of the same
category.7 So there is a possible world w * that differs from w only in that
the roles of G and G* have been switched.8 At w * it is not true that every F
is a G; so the law of w that every F is a G holds only contingently.9
Consider an example. Let w be a decently complex world at which
matter takes the form of point-particles and at which the only non-
geometric fundamental properties are mass and electric charge.10 Let us
further assume that the motions of the particles of w instantiate regularities
in accord with Newton’s “laws” of motion with the forces given by
Newton’s “law” of universal gravitation and by Coulomb’s “law.”11 Let
Pm, q be the set of point-particle worlds at which the non-geometric
fundamental properties are mass and charge. How large is Pm, q ? According
to Lewis, it is very large. It includes worlds at which the particle motions
violate Newton’s laws of motion and worlds at which they comport with
these laws, but only for non-standard forces (maybe the forces go as the
inverse cube of distance, or like charges attract while opposite charges
repel). It also includes a world w * which differs from w only quiddistically:
the particle motions at w and w * are identical, but the mass and absolute
6
“Ramseyan Humility,” p. 209.
7
Examples of categories: monadic properties; n-adic relations; magnitudes (these can be
scalar-valued, tensor-valued, etc.); relational magnitudes. See ibid. 205.
8
Such a pair of worlds is said to differ quiddistically, by analogy with pairs of worlds that
differ haecceitistically in virtue of being qualitatively identical but differing as to which
individual plays which role.
9
For this sort of example—and for an account of why Lewis accepts quiddistic differ-
ences but rejects haecceitisitic differences—see Lewis, “Ramseyan Humility,” §4.
10
Here and throughout this discussion, fundamental non-geometric properties are taken
to correspond to determinable quantities (mass, charge, spin), rather than to particular values
of such quantities. There are some subtle issues in the neighbourhood concerning which roles
of Lewisian fundamental properties can be fulfilled by determinables; see Hawthorne,
“Quantity in Lewisian Metaphysics.”
11
The scare quotes emphasize that I have not said anything about the laws at w, just about
regularities.
value of charge of each particle at w * are equal to the absolute value of

charge and mass, respectively, of its counterpart at w.12
One feature of Lewis’s view is that it supports very few necessary truths
about any given property. But, crucially, there are a few such truths—for
instance, that mass is always representable by an assignment of non-
negative numbers to objects and that no object-stage is ever assigned
two masses.
Nomic Necessitarianism
In the contemporary literature, necessitarianism about laws is arrived at via
two main routes. The first traces out the consequences of taking properties
to be individuated by their causal powers—so that if it is a law at a certain
world that being F has the capacity to cause something to be G under
given conditions, then the same law must hold at any world with the
appropriate cast of characters.13 The second takes as its point of departure
the Armstrong–Dretske–Tooley account of laws, under which laws are
understood as contingent relations of necessitation between universals, but
amends that account by taking the relations of necessitation between
universals to themselves be necessary, in virtue of being rooted in the
identities of the properties involved.14
Of course, each of these routes has a controversial point of departure.
But we can side-step some controversy by focusing on the common core
of necessitarian approaches: the view that the identity of a fundamental
property is so tightly bound up with the nomic role of that property that
two worlds can differ in their laws only if they differ as to the fundamental
properties instantiated.15
12
So expressions for the regularities at w* arise by everywhere swapping m and jqj in
expressions describing the regularities at w.
13
For this approach, see Hawthorne, “Causal Structuralism;” and Shoemaker, “Causality
and Properties” and “Causal and Metaphysical Necessity.” Under a related approach, appeal is
made to dispositions or propensities in addition to (or in place of ) causal relations; see Bigelow
et al., “The World as One of a Kind;” Bird, Nature’s Metaphysics, ch. 3; and Ellis, “Causal
Powers and Laws of Nature.”
14
See Fales, “Are Causal Laws Contingent?;” Swoyer, “The Nature of Natural Laws;”
and Tweedale, “Armstrong on Determinable and Substantial Universals.”
15
For further positions along these lines, see Kneale, Probability and Induction; and Sellars,
“Concepts as Involving Laws and Inconceivable Without Them.” For the early modern roots
of nomic necessitarianism, see Ott, Causation and Laws of Nature in Early Modern Philosophy,
§§13–22.
Consider how this would look in application to the mass-charge point-

particle world w considered in the discussion above of Lewis on properties.
Any approach to laws will be committed to the following claim: (0) one is
entitled to conclude from the description of w that the laws there are given
by Newton’s laws of motion, Newton’s law of universal gravitation, and
Coulomb’s law. Necessitarians add two characteristic theses: (1) the laws at w
obtain in virtue of facts about the properties mass and charge; (2) any world
at which mass and charge are the non-geometric fundamental properties
shares its laws with w. We can rephrase (2) as follows: the set of worlds at
which the law-proposition of w is true is just the set Pm, q (the set of worlds at
which mass and charge are the non-geometric fundamental properties).
More generally, we can characterize necessitarianism as the conjunction
of three theses.
N0. Physics is admirably adapted to discover the laws of nature at
worlds like our own.
N1. The laws of nature at a world are made true by facts about the
non-geometric fundamental properties at that world.
N2. The laws of nature at a world are given by a proposition
corresponding to the set of worlds at which fundamental prop-
erties instantiated exactly match those of the given world.
Three quick remarks by way of clarification. (a) Necessitarians are likely to
view N1 as fundamental and N0 and N2 as derivative—and advocates of
competing approaches are liable to view necessitarians as having an obli-
gation to argue for N0 rather than simply assuming it. The details (and
prospects for success) are liable to depend on how N1 is understood. (b)
Some necessitarians found N1 on sweeping claims about the nature and
individuation of properties. This raises the question whether the identity
of geometric properties and relations are also to be understood as bound up
with laws of nature.16 (c) As formulated here, necessitarianism envisions no
particular connection between the laws of two worlds, one of whose set of
fundamental properties is a proper subset of that of the other. This isn’t
very satisfying—it seems obvious that if every mass-and-charge world is
16
For discussion of this question, see Bird, Nature’s Metaphysics, ch. 7; and Ellis, “Response
to David Armstrong,” §2. For the remainder of the present discussion of necessitarianism
about laws of nature, it should be taken for granted that in speaking of fundamental
properties, I have in mind only non-geometric fundamental properties.
governed by the Newtonian laws and Coulomb’s law, then any world at
which mass were the only fundamental property would be a world
governed by the Newtonian laws. But this seems obvious only because
mass plays a special role in the Newtonian laws of motion—What would
the laws be at a world at which charge was the only property instantiated?
It is not obvious in the general case how to pass from knowledge of what
the laws are at a given world w to knowledge of what the laws are at a
world whose set of fundamental properties is a proper subset of the set of
fundamental properties of w.
The positive case for nomic necessitarianism hangs entirely on the
question of what kind of role causal profiles and nomic roles play in the
individuation of properties. Luckily, for present purposes we can set this
question aside.17 Our interest in nomic necessitarianism is limited to
understanding the general contours of the view—how it hangs together
and what sort of resources it has to reply to certain objections. I turn now
to several such issues (but one of the most serious objections is deferred
until Section 5 below).
Whence the Name? In what sense are the laws of nature necessary for
necessitarians? Here it is essential to distinguish between two species of the
genus. Strong necessitarians add to the core commitments N0–N2 the thesis
that the same fundamental properties are instantiated at every possible
world.18 For weak necessitarians, on the other hand, the fundamental
properties instantiated vary from world to world with there being some
worlds at which fundamental properties alien to our world are instantiated.19
Under any form of necessitarianism, the laws at a world w are given by a
proposition that holds at just those worlds that share their fundamental
properties with w. Under strong necessitarianism, the fundamental prop-
erties at each world are just the fundamental properties of the actual
17
Note, though, that many critics of necessitarianism grant that properties are individu-
ated in part by nomic roles. See e.g. Fine, “The Varieties of Necessity,” §2; Loewer, “Humean
Supervenience,” §VI; Lowe, The Four-Category Ontology, ch. 10; and Roberts, The Law-
Governed Universe, pp. 59 and 74.
18
See e.g. Bird, Nature’s Metaphysics; and Shoemaker, “Causality and Properties” and
“Causal and Metaphysical Necessity.”
19
Of course, there is some logical space between strong necessitarianism and weak
necessitarianism: in principle, one could deny that there are any worlds featuring alien
properties while allowing that there are some worlds at which only some of our world’s
fundamental properties are instantiated.
world—so the actual laws are metaphysically necessary. Under weak

necessitarianism, the laws are metaphysically contingent, since worlds
differ in their fundamental properties.20
Nonetheless, it may be felt that nomic necessity has more oomph under
weak necessitarianism than it does under other views on which it falls short
of full metaphysical necessity. For weak necessitarians can say that the laws
hold because they have to, given the natures of the properties involved.
This sounds more impressive than saying that the laws just happen to
hold (as one says under the best-system and primitivist approaches) or that
the laws hold in virtue of relations of necessitation between universals,
which in turn just happen to obtain (as one says under the Armstrong–
Dretske–Tooley approach).21
We can bring this point home as follows. Rather than merely sorting
views into those according to which laws are metaphysically necessary
and those according to which they are contingent, let us try to make
comparative judgements about how close the laws come to being meta-
physically necessary under various views. Consider the following question:
Given two possible worlds, how much information do we have to know
about them in order to determine whether they share their laws? The
answer depends on our account of nomic possibility:
Under strong necessitarianism we need no information.
Under weak necessitarianism we need to know whether the worlds
instantiate the same fundamental properties.
Under the Armstrong–Dretske–Tooley approach, we need to know
whether the worlds instantiate the same fundamental properties and
whether the same contingent relations of necessitation between the
relevant universals obtain at the two worlds.
Under the best-system approach, we need to know not only which
fundamental properties are instantiated at the two worlds, but also
enough about the patterns of instantiation of these properties in order
to determine whether they underwrite the same simplest-strongest
theories.
20
However, the laws come out as metaphysically necessary under some closely related
views. See Remark 2.2 below.
21
For the accusation that the Armstrong–Dretske–Tooley approach is “second-order
Humean,” see Swoyer, “Nature of Natural Laws,” pp. 210 f.
Of course, a view renders the laws of nature metaphysically necessary if and

only if it agrees with strong necessitarianism that no information is required in
order to determine that two possible worlds share their laws. So the infor-
mation required for our task under a given view gives us one natural measure
of how close that view comes to rendering the laws metaphysically necessary:
strong necessitarianism achieves this; weak necessitarianism comes relatively
close; the Armstrong–Dretske–Tooley view not quite as close; the best-
system view nowhere near. Upshot: ‘weak necessitarianism’ is a misleading
name—but in one respect it is not too misleading.
Remark 2.1. Of course, there is also a sense in which laws are no closer to
being metaphysically necessary under weak necessitarianism than under the
best-system approach. Consider some set of sentences suitable to be the
axioms of a strongest-simplest theory: under standard weak necessitarian
views, as under the Lewisian approach, one expects there to be a world
whose law-proposition is expressed by these sentences.22 &
Remark 2.2. There are close relatives of weak necessitarianism under
which: (i) there are worlds with properties other than those of the actual
world; but (ii) there is a sense in which laws of nature are metaphysically
necessary. The trick is to adopt a non-standard view of the logical form
of laws.
One way to manage this is to take there to be a single set of laws shared by
all worlds, with these laws having a conditional form expressed by sentences
like “If the fundamental properties are X, then the constraints on their
instantiation are thus and so.”23 Of course, roughly speaking, adopting this
approach means accepting that each world is governed by a plethora of
vacuous laws—a conclusion that is usually taken to be unattractive, on the
grounds that it is at variance with the pronouncements of physics.24
Another possibility is to take laws to be something like rules that govern
the evolution of instantaneous states (so that laws are non-propositional).25
22
There is a delicate issue here: What if the law sentence implies that there are distinct
properties that fill structurally identical causal roles? For discussion, see pt. 3 of Hawthorne,
“Causal Structuralism.”
23
For an approach of this kind, see Ellis, “Causal Powers.”
24
For worries concerning theories that recognize a plethora of vacuous laws, see Earman,
“Laws of Nature,” §2; Fine, “Varieties of Necessity,” §2; and Roberts, Law-Governed Universe,
§3.3.1.
25
For non-necessitarian approaches that treat laws in this way, see Vallentyne, “Explicating
Lawhood” and Maudlin, The Metaphysics within Physics, ch. 1.
So long as one is only concerned with worlds at which the list of

fundamental properties instantiated does not change over time and one
takes worlds that agree as to the fundamental properties to agree as to the
rule governing dynamical evolution, then one finds that laws are meta-
physically necessary: the laws/rules of the actual world are not violated at
any world (and likewise for the laws/rules of other worlds). &
Troublesome thought experiments. One of the main knocks against

necessitarianism about laws is its prima facie absurdity. Surely there are
possible worlds that share their fundamental properties while differing in
their laws! Indeed, it seems easy to imagine each of the following.
(1) A world in which the attraction between massive bodies goes
inversely as the cube rather than the square of distance.
(2) A pair of worlds w1 and w2 , each containing a single massive particle
in inertial motion, with w1 governed by the familiar Newtonian
laws (so that if a second massive particle were introduced, each
would accelerate) and w2 governed by Democritean laws (so that in
multi-particle worlds, only collisions interrupt inertial motion).
The world in (1) shares its fundamental properties with our world, but
differs in its laws. The worlds in (2) differ from one another in their laws
while sharing not only their fundamental properties, but also a single
pattern of instantiation of these properties.
The standard necessitarian response to examples of this sort is to remind
us that Kripke has shown that we are susceptible to systematic illusions of
imaginability.26 We seem to be able to imagine a world in which the
morning star and the evening star are distinct bodies. But there is no such
world. So at best we imagine a world that might be misdescribed as being a
world in which Phosphorus is not identical to Hesperus. Likewise, says the
necessitarian, it is impossible to imagine a world in which massive bodies
obey a non-standard law of gravity as in scenario (1) above. When we take
ourselves to have done so, either we fail to imagine any world, or we
imagine a world in which the standard law is obeyed but in which observers
are warranted in positing a non-standard law, or we imagine a world in
which attraction really does go inversely with the cube of distance, but in
26
For an especially thorough treatment along these lines, see Shoemaker, “Causal and
Metaphysical Necessity.”
which bodies possess not mass but some alien property, schmass.27 Similarly,
when it comes to scenario (2) either we imagine only one world, or we
imagine two that in fact differ in their fundamental properties.
None of this is very palatable. But presumably it is a small enough price
to pay if one has been driven towards nomic necessitarianism by general
views about the individuation of properties.
The Threat of Weak Laws. Advocates of each of the standard philosoph-

ical accounts of laws of nature recognize as a goal showing that on their
accounts the sciences can be expected to discover the laws at our world.28
This is no easy task—extant philosophical accounts of laws of nature make
it all too easy to worry that the sciences are liable to mislead us. For primitivists
about laws and advocates of the Armstrong–Dretske–Tooley approach this
worry takes the following form.29 Because our world involves a lot of
interesting regularities, scientists tell us that it is governed by interesting, strong
laws. But according to primitivists and followers of Armstrong–Dretske–
Tooley both, for every world full of regularities at which many laws hold,
there is a corresponding Hume world that instantiates exactly the same regu-
larities, but in which the laws are weak or non-existent. Because it has access
to laws only via observed regularities, there is no way that science can tell us
whether we live at a world with strong laws or at a Hume world.
This particular worry doesn’t arise for necessitarians: on their view, two
worlds that instantiate the same regularities will also instantiate the same
fundamental properties—and so will agree in their laws. But it doesn’t take
much work to find worries in the same neighbourhood that do apply to
necessitarianism.
Indeed, under necessitarianism, no matter what the laws are at our
world, there will be worlds that share our laws but at which there are
27
Of course, the third of these option is not open to strong necessitarians.
28
Indeed, the claim that their own approach to nomic possibility enjoys decisive epi-
stemological advantages over its rivals is made by advocates of most of the major approaches to
laws of nature. See e.g. Armstrong, What is a Law of Nature?; Dretske, “Laws of Nature;”
Earman and Roberts, “Contact with the Nomic. Part II;” Foster, “Induction, Explanation,
and Natural Necessity;” Kneale, Probability and Induction; Roberts, Law-Governed Universe;
Shoemaker, “Causality and Properties;” and Tooley, “The Nature of Laws.” For further
discussion, see Loewer, “Laws and Induction.”
29
It takes a rather different form under the best-system approach; see van Fraassen, Laws
and Symmetry, §3.5. For a sophisticated version of the argument sketched here, see Earman
and Roberts, “Contact with the Nomic. Part II” and ch. 4 of Roberts, Law-Governed Universe.
strong regularities that hold accidentally rather than by law. Consider the
mass-and-charge point-particle world w that we have been using as an
example. Suppose that the laws at w are what one would expect:
Newton’s laws of motion with forces given by Newton’s law of univer-
sal gravitation and by Coulomb’s law. According to necessitarians, these
will be the laws at every world at which mass and charge are the
fundamental properties. Now consider a world w * of this kind in
which the particles and their attributes are carefully arranged so that no
particle ever accelerates. Universal non-acceleration is a regularity that
holds accidentally at w * .
What reason do we have to believe that the regularities at our world
that we are inclined to view as laws hold in virtue of facts about our
fundamental properties rather than by accident? For all we have said so far,
there may exist fundamental properties m* and q* from the same categories
as mass and charge whose natures underwrite very weak laws—so weak
that the class of worlds Pm* , q* is in effect as capacious as Lewis takes Pm, q to
be. How could we ever determine that the fundamental properties at our
world are m and q (with their strong laws) rather than m* and q* (with their
dismayingly weak laws)?
Even strong necessitarians face a version of this challenge. Let it be granted
that the same fundamental properties are instantiated at every possible world
and that it follows that the same laws hold at each world. Still, we can
ask what warrant we have for believing that the properties in question
underwrite strong laws (putting substantive constraints on the pattern of
instantiation of properties) rather than weak ones (involving only such
constraints as follow from the logical type of the fundamental properties).
One might be tempted at this point to argue as follows:
Suppose that we find ourselves in a world in which particles never

accelerate. This holds either by law or by accident. In the former case,
it obtains at every world permitted by the laws, in the latter only at a
subset—and, if the laws involved are of the sort we are familiar with,
this subset must be very small indeed. We should assume that our world
is generic rather than special among the worlds permitted by its laws. So
we are warranted in assuming that particles move inertially by law. And
similarly for other regularities. So although we cannot prove that our
fundamental properties support strong laws, we are certainly warranted
in believing that they do so.
The problem with this argument is that it applies to all regularities, even
those too complex and gerrymandered to be considered candidates for
lawhood by physics—there are ever so many such at our world, highly
complicated regularities that we would never dream of taking to be laws.
So the methods of physics leave us with laws much weaker than this line of
thought would suggest. What is needed here is a rationale for taking our
world to have moderately strong laws.
3 Compatibility Properties
Before turning to the geometric case, it will be helpful to isolate an
important feature of nomic necessitarianism: roughly and loosely speaking,
it implies that some intrinsic physical properties encode modal information.
The notion of an intrinsic property is intuitively clear—to say that a
property is intrinsic is to say that whether an object possesses that property
depends only on the object itself, and not on the world external to the
object. There are a number of competing analyses of the notion of
intrinsicality.30 Many of these appear to be serviceable—they agree with
one another and with common intuitions in many cases. But none is
generally taken to be a complete success. I mention two of the more
promising—though nothing that follows depends on the sort of questions
of detail over which competing analyses disagree.
i. P is intrinsic1 if whether a thing instantiates P or not-P cannot be
changed by adding something to its containing world.31
ii. An object is lonely if there is no object at its world wholly distinct
from it, otherwise it is accompanied. A property P is intrinsic2 if neither
P nor not-P imply either loneliness or accompaniment (i.e., there is
a possible world at which an accompanied object has P, a world at
which an accompanied object has not-P, a world at which a lonely
object has P, and a world at which a lonely object has not-P).32
30
For discussion and references, see Weatherson, “Intrinsic vs. Extrinsic Properties.”
31
For an analysis of this sort, see Yablo, “Intrinsicness;” for an alternative that turns on
subtraction rather than addition, see Vallentyne, “Intrinsic Properties Defined.” These
proposals require that objects exist at more than one possible world. For discussion of how
Lewis’s framework can be extended to allow trans-world individuals, see McDaniel, “Modal
Realism with Overlap” and Parsons, “Is Everything a World?”
32
This is a simplified version of the proposal of Langton and Lewis, “Defining ‘Intrinsic’,”
suitable for those who don’t have to muck around with worries about disjunctive properties.
As usual, we say that objects are duplicates if they share all of their
qualitative intrinsic properties.33
We say that object x has the property of compatibility with property P if
there is a duplicate of x that exists in a world at which P is instantiated.
Informally: x is compatible with P if the property of being just like x is
coinstantiated with the property P at some world.
Being such that some property is instantiated at your world is in general
one of your extrinsic properties. But being compatible with the instanti-
ation of a given property is an intrinsic property. For let x be a possible
object and consider the class d consisting of all the duplicates of x at
various worlds. For any property P, either some member of d has P or no
member of d has P. In the first case, every duplicate of x is compatible
with P; in the second case, no duplicate of x is compatible with P. Either
way, duplicates never differ concerning whether they are compatible with
P—which is to say that compatibility with P is an intrinsic property.
This much follows just from the definition of compatibility. What sort
of compatibility properties actual things have will of course depend on the
extent of the space of possible worlds.
Let Bucephalus be an actual horse and Porky an actual pig.34 Then, of
course, Bucephalus is compatible with porcinity and Porky is compatible
with equinity. But the question of their compatibility with the absence of
these properties is more subtle.
Consider first how things look under lavish views of the extent of
logical space, under which something like Lewis’s principle of recombin-
ation obtains. Equinity in Bucephalus is one thing, porcinity in Porky
another. So there will be a possible world containing just a duplicate of
Bucephalus and another containing just a duplicate of Porky. It follows
Their official proposal is: a property P which is neither a disjunctive property nor the negation of
a disjunctive property is basic intrinsic if it satisfies the condition in the text; the intrinsic properties
are those that never differ between objects sharing all of their basic intrinsic properties.
33
This is the most common notion of duplication. (1) Lewis sometimes characterizes
duplicates in terms of perfectly natural/fundamental properties rather than qualitative intrinsic
properties (Plurality, 61). The difference is inessential for present purposes, since Lewis takes
perfectly natural/fundamental properties to be intrinsic and to provide a supervenience base
for the qualitative intrinsic properties (see Langton and Lewis, “Defining ‘Intrinsic’,” §IX and
Lewis, “Ramseyan Humility,” §2). (2) For worries about the standard notions of intrinsicness
and duplication, see Bader, “Towards a Hyperintensional Theory of Intrinsicality.” On
Bader’s account, the argument of the paragraph after next would not go through.
34
Here and below, being an animal of a certain kind requires only having atoms arranged
in a certain fashion (in particular, nothing about origin or history is involved).
that under such views Bucephalus is compatible with the absence of

porcinity, Porky with the absence of equinity.
At the other extreme, we have the hyper-necessitarian view according
to which the actual world is the only possible world. Under this view,
nothing actual has a duplicate at a world that differs in any way from the
actual world. It follows that under this view Bucephalus is incompatible
with the absence of porcinity and Porky is incompatible with the absence
of equinity.
There is of course plenty of territory intermediate between these ex-
tremes. Under nomic necessitarianism whether an actual object is compat-
ible with the absence of an actually instantiated property depends on the
precise form of the laws of nature. Bucephalus is compatible with the absence
of porcinity only if there is a world with a duplicate of Bucephalus at which
porcinity goes uninstantiated—and for nomic necessitarians, whether or not
such a world exists depends on what the laws of physics are at the actual
world. If, for instance, it should turn out to be a surprising consequence of
quantum mechanics that atoms cannot be arranged horse-wise at a world
unless some other atoms are arranged pig-wise elsewhere at that world, then
Bucephalus would not be compatible with the absence of porcinity.35
Now, under the views lying at the extremes of our spectrum, compati-
bility properties do little to distinguish between possible objects. Under the
Lewisian view, every possible object x is compatible with every intrinsic
property P: let y be a possible object that instantiates P; then by recombin-
ation, there exists a world containing a duplicate x* of x and a duplicate y* of
y; since P is intrinsic and is instantiated by y, it is also instantiated by y* —so
a duplicate of x exists at a world where P is instantiated.36 And for the hyper-
necessitarian, for any property P, any two possible (that is, actual) objects are
either both compatible with P or both incompatible with P—depending
only on whether or not P is actually instantiated.
But things are quite different under necessitarianism about laws. If
quantum mechanics were to have the consequence that no horses can
exist unless a pig exists but to have no other surprising consequences of this
35
Note under this scenario, equinity would not be intrinsic2 (because it would imply
accompaniment). But it would presumably be intrinsic1 (since whether the atoms of an object
are arranged horse-wise is, presumably, something that cannot be changed by adding things to
an object’s world).
36
Of course, this does not work for extrinsic properties: Loki is not compatible with the
property of living in a godless world.
sort, then the property of compatibility with the absence of porcinity

would distinguish between Bucephalus (who lacks this property) and his
opposite number among dogs, Grisbi (who possesses this property). Fur-
ther, under weak necessitarianism, one expects that there should be alien
forms of matter governed by alien laws incompatible with the pig-wise
arrangements of ordinary atoms. Creatures formed of such matter would
differ from Bucephalus and Grisbi in being incompatible with porcinity.
By placing hefty constraints on recombination, nomic necessitarianism
renders compatibility properties non-trivial: when two objects differ in
their compatibility properties, this tells us something about the laws of
nature.
4 Necessitarianism about Geometry

Relationalists in search of a grounded, ambitious account of geometric
possibility can exploit the nexus between restrictions on recombination
and the encoding of modal information in non-trivial compatibility prop-
erties.
Let us, as usual, take a relationalist world to consist of material points
instantiating certain properties and relations. A mereological sum of such
points is a region. We say that a relationalist world w is full if for every region
R of w, all of the geometric properties compatible with R are instantiated
at w. Two objects (worlds or regions) are geometric duplicates if they share all
of their intrinsic geometric properties (including, of course, their compati-
bility properties).37 We lay down the following postulates.38
Postulate I. The geometric facts about a relationalist world w are
specified by specifying the distance relations between material points
at w and the compatibility properties of the regions of w. (So, in
particular, the complete material configuration at a full world deter-
mines a metric geometry.)
Postulate II. For every metric space (X, d) representing a possible
geometry of space, there is a full relationalist world w whose material
points instantiate the metric geometry of (X, d ).
37
So O1 and O2 are duplicates if and only if there is a correspondence between their parts
that relates parts instantiating the same intrinsic properties and which preserves all intrinsic
geometric relations of fusions of parts.
38
Recall that we continue to restrict attention to static worlds.
Postulate III. Within the set of full relationalist worlds, all geometric
facts supervene on metric facts. That is: regions R1 and R2 of full
worlds w1 and w2 are geometric duplicates if and only if there is an
isometry f : w1 ! w2 that maps R1 to R2 .
Postulate IV. For every region R in a full relationalist world, there is
a relationalist world whose material configuration is a geometric
duplicate of R.
Postulate V. Every region of a relationalist world is a duplicate of a
region of a full relationalist world.
Postulate I tells us that once we know the facts about distances and
compatibility at a relationalist world, we know everything about the
geometric properties and relations instantiated there.39 Postulate II is a
sort of principle of plenitude, telling us that for every way space could be
we have a full relationalist world where space has that structure.40 Postu-
late III tells us that regions in full worlds are duplicates of one another if
and only if: (i) their worlds instantiate the same metric geometry; and (ii)
the regions determine the same superposability class in this shared metric
geometry. Postulates IV and V govern relations of duplication between
regions at full and non-full worlds. Of course, geometric duplicates share
all of their compatibility properties as well as instantiating the same pattern
of distance relations.
Geometric duplication is an equivalence relation. So if the material
configuration at a world w is a duplicate of regions R1 and R2 in full
worlds w1 and w2 , then R1 and R2 are duplicates of each other. So w1 and
w2 instantiate the same metric geometry and R1 and R2 correspond to the
same superposability type in this geometry. So it makes sense to speak of
the metric geometry and superposability type of w.
It is natural to say that a given pattern of distance relations (or distance
relations plus compatibility properties) is geometrically possible relative to
w if it is instantiated in the metric geometry of a full world that contains a
duplicate of the material configuration of w. Similarly, it is natural to say
that a given pattern of distance relations (or distance relations plus com-
39
Note that Postulate I does not say anything about which properties or relations are
basic—in particular, it does not say that distance relations are somehow more basic than, say,
non-quantitative relations such as betweenness and congruence.
40
So what we have here is more closely analogous to weak nomic necessitarianism than to
strong nomic necessitarianism.
patibility properties) represents a geometrically possible extension of the

material configuration C of w if one can find a world featuring that pattern
of relations which contains a duplicate of C.
The resulting account of geometric possibility is of course grounded: if
two relationalist worlds are geometric duplicates, then they agree about
geometric possibility. The account is also ambitious. For let w be a sub-
stantivalist world with material configuration C. Then there is a full
relationalist world w * whose material points instantiate the same metric
geometry as do the points of space at w. And there is of course a region R
in w * that corresponds to the same superposability type in w * as C does in
w. So there exists a relationalist world w ** whose total material configur-
ation is a duplicate of R. The worlds w, w * , and w ** agree about what
configurations are geometrically possible. And w and w ** agree about
which (qualitatively specified) material configurations are geometrically
possible extensions of their respective material configurations.41
Examples
The crucial role that compatibility properties play in encoding the geometry
of space at a world can be illustrated by considering one-particle worlds.
Example 4.1 (One-Particle Worlds). Let w1 , w2 , and w3 be one-particle
worlds, with w1 having the spatial geometry of the Euclidean plane, w2
having the spatial geometry of Euclidean three-space, and w3 having
spherical spatial geometry. The particles at our three worlds are not
geometric duplicates of one another: each is a duplicate of a particle in a
full world sharing the spatial geometry of its own world (and particles in
full worlds with distinct spatial geometry can never be duplicates). So the
particles in our three worlds must differ in their intrinsic properties. These
properties are exhausted by specifying the distance relations and the
compatibility properties—and since the former are trivial at one-particle
worlds, we know that our particles differ in their compatibility properties.
Indeed, the particle of w1 has the following properties: (i) compatibility
with a material configuration with the geometry of the Euclidean plane;
(ii) incompatibility with any material configuration that includes the
41
It is also possible to capture the sort of non-qualitative but relationalistically acceptable
information discussed in fn. 14 of Ch. IV above. This would require keeping track not just of
which regions are duplicates of one another, but also of duplications (the trans-world relations
that match up corresponding parts of duplicate regions).
Euclidean plane as a proper part. The particle at w2 has (i) but not (ii) while
the particle at w3 has (ii) but not (i).42 &
The compatibility properties of the material configuration at a world play a

crucial role in telling us what the geometry of space is at that world (they tell
us whether the world is full—and if not, what sort of full worlds share its
geometry). But the compatibility properties also encode information about
the “location” properties of matter. For consider a full relationalist world w
containing regions R1 and R2 that are congruent but not superposable.
There are non-full relationalist worlds w1 and w2 whose material configur-
ations are duplicates of R1 and R2 , respectively. The compatibility proper-
ties instantiated at these worlds differ. They tell us, e.g., that if the space at w1
were filled out with matter, the material configuration of w1 would play the
geometric role played by R1 rather than that played by the congruent region
R2 . It is important to note that in this role compatibility properties may
encode geometric information of a very familiar sort.
Example 4.2 (Cylinder). If we consider two points separated by a small
distance in a world with the spatial geometry of a cylinder, then their
compatibility properties encode: (i) the geometry of space; and (ii) the
inclination between the axis of the cylinder and a straight line joining the
points, by telling us what sort of geometric role would be played in a full
cylindrical world by the material line joining duplicates of our two
points. &
Example 4.3 (Punctured Plane). Let w be a full relationalist world with
the geometry of the punctured Euclidean plane. Let x1 be a point in w
close to the missing point and let x2 be a point distant from the missing
point. Let w1 and w2 be one-particle relationalist worlds, with the particle
u1 in w1 being a geometric duplicate of x1 and the particle u2 in w2 being
a geometric duplicate of x2 . The material configurations of w1 and w2 are
duplicates of non-superposable regions of the same world. They of
course agree in their metric properties. Their compatibility properties
determine the same spatial geometry but correspond to different “loca-
tions”: each of u1 and u2 is compatible with the material geometry being
42
What about empty worlds? By Postulate IV above the material configuration at an
empty world would have to be a duplicate of a material configuration at a full world. This
would appear to be possible only if we count the null individual as a region (see § 6 of Ch. IV
above for discussion of this sort of ploy).
given by the punctured plane—but u1 is compatible only with the missing

point being nearby, u2 compatible only with the missing point being far
away. &
5 Super-Grounding and Necessitarianism

Recall that in the terminology of Chapter II above, an account of a species
of modality is grounded if it has modal facts supervening on intrinsic facts
about worlds (i.e., according to it duplicate worlds never differ in modal
facts). A grounded account is super-grounded if, further, the properties and
relations constituting the supervenience base are themselves free of the
taint of modality.
Sales pitches for best-system-style treatments of laws of nature often
place great stress on the fact that the item offered is super-grounded,
reducing nomic possibility to non-nomic facts.43 And one might make
the same sort of claim for the best-system treatment of geometry, under
which facts about geometric possibility supervene on facts about the
pattern of distance relations instantiated.
The varieties of necessitarianism about laws and geometry considered
above are certainly grounded, since they have facts about modality super-
vening on the fundamental properties instantiated (in the nomic case) or the
intrinsic properties instantiated (in the geometric case). But it is natural to
doubt whether they are also super-grounded. Nomic necessitarians in effect
encode modal information in fundamental physical properties (mass, charge,
and the like) while the form of geometric necessitarianism sketched above
features compatibility properties. Indeed, both views involve robust con-
straints on recombination—so knowing that an object at a world instantiates
certain properties tells you a lot about what else can be going on at that world.
One reason to be interested in the question of super-grounding for
necessitarianisms is as follows. Once one sees how these views work, it
opens up the possibility that one can show that any newfangled modal
notion one cares to introduce can be grounded if one is willing to accept
the requisite restrictions on recombination. A natural place to make a stand
against this strategy is on the question of super-grounding: one could grant
that our necessitarian approaches are grounded, but maintain that super-
43
See e.g. Schaffer, “Causation and Laws of Nature.”
grounding is the desirable feature in the neighbourhood—and insist that

while the nomic and geometric best-system approaches are super-grounded,
the corresponding varieties of necessitarianism are not.44
In this section I aim to do what I can to undermine the appearance that
best-system approaches enjoy a clear and impressive advantage of this kind
over necessitarian approaches.
Ground Rules
The question whether an account is super-grounded is substantive and
sharp only to the extent that there is a substantive and sharp distinction
between modally tainted and modally untainted properties and relations.
On some philosophical accounts of properties the distinction between
modal and non-modal properties evaporates—and with it, the notion that
super-grounding is a stronger requirement than grounding.45 For present
purposes we can set such approaches aside and take it for granted that there
is a non-trivial modal/non-modal distinction for properties. We still face
the problem of attempting to make this distinction sharp enough to
underwrite a usable notion of super-grounding. This turns out to be a
slippery matter.46
In practice, in most discussions of this terrain we are simply offered a
few examples that are supposed to fix the modal/non-modal distinction.
We are typically told something like: (i) dispositional properties are clear
cases of modal properties; and (ii) properties like mass, charge, and distance
are clear examples of non-modal properties. Just about everyone who
recognizes a modal/non-modal distinction is willing to grant (i). But the
status of (ii) is sensitive to the account of laws in play. Indeed, advocates of
the best-system approach to laws tend to claim that while mass and the like
are non-modal under approaches like the sensible Lewisian account, these
44
Perhaps geometric necessitarianism is worse off in this respect than is nomic necessitarian-
ism. In the nomic case, restrictions on recombination have independent motivation in certain
accounts of properties. But in the geometric case the restrictions on recombination are more
or less posited for the sake of grounding geometric possibility in facts about intrinsic properties.
45
According to some views, all properties have both modal and categorical aspects;
according to others, all properties are dispositional and hence modal. For representative
examples of these two strands of thought, see Heil, “Dispositions” and Bird, Nature’s Metaphys-
ics. For another view in the same neighbourhood, see Cartwright, The Dappled World.
46
To get a feeling for the difficulty of delineating nomically-involved facts from others,
see: Carroll, Laws of Nature, §1.1; Earman and Roberts, “Contact with the Nomic. Part I,”
§3.4; Skow, “Earman and Roberts on Empiricism about Laws;” and Roberts, “Reply to
Skow.”
properties develop a distinctly modal taint if called upon to do the sort of

work that nomic necessitarians expect of them.47 So the typical approach
to the modal/non-modal distinction via ostensive definition is unsuited to
our present purposes.
In the absence of a full analysis of the notion of a modal property, let us
try to make do with the following rough heuristic principle:
A property or notion is modally untainted if and only if it would be

acceptable to (some) anti-realists about modality.48
Admittedly, acceptance of this principle doesn’t exactly lead to immediate
progress.
For one thing, the notion of anti-realism about modality is itself unclear.
We can agree that anti-realism is the doctrine that there are no necessities
in nature, or the doctrine that the only necessities are verbal necessities, or
something of the like.49 But one way to aim to qualify as an anti-realist
about modality is to argue that the true generalizations that we label
necessary are distinguished only in possessing some special but unobjec-
tionable property. And in order for the resulting view to count as anti-
realism about modality, the property latched on to had better be . . . non-
modal in nature.
For another thing, the edicts of paradigm anti-realists provide little
explicit guidance for our present project. Unsurprisingly, among anti-realists
one finds a number of distinct strands of opinion concerning which prop-
erties are acceptable. One strand is baldly opportunistic—surely it can’t hurt
to take on board this or that apparently modal notion, so long as one does
not make a habit of it.50 According to what is probably the dominant strand
of thought, it is harmless to accept those apparently modal terms, such as
dispositional terms, that can be paraphrased away via conditions stated in
ordinary, unproblematically non-modal, scientific vocabulary.51 But this
reasonable-sounding requirement is of course of no help in the present
47
See e.g. Earman and Roberts, “Contact with the Nomic. Part I,” pp. 11 f. and Loewer,
“Humean Supervenience,” §vi.
48
For this sort of principle, see e.g. McDaniel, “Modal Realism with Overlap,” §ii.
49
For views of this kind, see e.g. Quine, “Necessary Truth;” and van Fraassen, The
Scientific Image, §6.5, and Laws and Symmetry, p. 354 n. 1.
50
See the epigraph to this chapter. This passage is quoted approvingly by van Fraassen,
Introduction to the Philosophy of Time and Space, §vi.6.c.
51
See e.g. Monton and van Fraassen, “Constructive Empiricism and Modal Nominalism,”
§4; and Quine, “Natural Kinds,” “Necessary Truth,” and Word and Object, §46.
context, where what is at issue is whether any of the properties trafficked in

by science are non-modal when understood along necessitarian lines.52
Nonetheless, our heuristic will turn out to be of some help.
Super-Grounding: the Nomic Case
On what seems to be the standard view concerning such things, there are
three sorts of regularities concerning mass at the actual world: (i) mere
accidents (e.g., concerning the sizes of gold spheres); (ii) physical necessities
(e.g., that all electrons have the same rest mass); and (iii) metaphysical
necessities (e.g., that masses of objects can be represented by real numbers
and that no object-stage has more than one mass).
Roughly speaking, best-system analysts buy a version of the standard
view in which the distinction between the first and second of these
categories is somewhat deflated (theirs is a reductive account of nomic
necessity, on which laws are regularities that exhibit certain theoretical
virtues) while nomic necessitarians buy a version in which the distinction
between the second and the third categories collapses completely (on their
account it is no more possible for there to be a world without gravitational
attraction whose fundamental properties match those of the actual world,
than it is for there to be a world at which mass is representable by zillion-
tuples of quaternions but not by real numbers). Both parties see there as
being a distinction between brute regularities and regularities that hold of
metaphysical necessity—and for this reason it is in fact far from clear that
the best-system approach, as it is normally conceived, is better off than
nomic necessitarianism when it comes to the question of super-grounding.
Or so I shall argue.
Let us begin with the best-system approach. Clearly Lewis’s own
account of properties and laws is not super-grounded. In explaining his
principle of recombination, which implicitly characterizes the principle of
individuation of properties, Lewis relies on the idea that each property
belongs to a “category.”53 For instance, he takes the property of having
positive charge to be a monadic property—at every possible world. This is
a notion that anti-realists about modality will consider unintelligible—so
52
A third neo-Kantian strand of thought can be found in van Fraassen. It is sketched
in broad outline in §§iii.4 and vi.6.c of Introduction and in §6.5 of Scientific Image. But, as
becomes clear in §2 of the Postscript to Introduction, the programme is radically incomplete—
one important stumbling block being the treatment of geometric relations.
53
See fn. 7 above.
Lewis’s account fails to be super-grounded because it involves an account

of properties on which fundamental physical properties are modal.
This observation won’t much worry advocates of the best-system
account—it is not exactly news that if you combine Lewis’s account of
laws with other elements of his system, the resulting picture is unaccept-
able to anti-realists about modality. The question is whether there are
versions of the best-system account under which fundamental physical
properties are non-modal.
It is tolerably clear that there are such accounts. Anti-realists about modal-
ity are perfectly happy talking about physical properties and about the pattern
of instantiation of these properties at our world. Of course, more is required
in order to state the best-system view: that view singles out some of the true
regularities at our world as especially interesting, in virtue of figuring in
certain theories stated in a certain sort of language. Making sense of this
requires being an anti-nominalist in two senses: being willing to countenance
abstract objects, such as the theories in question (the ideal theory need never
actually be written down); and accepting that some, but only some, systems
of predicates correspond to joint-carving sets of properties (unless one
restricts the languages in which the theories of interest are formulated, the
account cannot get off the ground). Actual anti-realists about modality tend
to be nominalists—but presumably it is possible to combine anti-realism
about modality with anti-nominalism in a coherent way. The result would
be a version of the best-system account that drew a distinction between
mere regularities and special regularities—with the latter distinguished,
presumably, by their capacity to play certain explanatory roles, and so on.54
But if that is all that the best-system account is taken to be—a machine
that anti-nominalists can use to sort regularities at our world into two
piles—then it will collapse the ordinary distinction between metaphysical
54
Quine (“Reply to Parsons,” pp. 397 f.) and van Fraassen (Laws and Symmetry, pts. i and
ii) are enemies of laws of nature. But for Quine our attributions of nomic necessity are
traceable to “what passes for an explanatory trait or the promise of it” (“Necessary Truth,”
p. 76). On van Fraassen’s approach, explanations are answers to why-questions, in which
answers those propositions that hold in every model of our current theory play a special role:
“The only genuine empiricist course . . . is to deny that explaining something consists in
showing why it had to be the way it is—tout court. We must say instead that . . . the criteria
for what is a good answer are context-dependent. To be specific, I think that the important
contextual factor here is the background of accepted scientific theory. This must play the
role . . . played in Aristotle’s account by objective necessities in nature” (“A Re-Examination
of Aristotle’s Philosophy of Science,” p. 43; for elements of van Fraassen’s implementation of
this program, see ch. 5 of Scientific Image and pts. ii and iii of Laws and Symmetry).
and physical necessities. That mass is representable by real numbers is a

regularity at our world that will be sorted into the same pile as claims
traditionally held to be merely physically necessary. This sort of collapse is
of course familiar as a feature of nomic necessitarianism—a feature that
many advocates of the best-system approach find objectionable.55
Can one construct a more robust version of the best-system account
that is super-grounded but which preserves the standard tripartite distinc-
tion between regularities at our world?
I believe that one can. Here is one way to proceed. Consider the notion
of an essential property. Is the thesis that I have some of my properties
essentially and that some truths concerning me are essential truths (i.e.,
follow from my possession of my essential properties) compatible with
anti-realism about modality?
Certainly not, if, as is perhaps most usual, we understand my essential
properties to be those properties that I possess in every possible world
in which I exist and understand essential truths about me to be of
the form: necessarily, if I exist then such and such obtains. For it is hard
to see how an anti-realist about modality could make sense of these
notions.
But there is another approach to understanding essential properties and
essential truths: one can take the essence of a thing to be something like its
real definition—roughly, the conjunction of those true propositions about
it that tell us about its nature.56 And one can likewise take the essential
properties of a thing to be those that follow from its essence and take the
essential truths concerning a thing to be those that hold in virtue of its
essence.
A realist about modality who adopts this second approach to under-
standing essential properties and truths will of course think that essential
truths generate necessary truths—any truth that holds of me in virtue of
my essence will hold of me at any world at which I exist. But, as Kit Fine
has forcefully argued, it by no means follows that essential truths should be
identified with modal truths: at every world at which either exists, Socrates
is a member of the singleton set {Socrates} and this set has Socrates as
55
See e.g. Langton and Lewis, “Defining ‘Intrinsic’,” §v.
56
This venerable idea is developed in a contemporary form in Fine, “Essence and
Modality” and “Senses of Essence.”
a member; but while it is of the nature of the set to have Socrates as a

member, it is no part of his nature to be a member of any set.57
The fact that under this approach essential truths are not just modal
truths even for realists about modality leaves the door open for anti-realists
about modality to accept essences of the sort envisioned.58 Stripped of any
modal consequences, the cash value of positing such essences is to allow
one to distinguish between two kinds of truths—those that can be
explained as holding in virtue of the natures of the objects involved, and
others.59
If this is accepted, then it opens the way for a best-system approach
acceptable to anti-realists about modality but under which one postulates
essential truths about the fundamental physical properties—such as that
mass is representable by real numbers. One would retain the usual tripartite
distinction among regularities at the actual world under this approach,
even if it were combined with anti-realism about modality: first-class
regularities admit one kind of explanation (the identities of the properties
involved); second-class regularities of another (they follow from the prin-
ciples of the ideal theory); third-class regularities are those that must be
taken as surd.
But then there is nothing to stop nomic necessitarians from proceeding
likewise.60 Some varieties of nomic necessitarianism attribute essences to
fundamental properties.61 Under such approaches, the laws at a world are
to be thought of as those constraints on the pattern of instantiation of
fundamental properties that are implied by the conjunction of the prop-
osition that says just which fundamental properties are instantiated with
the essential truths about those properties. If we are allowing best-system
57
For this and other examples, see “Essence and Modality,” pp. 4 ff. Van Fraassen in effect
observes that if one wants to found scientific explanations on essential truths, then the
asymmetries of explanation provide motivation for taking the essential truths to form a proper
subset of the necessary truths; “Re-Examination,” §§1.5 and 2.2.
58
Kment, too, takes Fine-style essential properties to be non-modal; see §1.3 of “Coun-
terfactuals and the Analysis of Necessity.”
59
I assert only that the combination of anti-realism about modality with a certain sort of
essentialism is consistent, not that it would be likely to appeal to actual partisans of either
view.
60
Of course, some forms of nomic necessitarianism are pretty clearly not super-grounded.
According to some necessitarians, fundamental properties are dispositional; see e.g. Bird,
Nature’s Metaphysics. Necessitarians and non-necessitarians alike agree that, if taken to be
dispositional, fundamental properties must be viewed as modally tainted; see e.g. ibid., §4.3.1
and Loewer, “Humean Supervenience,” §vi.
61
See e.g. Bigelow et al., “The World” and Ellis, “Causal Powers.”
analysts to attribute essences to fundamental properties in a way consistent

with anti-realism about modality, then we must do likewise for essentialist
necessitarians. Again, the cash value of the distinction between an essential
truth about a fundamental property and a mere truth about that same
property will be the distinction between a regularity that can be explained
as holding in virtue of the nature of the properties involved and a
regularity which cannot be so explained.
What is the upshot? There are best-system approaches that are clearly
super-grounded. But these involve collapsing the usual tripartite distinc-
tion between regularities at the actual world, in a way that few advocates
of the best-system approach are liable to relish. There are also best-system
approaches that maintain this tripartite distinction via appeal to essential
properties but which, arguably, are super-grounded. But if such
approaches are deemed super-grounded, then it must also be admitted
that there exist super-grounded forms of nomic necessitarianism—which
admission goes a long way towards undermining the interest of the notion
of super-grounding. The advantage that best-system analysts are supposed
to enjoy over necessitarians in regard to super-grounding proves to be
elusive.
Might there be other ploys by which best-system analysts could achieve
super-grounding while maintaining the tripartite distinction between
types of regularities without opening the door to nomic necessitarians?
This seems unlikely to me. Under any such ploy, the regularities at our
world that are traditionally taken to be metaphysically necessary will have
to be assigned some feature that the other regularities lack. This feature
must be acceptable to anti-realists about modality, but must imply meta-
physical necessity in the setting of realism about modality. There may well
be notions other than that of an essential truth that will do the job. But it is
hard to see what would prevent nomic necessitarians from co-opting any
such notion for their own purposes. For such prevention would seem to
require that it be incoherent for someone to maintain that the regularities
ordinarily deemed physically necessary at our world fall under the notion
in question. That is a pretty high standard to meet—so long as one restricts
attention to notions that have independent motivation.
Super-Grounding: The Geometric Case
Under the necessitarian account, the intrinsic geometric properties of the
material configuration of a relationalist world include (and are fixed by
specifying) the compatibility properties of the points making up that

configuration and the distance relations between those points. Together,
facts about distances and facts about compatibility suffice to determine
both the structure of space (the metric geometry that we use to settle
questions about which material configurations are geometrically possible)
and the location of matter (the superposability type within the metric
geometry encoding the structure of space—this determines which
material configurations are geometrically possible extensions of the
given one). It is natural to split our investigation into two corresponding
parts.
The Structure of Space. Let us begin with the question whether the
necessitarians’ distance relations and compatibility properties must be
viewed as modally-tainted in virtue of encoding information about the
structure of space.
It will be helpful to consider the corresponding question regarding the
sort of best-system approach to geometric modality considered in Chapter
III above. The best-system approach is naturally regarded as being the
approach with the strongest claim to being super-grounded. Let us ask
exactly what commitments advocates of the best-system approach would
have to take on board in order to secure this claim.
Intuitively speaking, within the best-system approach, at any world
some claims about the pattern of distance relations instantiated hold of
metaphysical necessity, some of (mere) geometric necessity, while some
are contingent. Consider, for instance, a world w with the spatial structure
of the Euclidean plane that happens to contain four points corresponding
to the vertices of a square. That the distance relations at w satisfy the metric
space axioms is a matter of metaphysical necessity. That the distance
relations between any four points of w have a vanishing Cayley–Menger
determinant (see p. 12 above) is a matter of (mere) geometric necessity.
That w includes three points corresponding to the vertices of a right-
angled triangle is a contingent matter.
Under the standard best-system approach, the metric space axioms have
a very different status from other regularities. As in the nomic case, if we
want to construct a super-grounded best-system account, we must either
drop this feature or find some way of marking the special status of the
metric space axioms that can be safely combined with anti-realism about
modality. A natural way to implement the latter strategy is to take the
satisfaction of the metric space axioms by distance relations to be an

essential truth rather than a modal one. Then the distinction that we are
after can be understood as follows:
The metric space axioms hold at w in virtue of the nature of distance,

whereas many other regularities in the pattern of distance relations at
w admit of no such explanation.
If this is deemed to be consistent with anti-realism about modality then
the distance relations of this sort of best-system approach should be
counted as non-modal, and the approach as a whole can be seen to
be super-grounded.
Now let us return to the problem we began with: Must the necessitar-
ians’ distance relations and compatibility properties be viewed as modally-
tainted in virtue of the way in which they encode information about the
structure of space?
Here is one (ugly) way to see that that the answer is: No. Up until now,
we have proceeded as if all parties were committed to working with generic
distance relations: that is, we took it that if the numerical values of (the
ratios of ) the distances between the set of points at world w1 were the same
as for the set of points at world w2 , then the exact same relations were
instantiated at both worlds. But there is an alternative approach under
which distance relations are taken to be specific: the exact same distance
relations can be instantiated at two worlds only if the worlds share their
geometry—otherwise the relations are merely analogous (in the sense that
they can be given the same numerical representations).62 Geometric
necessitarians who resort to specific distance relations can view each such
relation as possessing a rich essence that encodes the geometry of space at
the worlds at which it is instantiated. The facts about geometric possibility
that encode the geometry of space at a given world will then supervene on
a non-modal base at that world (namely, the fact that a certain specific
distance relation is instantiated there).
For example, we could take there to be a specific notion of distance
instantiated at just those worlds at which space is two-dimensional and
62
By way of motivation, consider the way that (weak) nomic necessitarians insist that the
exact same fundamental property cannot figure at worlds with different laws, even it would
be natural to give the same mathematical representation of the pattern of instantiation of
fundamental properties at the two worlds.
Euclidean. It would then be an essential truth about this specific distance

relation that at any world at which it is instantiated, the metric space
determined by the material configuration is embeddable into the Euclid-
ean plane.63 Of course at any such world the metric space determined by
the material configuration will be embeddable into metric spaces other
than the Euclidean plane—but under the current proposal, this fact would
be a brute one, not explicable in terms of the nature of the distance relation
at that world.
So far, the situation is much the same as in the nomic case: it appears that
only those best-system analysts who are willing to deny the standard
distinctions among regularities at the actual world enjoy an advantage
over necessitarians on the question of super-grounding.
The Location of Matter. The question remains whether the geo-

metric necessitarians’ compatibility properties must be counted as modally
tainted in virtue of fixing the location properties of the material configur-
ations at relationalist worlds.
Now, it is clear that at least some of the location information encoded in
compatibility properties of material configurations can be thought of as
inherent in geometric relations with just as much claim to be non-modal as
distance relations. For consider our favourite example of a situation in
which specifying the distance relations between the points of a material
configuration fails to specify all location properties: specifying that three
points in the (unit) elliptic plane have pairwise separation of 3 units fails to
specify their location since it leaves open whether the three points are or
are not collinear. So in this sort of example specifying the compatibility
properties of material points plays a substantive role in addition to that
played by specifying distance relations—but this role could equally well be
fulfilled by specifying facts about collinearity. And there seems to be no
special reason why facts about collinearity at an elliptic world should
automatically be judged to be modal in nature.64
63
A metric space is isometrically embeddable in the Euclidean plane if and only if the
distance relations instantiated by every five-point set of the given metric spaces satisfy a
certain algebraic condition. See Blumenthal, Theory and Applications of Distance Geometry,
§§38–43.
64
Indeed, as has been mentioned before, some authors take it for granted that relationalists
should take non-quantitative relations such as collinearity as the basis for a modal construction
of quantitative relations; see Field, “Can We Dispense with Space-Time?”
The question, of course, is whether in every case necessitarians about

geometric possibility can find non-modal stand-ins for all of the roles of
compatibility properties in fixing locations. We can think of compatibility
properties as being useful placeholders for more interesting (and less
problematic) geometric properties: the utility of compatibility properties
in developing the necessitarian approach above lay in the fact that they
allowed us to develop a general scheme that treated every world on the
same footing. But in addressing the question of super-grounding we will
want to look more closely at the particular fundamental properties at each
world, and ask whether they can be thought of as non-modal.
Some cases will certainly be less clear than our example in the elliptic
plane. Consider a world with the geometry of the punctured plane,
occupied by a single material point. Fixing the location of the material
point means specifying its distance from the “missing point of space.” Are
we entitled to simply posit a geometric fact that does this work, and then
to insist that it is non-modal?
I am not sure whether location properties pose an insuperable obstacle
to considering geometric necessitarianism to be super-grounded. In order
to adjudicate this question one would have to have a much better sense
than I do for what sorts of properties one needs in general to fix the
superposability type of a region in a metric space.65
6 Summation
Under any ambitious approach to geometric possibility, the facts about
geometric possibility at a world w with material configuration C corres-
pond to the facts about geometric possibility at a full world w * whose
material configuration includes a region C * just like C.
The precise relation between C and C * depends on the details of the
account. Under the primitivist approach developed in Section 4 of Chapter
IV, C and C * were composed of the same material points, standing in the
same distance relations. Under the variant primitivist approach of Remark
5.2 of Chapter IV, C stood in a primitive trans-world relation of super-
65
A good place to begin thinking about this topic would be with the spaces and properties
considered in Brehm, “The Shape Invariant of Triangles and Trigonometry in Two-Point
Homogeneous Spaces” and in Brehm and Et-Taoui, “Congruence Criteria for Finite Subsets
of Complex Projective and Complex Hyperbolic Spaces” and “Congruence Criteria for
Finite Subsets of Quaternionic Elliptic and Quaternionic Hyperbolic Spaces.”
congruence to C * (and to every region at w * superposable with C * ). Under

either of these approaches there are worlds whose material configurations
are geometric duplicates of C—in virtue of instantiating the same pattern of
distance relations or other basic relations—but at which the facts about
geometric possibility are different. So these are ungrounded approaches.
Under the approach developed in the present chapter, C and C * are
held to be geometric duplicates of one another and there is no world
whose material configuration is a duplicate of C but at which the facts
about geometric possibility differ from those at w. This gives us a grounded
account—but only at the cost of requiring the existence of a family of
intrinsic geometric properties much richer than those required by primi-
tivists. Along with the closely related question of super-grounding, the
question whether the resulting picture is an attractive one will have to
wait, one would suppose, until we have a more detailed picture of just
what these properties are like at generic relationalist worlds.
Conclusion
Substantivalists take space to be composed of parts standing in geometric

relations to one another and take the geometric relations between material
bodies to be derivative on the geometric relations between the parts of
space that they occupy. Relationalists deny that space consists of parts and
take the geometric relations between bodies to be direct rather than
mediated by relations between the parts of space occupied.
Both parties take space at a given world to have a determinate structure:
to be finite or infinite, to be three-dimensional or otherwise, etc. Sub-
stantivalists have a straightforward account of the content of claims about
the structure of space—they are claims about the pattern of relations
instantiated by points of space.
Relationalists can adapt this strategy to their needs in the case of worlds
in which matter forms a plenum (a case of central interest to traditional
relationalists). But modern relationalists will also want to countenance
worlds in which matter is sparse. What does it mean, for instance, to say
that space is infinite in extent at a world in which there is an upper bound
on the distance between material points? Here it seems that relationalists
are driven to appeal to a notion of geometric possibility—space is un-
bounded in extent if for every real number, it is geometrically possible for
there to be a pair of points whose distance exceeds that number.
It is natural for modal relationalists to seek an account under which
geometric possibility is grounded in ordinary geometric facts, in the sense
that it requires that if the material configurations at two worlds are
duplicates of one another, then the facts about geometric possibility are
the same at those two worlds. It is also natural for relationalists to seek an
account that is ambitious, in the sense that according to it, corresponding to
every substantivalist world with a given material configuration, there is a
relationalist world whose material configuration has the same geometry as
conclusion 135
that of the substantivalist world, and at which the (relationalistically

respectable) facts about geometric possibility match those of the substan-
tivalist world. And it is also natural for relationalists seek a metric account,
under which material configurations corresponding to the same pattern of
distance relations are geometric duplicates of one another.
But these natural desiderata are of course mutually incompatible so long
as substantivalists recognize the existence of one-particle worlds that differ
as to their spatial geometry. My strategy has been to consider accounts of
geometric possibility that give up on one of the desiderata while holding
on to the other two.
In Chapter III, I considered a grounded, metric account of modal
relationalism modelled on the best-system account of nomic possibility.
Any grounded, metric account will be be built upon a rule that allows one
to pass from the list of the instantaneous metric geometries instantiated by
matter at a world to the metric geometry that encodes the structure of
space at that world. In general, there is no privileged solution to this sort of
selection problem. Advocates of the best-system approach advance the
bold empirical conjecture that there are intersubjective standards of sim-
plicity and balance-of-simplicity-and-strength that allow us to solve the
selection problem by assigning to each world the strongest-simplest metric
geometry consistent with the geometries of the instantaneous material
configurations. I regard this approach as unpromising because its central
empirical conjecture is largely unsupported by evidence and is in fact itself
quite implausible. It is also marred by several unattractive features—first
among these being the attribution to geometric notions of a startling
species-relativity. Further, it is hard to see how any reasonably well-
motivated, grounded, metric account of geometric possibility could
avoid incorporating flaws of this general sort.
Chapter IV was primarily concerned with ambitious, metric approaches
under which the intrinsic geometric facts at a relationalist world are taken
to be exhausted by the facts about distances between its points, but in
which one takes as primitive a notion of geometric possibility that allows
relationalists to match substantivalists stride for stride. These approaches
have the advantage of being straightforward—and the disadvantage of
looking rather more like theft than like honest toil. The necessitarian
approach developed in Chapter V is a close relative of the primitivist
approaches—but manages to be both grounded and ambitious by taking
136 conclusion
the intrinsic geometric facts at any world to be rich enough to encode full
information about geometric possibility at that world. I suspect however,
that some readers will feel that the groundedness of the necessitarian
approach is not so much a positive credential as a sleight of hand.
As will be clear enough from the foregoing, I think that unambitious
approaches are the least compelling of the options discussed. And while I
do not think that there is at present that much to choose between the
primitivist and necessitarian approaches, I also find the latter more intri-
guing—in part because it suggests interesting questions concerning what
sort of fundamental properties need to be invoked at each world in order
to determine the superposability type of each region (these questions are
closely related to the question whether the base of properties upon which
facts about geometric possibility supervene under the necessitarian ap-
proach is itself non-modal).
Much of the traditional interest of the substantival-relational debate
derives from its close connection with questions about the relativity of
motion and the proper form for dynamical theories in classical physics. In
the classical setting, it is natural for relationalists to take claims concerning
the state of motion of a body to be (explicitly or implicitly) claims about
the way it changes its position relative to some reference bodies. But for a
substantivalist another option is available: a body is at absolute rest if the
part of space that it occupies does not change; it is in uniform motion if
the amount of space it crosses varies linearly with time; etc. For Newton,
the availability of a notion of absolute acceleration was an invaluable
advantage—it provided the foundations for his laws of motion. At the
same time, the availability of an absolute notion of velocity was something
of an embarrassment: within Newtonian physics the absolute velocity of a
system is a real but empirically inaccessible quantity.
How much of Newton’s physics can be reformulated in terms accept-
able to relationalists? The natural starting point is to see what can be done if
one takes the relative distances and relative velocities between material
particles as one’s dynamical variables. The well-known answer is that in
terms of these variables one can write down a theory that permits exactly
the same relative motions of bodies as does Newton’s theory under the
assumption that the total angular momentum of the system vanishes.
There are various strategies that relationalists can adopt if they are
interested in capturing all of the relationalistically respectable content of
conclusion 137
Newton’s theory.66 Here I would just like to observe that in general

settings specifying the relative distances between material points fail to
determine the superposability type of the configuration formed by those
points (i.e., in generic metric spaces, the distances between an arbitrary set
of points fails to determine their embedding up to a symmetry of the
ambient space). So in general relationalists need to specify some informa-
tion beyond the relative distance between material points in order to
specify an instantaneous configuration of those points. This suggests that
relationalists interested in formulating theories of motion for worlds with
spatial structures for which the congruence of configurations does not
guarantee their superposability may face some interesting decisions in
selecting their basic dynamical variables.
Let me end by addressing a question the reader may well feel is long
overdue: How does all of this look if we are interested in spatiotemporal
rather than spatial geometry?
In principle, the situation is much the same. Substantivalists about
spacetime take spacetime to be an entity consisting of parts structured by
spatiotemporal relations, and take the spatiotemporal relations between
material events to be derivative upon the relations between the spacetime
points that they occupy. Relationalists about spacetime deny the existence
of points of spacetime, and take material events to stand in direct spatio-
temporal relations with respect to one another.
Relationalists about spacetime find themselves in a predicament that
directly parallels the familiar predicament of relationalists about space.
Substantivalists need only employ a very small number of spatiotemporal
relations. In the setting of special relativity, for instance, they can get away
with a single binary relation of causal connectability (if they are happy to
abandon trans-world comparisons of scale). But a relationalist who tries to
make do with such a sparse family of properties runs into difficulties: in
general knowing the pattern of instantiation by material events of just a
few geometric relations will fall well short of determining the spacetime
geometry of a world or of determining the embedding up to a spacetime
isometry of those material events.
66
e.g., they can rely on differential equations involving relative accelerations as well as
relative distances and velocities or they can introduce primitive variables to play the role that
total angular momentum about the centre of mass plays in the standard formulation.
138 conclusion
Something more is required. And here it is again natural for relationalists

to appeal to a notion of geometric possibility—and for ambitious relation-
alists to adopt a primitivist or necessitarian account of this notion.
Why, then, have I focused on the spatial case up until now, rather than
on the more interesting spatiotemporal case? For the simple reason that we
have a much better idea what we are talking about when we speak about
space rather than spacetime. We have been able to proceed at a very
general level because we had available the notion of a metric space and
the idea that at any possible world with spatial structure it must make sense
to speak of (ratios of) distances between points. But in the spatiotemporal
setting, we have no such unifying framework. We have examples of
spacetime structures of course: a handful of interesting pre-relativistic
spacetime structures that have proved helpful in thinking about classical
physics, and the vast family of Lorentzian geometries from which the
spacetimes of general relativity are drawn. But there is no obvious way
to think of these two families as unified: the basic relations appropriate to
pre-relativistic spacetimes are quite different from those appropriate to
Lorentzian spacetimes; and there is no known way to generalize these two
families to yield a family that stands to our intuitive concept of spacetime as
the family of metric spaces stands to our intuitive concept of space.67 So
the spatial setting supported a much more general and precise discussion of
modal relationalism than the spatiotemporal setting would have.
67
For an attempt to isolate a notion that stands to Lorentzian spaces in something like the
way that the concept of a metric space stands to our intuitive concept of space, see Busemann,
“Timelike Spaces.”
Appendix A
Simplicity and Ontology
The desideratum of simplicity . . . does play a considerable role in the

critical appraisal of theories. But its satisfaction clearly has no bearing
on the question of their truth.
Hempel
It is widely felt that relationalism involves a more baroque metaphysics

than does substantivalism. Sadly, nothing in the present work is liable to
dispel that feeling.
This leads to an apparently serious worry. It is generally felt that the
debate between substantivalists about space and relationalists about space
has morals applicable to the investigation of the geometric ontology of
the actual world. But if one of those morals is that substantivalism about
space enjoys a dominating advantage in simplicity over its rival, and if there
is no special reason to think that this advantage will disappear when other
contexts are considered, then it may seem difficult to see why one should
spend much time worrying about the details. For it is also generally felt
that the point of investigations of ontology is to discover the truth about
our world and that the proper method of ontology is that suggested by
Quine—to accept as true the simplest overall theory of our world. But it
would seem to follow from what has been said that: (i) if we knew that we
lived in a spatially structured world, then we would have overwhelming
reason to believe that our world was a substantivalist one rather than a
relationalist one; and, (ii), to the extent that one can raise a question
analogous to that at issue between substantivalists and relationalists about
space concerning the geometric ontology of the actual world, we have
overwhelming reason to think that our world is relevantly similar to
substantivalist worlds rather than to relationalist worlds. If all of this is
accepted, then any interest that an extended discussion of relationalism
might otherwise have had threatens to evaporate.
140 APPENDIX A
I myself reject the line of thought just sketched because I reject

the orthodox Quinean view of the force of simplicity considerations in
ontological investigations. A thorough discussion would be very involved.
Here I limit myself to a brief discussion of the roots of the orthodox view
and a quick account of some considerations that I take to undermine it.
Before beginning, I should emphasize that what follows is not intended
as a denunciation of ontological inquiry as it is conducted by metaphys-
icians and by philosophers of physics. For, on the one hand, only some
discussions falling under this heading turn crucially on the aspects of the
Quinean method that I am complaining about.1 And, on the other hand, at
most my discussion gives reason to doubt that this method is especially
likely to lead to the truth. But locating the truth is not the only way
for ontological investigations to contribute to understanding.2 An irenic
formulation of my point: for certain central types of ontological investiga-
tions, the most perspicuous terms of praise for favoured theses are ‘appeal-
ing,’ ‘interesting,’ ‘well-motivated,’ and the like, rather than the traditional
and more impressive-sounding ‘likely,’ ‘plausible’ ‘well-supported,’ etc.
Orthodoxy
Let us begin with an example of the sort that provides motivation for the
orthodox approach. Consider a world w whose inhabitants set out to
determine the spatial structure of their world. Their inquiry splits naturally
into two stages.
In the first stage, the physicists of w perform a series of measurements of
distances and angles. Each set of measurements cuts down the family of
mathematical geometries (metric spaces) that are candidates to represent the
spatial geometry of w. But at any stage only finitely many measurements
have been made and infinitely many candidates remain. Nonetheless, even-
tually the physicists of w find that among the remaining candidates one
enjoys an enormous advantage in simplicity over its rivals. At this point they
close their investigation, concluding (provisionally) that the spatial structure
1
Some philosophical theses in this neighbourhood are in effect highly speculative empir-
ical conjectures—e.g., various conceptual analyses and some claims concerning the proper
interpretation of physical theories.
2
On this point, see Jauernig, “Must Empiricism be a Stance, and Could it Be One?” and
Dorr, “Review of Every Thing Must Go.”
APPENDIX A 141
of w is represented by this remarkably simple candidate. For definiteness, let

us assume that the structure in question is that of Euclidean space.
The second stage of the investigation is conducted by philosophers (for
convenience, let us assume that they are all committed substantivalists).
They observe that their colleagues in physics investigated only the metric
structure of their world’s spatial geometry, not touching on the question of
how that structure is instantiated. There are many ways to axiomatize
Euclidean geometry: you can take points of space as your individuals, and
take distance relations or non-quantitative relations (such as betweenness
and congruence) as your undefined relations; or you can take spheres as your
individuals and take tangency as your sole undefined relation; etc. Each of
these options appears to correspond to a different way that a world with
Euclidean space could be structured at the fundamental level: a Euclidean
world in which all of the primitive relations are binary would seem to be
different from a Euclidean world in which none of the primitive relations
are binary; a Euclidean world in which there are unextended things would
seem to be different from a Euclidean world in which there are punctual
objects, and so on.3 How are we to choose between these options?
Quine famously suggested that we should regard the two phases of
our imagined project as being methodologically homogeneous with one
another:
Our acceptance of an ontology is, I think, similar in principle to our acceptance of a
scientific theory, say a system of physics: we adopt, at least insofar as we are
reasonable, the simplest conceptual scheme into which the disordered fragments
of raw experience can be fitted and arranged. Our ontology is determined once we
have fixed upon the over-all conceptual scheme which is to accommodate science
in the broadest sense; and the considerations which determine reasonable con-
struction of any part of that conceptual scheme, for example, the biological or the
physical part, are not different in kind from the considerations which determine a
reasonable construction of the whole.4
3
Readers who see only pseudo-questions here should substitute an example more to their
liking.
4
“On What There Is,” p. 190. Of course, there is ample reason to doubt that Quine took
ontological theses as seriously as do many contemporary ontologists: “Our scientific theory
can indeed go wrong, and precisely in the familiar way: through failure of predicted
observation. But what if, happily and unbeknownst, we have achieved a theory that is
conformable to every possible observation, past and future? In what sense could the world
then be said to deviate from what the theory claims? Clearly in none, even if we can
somehow make sense of the phrase ‘every possible observation.’ Our overall scientific theory
demands of the world only that it be so structured as to assure the sequences of stimulation
142 APPENDIX A
Further measurements will not help the denizens of w to choose between

the various ways that Euclidean geometry could be instantiated. But con-
siderations of simplicity may well: if they can identify one of the many ways
of building Euclidean space as being distinctly simpler than the others, then
it seems that the philosophers of w would be justified in concluding that this
option is instantiated at their world, just as physicists of w were justified in
concluding that their world had Euclidean structure when they found that
that was the simplest option consistent with the empirical constraints.
This methodological picture has a number of sources in Quine’s other
commitments. On the one hand, he takes the package consisting of
hypothetico-deductivism and a preference for simpler hypotheses to
more or less exhaust what we can say about the scientific method.5 For
Quine, it follows that considerations of simplicity constitute a sort of
evidence according to the standards of science (and of common sense).6
These themes are in evidence in the following passage.
The scientist introduces system into his quest and scrutiny of evidence. System,
moreover, dictates the scientist’s hypotheses themselves: those are most welcome
which are seen to conduce most to simplicity in the overall theory. Predictions,
once they have been deduced from hypotheses, are subject to the discipline of
evidence in turn; but the hypotheses have, at the time of hypothesis, only the
considerations of systematic simplicity to recommend them. Insofar, simplicity
itself—in some sense of this difficult term—counts as a kind of evidence; and
scientists have indeed long tended to look upon the simpler of two hypotheses as
not merely the more likeable, but the more likely.7
Further, Quine insists that the notion of evidence owes its intelligibility to
its ties to everyday use and (especially) to its ties to scientific use—and he
thinks that to question integrity of this notion of evidence is to fall into a
peculiarly philosophical sort of error.8
Something like Quine’s method is followed throughout a wide
swath of contemporary philosophy of physics and analytic metaphysics.
that our theory gives us to expect. More concrete demands are empty . . .” (“Things and
Their Place in Theories,” pp. 246 ff.).
5
See Salmon’s scandalized report, Reality and Rationality, p. 91 n. 12. Quine sometimes
mentions theoretical virtues other than simplicity that can count in favour of hypotheses, but
these further virtues then turn out to be reducible to or trumped by considerations of
simplicity. See e.g. Quine, Word and Object, §§5 f.; and “Posits and Reality,” §I.
6
See “The Scope and Language of Science,” §II; “Posits and Reality,” §III; and “On
Simple Theories of a Complex World.”
7
“Scope and Language,” p. 198. 8
De Caelo, §§I f.
APPENDIX A 143
Although there are very few thoroughgoing Quineans about these days,
many seem to follow Quine in:
(i) Taking ontological inquiry to be (roughly) methodologically
homogeneous with scientific inquiry, in that both depend to a
very large extent on simplicity considerations.
(ii) Holding that one cannot question whether our preference for
simpler accounts in philosophical contexts is grounded in a desire
to know the truth (rather than, say, in aesthetic or pragmatic
considerations) without falling into some sort of sceptical doubt
about the pronouncements of the empirical sciences.
In what follows, I aim to undermine the faith of the orthodox on both of
these points.
Questioning Homogeneity
The hypothetico-deductive method tells us that a hypothesis H is confirmed
by a collection of data E if the members of E can be derived from H
(supplemented, if necessary, by plausible auxiliary hypotheses). Consider a
very simple application: our problem is to determine the curve y ¼ f(x) that
describes the relation between two variables of interest, x and y. So a hypo-
thesis is a guess about what this true curve is. We idealize and assume that our
measurements are perfectly accurate. So a data set consists of finitely many
points in the x-y plane that lie on the true curve: (x1 , f (x1 )), (x2 , f (x2 )), . . . ,
(xn , f (xn )). The hypothetico-deductive method tells us that a given curve is
confirmed by a data set if the points in the data set lie on that curve.
On its own, the hypothetico-deductive method is hopeless as a char-
acterization of the scientific method. This is clear already in our curve-
fitting case: there will be infinitely many curves consistent with any given
data set, of which we regard very few as corresponding to hypotheses
substantially confirmed by the data—but the hypothetico-deductive
method draws no distinction among the curves consistent with the data.
But notice that (in at least some cases) people have pretty strong (and, in
some cases, fairly intersubjective) intuitions about which of the curves
consistent with a given data set correspond to the simplest hypotheses—
and they tend to regard the simplest hypotheses as enjoying a high degree
of confirmation relative to the others. For this reason, it would be a giant
improvement in our account of the scientific method to move from a
144 APPENDIX A
commitment to the hypothetico-deductive method alone to commitment

to a package consisting of the hypothetico-deductive method along with a
principle of simplicity that told us to apportion our preference among the
hypotheses consistent with our data in proportion to their simplicity
(whatever that is . . . ).
Now, relative to this package, ontological and scientific inquiry do
indeed appear to be more or less methodologically homogeneous. But,
Quine aside, very few people view this package as providing more than
a very rough first approximation to the scientific method. And whenever a
property of interest holds relative to a certain approximation, one is
obliged to ask whether it would continue to hold if the approximation
were improved. In our case, this means that those eager to view
ontological investigations as being methodologically homogeneous with
scientific investigations are obliged to ask: Would this homogeneity sur-
vive a transition to a more adequate account of the scientific method?
In fact, it is difficult to answer this question. For there is nothing
approaching a consensus among philosophers concerning the correct
account of the scientific method. But for present purposes it is worthwhile
observing that most serious philosophical accounts of the scientific method
fall into one or more of the following categories. (1) Accounts under
which no notion of simplicity or the like plays any privileged role.9 (2)
Accounts under which some explicit notion of simplicity, parsimony, or
the like plays an important role—but under which it is very implausible
that the notion in question plays any role in discussions of ontology.10 (3)
Accounts under which a limit is placed on the circumstances under which
the scientific method can help us to choose between hypotheses that are in
some sense empirically equivalent—and in which this limit is drawn
in such a way as to leave ontological hypotheses outside looking in.11
9
In this category falls the dominant Bayesian approach, in its standard subjectivist variant;
see e.g. Howson and Urbach, Scientific Reasoning, §11.k. Some objectivist Bayesians seek to
build a bias in favour of simple hypotheses into the notion of rationality; for critical discussion,
see e.g. Kelly, “Ockham’s Razor, Hume’s Problem, Ellsberg’s Paradox, Dilation, and Opti-
mal Truth Conduciveness,” §3 f.
10
In this category one finds e.g the accounts of Glymour, Theory and Evidence and Forster
and Sober, “How to Tell when Simpler, More Unified, or Less Ad Hoc Theories will Provide
More Accurate Predictions.” For a delineation of the obstacles to carrying the latter of these
accounts over to the ontological context, see Sober, “Parsimony and Predictive Equivalence.”
11
One can find such limits drawn (in the course of defences of scientific realism) by e.g.
Earman and Roberts, “Contact with the Nomic. Part II,” §6; Glymour, Theory and Evidence,
pp. 354 ff.; and Roush, Tracking the Truth, ch. 6.
APPENDIX A 145
Of course the observations of the preceding paragraphs come nowhere

near to providing a conclusive case against orthodoxy. For, as noted, there
is no philosophical consensus concerning the nature of the scientific
method. And there are accounts of the scientific method that are not
hostile to orthodoxy—such as the account that takes inference to the
best explanation to be the basic form of ampliative inference.12 But
there should at any rate be enough here to give the orthodox pause: the
claim that ontological and scientific investigations follow essentially the
same method is a bold one, and ought to be regarded as requiring
substantive justification.
Reliable and Unreliable Methods

Let us suppose, however, that under the true account of the method of the
sciences, an important role is played by simplicity considerations of much
the same sort as figure in ontological inquiry. Would it follow that one
could not downplay the force of simplicity considerations in philosophical
contexts (e.g., by tracing our preference for simpler philosophical accounts
back to our desire to work with appealing and tractable objects rather than
to our desire to know the truth) without having to give the same treatment
to the simplicity considerations that figure in scientific contexts? I claim
that it would not: we can regard our preference for simpler scientific
hypotheses as directly rooted in our desire to know the truth without
taking the same view of our preference for simpler ontological accounts.
The question at hand is one about the credentials of the method of
preferring simpler hypotheses. The credentials of a method must be judged
relative to the context in which it is employed. Sipping and tasting is a
good method for distinguishing water from wine—but a poor method for
distinguishing water from heavy water. Similarly, our preference for
simpler hypotheses has much better credentials in application to scientific
problems than in application to ontological problems.
12
For an account of this kind, see Harman, “The Inference to the Best Explanation.”
Note, however, that most advocates of the importance of inference to the best explanation
assign it a much more modest role. Lipton, for instance, sees a preference for explanatory
hypotheses as a heuristic by which computationally limited beings can achieve approximate
obedience to the dictates of Bayesianism; see “Is Explanation a Guide to Inference?” and
Inference to the Best Explanation, ch. 7.
146 APPENDIX A
What makes a method suited to a given problem? According to one

dominant strand in our thought, the mark of a good method is reliability.13
Let us think of a problem as a set of mutually exclusive contingent
hypotheses (not in general exhaustive). And let us think of a method for
a given problem as a rule that tells us for each possible set of evidence: (i)
how to evaluate the hypotheses of the problem relative to the given
evidence; and (ii) what steps to take to collect further evidence. In order
to consider a method suited to a given problem, we demand that it tends
to render our beliefs sensitive in some suitable sense to which of the
hypotheses under consideration is true, if one of them happens to be true.
Consider the method of sipping and tasting. What makes it a good
method for me to use in determining whether a given sample of liquid is
water or wine is that in cases where the sample in fact is one or the other,
the method makes my beliefs very sensitive to which it is. Correlatively,
what makes it a bad method for determining whether a given sample is
water or heavy water is that in the case where the sample in fact is one or
the other, the method doesn’t help me to detect which it is. Similarly, if
the problem is to determine whether there are three people in the yurt or
whether there are four people in the yurt, then looking and counting is a
good method—more or less guaranteed to yield a true belief in the cases
where one of the hypotheses under consideration is true. But if the
problem is to determine whether there are three people in the yurt (as
appears to be case) or whether an evil demon is deceiving me into
believing that there are three people in the yurt, then looking and count-
ing is a useless method—since it presents me with the same evidence and
advice no matter which of the hypotheses under consideration is true.
How to make the relevant notion of sensitivity precise is of course a
very difficult problem. But there are widely accepted sufficient conditions
for sensitivity that would appear to cover much of ordinary and scientific
reasoning.
Let us say that a method is fully reliable for a problem if it is guaranteed to
lead us to the truth eventually (i.e., if we follow the method, we are
guaranteed to be presented with a sequence of evidence sets that will lead
us to strongly favour the true hypothesis from some point onwards—if the
13
There are of course competing tendencies in our thought. Here I aim only to identify
one important respect in which our preference for simpler hypotheses is (part of ) a good
method when applied to scientific questions but not when applied to ontological questions.
APPENDIX A 147
true hypothesis is among those under consideration). Idealized example: if

my keys are somewhere in the house, any procedure that has me look
thoroughly in each room is fully reliable.
More generally, let us say that a method is asymptotically reliable for a
given problem if its continued pursuit is more or less guaranteed to lead us
arbitrarily close to the truth.14 Consider again the problem of curve-
fitting. Suppose that my hypotheses include all curves given by continuous
functions y ¼ f (x) for 0 < x < 1. Consider the method that tells me: (i) to
find a sequence x1 , x2 , . . . dense in the interval (0,1) (so that any 0 < x < 1
can be approximated arbitrarily well by the xi ) and to collect data by first
finding f (x1 ), then finding f (x2 ), then . . . ; and (ii) to conjecture at each
stage that the true curve arises by just connecting each data point to its
immediate neighbours by straight line segments. This method has the
feature that the sequence f1 , f2 , . . . , of curves that it leads me to conjecture
converges point-wise to the true curve f(x).15 Not every method for
handling this problem will share this feature. But it will be shared by any
of the intuitively appealing strategies for curve-fitting, N N such as those
underwritten by principles of simplicity. In this sense, then, our simplicity-
tropic approach is a good one for this sort of problem.16
Both full reliability and asymptotic reliability require a guarantee that
the truth be reached or approached ineluctably. Perhaps this requirement
is too strong. Perhaps we should require only that our method have a
certain probability of approaching the truth. Suppose, for instance, that
you know that your mother tells the truth 90 per cent of the time. She
gives you a coin and tells you that it is her favourite. You begin flipping the
coin. After each toss you calculate the relative frequency p of heads in your
14
Reichenbach argued that our inductive methods are asymptotically reliable and that this
observation provides a sort of solution to Hume’s problem of induction; see e.g. Experience
and Prediction, §§39 ff.
15
i.e., for any 0 < x < 1 and any > 0, there is an N 2 N such that j f (x) fk(x)j < for
k > N.
16
(1) Note that if a problem admits an asymptotically reliable method, then it admits many
of them. From a certain lofty perspective, all of these are on a par—so while it is rational for us
to adhere to some sort of principle of simplicity, it would not be irrational to follow some
other suitable method. But presumably if one takes into account facts about human psych-
ology (computational constraints) and sociology (the need for coordinated action), the field of
live options becomes quite narrow. (2) Note also that for certain problems, simplicity-based
methods enjoy a certain sort of optimality; see Kelly, “Ockham’s Razor.” (3) Note, finally,
that Bayesian agents exhibit a form of asymptotic reliability under certain circumstances; for
discussion and references, see §§6.3–6.6 and 9.4–9.6 of Earman, Bayes or Bust?
148 APPENDIX A
sequence of tosses and conjecture both: (i) that the coin that your mother
gave you has a chance p of coming up heads on any given toss; and (ii) that
your mother’s favourite coin has a chance p of coming up heads on any
given toss. It certainly seems like your method is a good one for deter-
mining the bias of the coin (it leads asymptotically to the truth with
probability one). Whether it is also a good method for determining the
bias of your mother’s favourite coin is a more controversial question.17
But there should not be any controversy concerning the method of
ontology. Consider a world w at which physicists have determined that the
spatial geometry is Euclidean and that the laws of nature are compatible
with either a substantival or a relational understanding of spatial ontology.
Philosophers at w will be interested in the question whether the spatial
ontology of w is in fact substantival or relational. But their method is
patently insensitive to the truth about this matter: they will look for the
simplest total theory (¼physics þ ontology) compatible with their experi-
mental evidence, then take this theory to be true. But what considerations
they weigh and what answer they arrive at will not depend in any way on
what the spatial ontology of their world in fact is—the same articles would
be published by philosophers situated at a world with the other spatial
ontology. In cases of this sort, the standard method of ontology does not
render its practitioners sensitive to which of the hypotheses under consid-
eration is in fact true. As far as sensitivity to the truth goes, the philosophers
of w might as well be pursuing a method that tells them to posit that w has
substantival ontology if and only if Goldbach’s conjecture is true—for, like
Goldbach’s conjecture, the claim that the simplest total theory compatible
with the physics of w is substantival is necessary if true and hence does not
depend in any way on whether w is in fact substantival.
To be led to the truth by a given method always requires an element of
luck—one needs to be at a world at which one of the hypotheses under
consideration obtains. But once this luck is in place, no further luck is
required if one is employing a (fully or asymptotically) reliable method—if
17
On the pro side, see Sober, Reconstructing the Past, §§5.3–5.5. On the con side, note that
there are competing methods that seem equally good but which will sometimes lead to
asymptotic disagreement with yours—e.g. the method that has you roll a ten-sided die before
beginning your coin flips in order to determine whether you are going to take your mother to
have been lying or telling the truth when she told you that the coin was her favourite. If both
methods count as rational, then a certain sort of objectivity fails—there are disagreements
between rational agents that cannot be resolved no matter how much evidence is taken into
account (if not in this example, then in structurally similar ones).
APPENDIX A 149
one plugs away long enough one will latch on to the truth. Not so with
the methods considered in the preceding paragraph—to reach the truth via
such a method requires one to be lucky through and through.
To the extent that a preference for simpler hypotheses is a feature of the
scientific method, this preference can be traced back to our desire for true
theories (because the method as a whole is such that in application to
scientific questions, it can be expected to lead us to the truth—if we stick
with it long enough and if the truth lies in the set of hypotheses we are
investigating). But in many philosophical contexts, no such justification
can be given for our preference for simpler hypotheses—so I think that we
ought to admit that this preference is traceable to aesthetic, pragmatic, and
psychological grounds rather than epistemic ones.
Appendix B
Limits of Sequences of Metric
Spaces
This appendix presents some technical notions (standards of convergence

for sequences of metric spaces) relevant to the Chapters I and II above.1
Gromov–Hausdorff Distance. Our first step is to define a notion of

distance between metric spaces. As a preliminary, we require a notion of
the distance between two subsets within a single metric space.
Let A be a subset of a metric space (X, d). For any r > 0, the r-penumbra
of A in X is:
Ur (A): ¼ {x 2 X: d(x, y) < r for some y 2 A},
(the set of points in X that are within r units of distance of a point in A).
Definition B.1 (Hausdorff Distance). Let (Z, dZ ) be a metric space with
subsets A, B Z. The Hausdorff distance between A and B in Z is
dHZ (A, B): ¼ inf {r > 0: A Ur (B), B Ur (A)}:
(Recall that the infimum (‘inf ’) of a set of real numbers is the greatest lower
bound for that set; correlatively, the supremum (‘sup’) of a set of real
numbers is the least upper bound for that set).
Definition B.2 (Gromov–Hausdorff Distance). Let X and Y be metric
spaces. We call r > 0 admissible if there is a metric space Z with a subspace
X 0 isometric to X and a subspace Y 0 isometric to Y, such that
1
The material that follows is largely drawn from Burago et al., A Course in Metric Geometry,
esp. ch. 7 and §8.1. For generalization to the Lorentzian case, see Bombelli and Noldus, “The
Moduli Space of Isometry Classes of Globally Hyperbolic Spacetimes.”
appendix b 151
dHZ (X, Y ) < r. The Gromov–Hausdorff distance between X and Y, dGH (X, Y ),
is the infimum of the set of admissible r.
Definition B.3 (Gromov–Hausdorff Convergence). A sequence {Xn } of
metric spaces converges in the Gromov–Hausdorff sense to a metric space X if
limn!1 dGH (X, Xn ) ¼ 0.
So dGH takes values in [0,1]. The diameter of a metric space X is defined as
diamX: ¼ sup{d(x, y): x, y 2 X}:
We say that X is bounded if diamX < 1, otherwise we say that it is

unbounded. The Gromov–Hausdorff distance between bounded metric
spaces is always finite. But if X is bounded, then for any metric space Y,
1
dGH (X, Y )$ jdiamX diamY j:
2
So the Gromov–Hausdorff distance between a bounded metric space and
an unbounded one is always infinite. It follows that no sequence of
bounded metric spaces converges in the Gromov–Hausdorff sense to an
unbounded metric space.
Gromov–Hausdorff distance satisfies some but not all of the metric space
axioms. Clearly, dGH (X, Y ) ¼ dGH (Y , X ). And one can show that the
Gromov–Hausdorff distance obeys the triangle inequality. But, as we
have just seen, the Gromov–Hausdorff distance between two metric spaces
can be infinite. And that dGH (X, Y ) ¼ 0 does not imply that X ¼ Y: e.g.,
distinct metric spaces can be at vanishing Gromov-Hausdorff distance from
one another if they are isometric of if one is a dense subset of the other.2
Compact Metric Spaces. But in one important special case the Gro-
mov–Hausdorff distance is as well-behaved as one could desire.
A metric space (X, d ) is totally bounded if for every > 0 one can find
finitely many x1 , . . . , xn 2 X such that every x 2 X is within distance
of one of the xi .3 A metric space is compact if it is complete and totally
bounded. Roughly and intuitively speaking, the compact path metric
2
Recall that Y is dense in X if for any point in X any metric ball at that point includes a
point of Y. Think of the rational numbers as a subset of the real numbers.
3
Total boundedness implies boundedness. The converse is not true in general (think of
the result of gluing together infinitely many copies of [0,1] by their initial points) but it does
hold for complete Riemannian manifolds; see e.g. Lang, Fundamentals of Differential Geometry,
corollary viii.6.7.
152 appendix b
spaces are those that are gap-free and finite in extent—spheres, tori, finite
graphs, etc.4
The Gromov–Hausdorff distance between two compact metric spaces
vanishes if and only if the spaces are isometric and (as noted above) the
distance between bounded metric spaces is always finite. So the Gromov–
Hausdorff distance makes the space of isometry classes of compact metric
spaces into a metric space.5
Whenever we have a metric space, we can ask whether it is complete—
whether any Cauchy sequence of points in the space has a limit in the
space. The space of (isometry classes of ) compact metric spaces is complete
relative to the Gromov–Hausdorff metric.6 The subspace of (isometry
classes of ) compact path metric spaces is likewise complete.7
But the space of (isometry classes of ) compact Riemannian manifolds is
not complete. To see this, note first that Gromov-Hausdorff convergence
does not preserve dimension: some sequences of cylinders have as their
limits line segments; some sequences of tori have as their limits circles; etc.8
The same sort of considerations show that if we consider a sequence of
Riemannian manifolds, each of which consists of a pair of spheres of a
given size connected by a cylindrical neck of a given length but with the
radius of the neck shrinking to zero as the sequence progresses, the
Gromov–Hausdorff limit of the sequence will be a pair of spheres con-
nected by a line segment. Here we have a sequence of Riemannian
manifolds whose limit is not a Riemannian manifold (because some parts
of the space are one-dimensional, others two-dimensional). In fact, every
compact path metric space arises as the limit of a Cauchy sequence of two-
dimensional compact Riemannian manifolds.9
4
Any finite metric space is also compact.
5
This space has the cardinality of the continuum; see remark 7.2.5 of Burago et al., Course.
6
See Gromov, Metric Structures for Riemannian and Non-Riemannian Spaces, §3:11: 12 þ or
Petersen, Riemannian Geometry, proposition 42.
7
Every convergent sequence of path metric spaces converges to a path metric space;
Burago et al., Course, theorem 7.5.1. So the space of compact path metric spaces is a closed
subspace of the complete space of compact metric spaces—and a closed subspace of a
complete space is complete.
8
Use the following useful fact: if f: X ! Y is a surjective map (so that every y 2 Y is the
image of some x 2 X under f ) then dGH (X, Y )# 12 disf , where the distortion of f: X ! Y is:
disf : ¼ sup(jdX (x1 , x2 ) dY ( f (x1 ), f (x2 ))j: x1 , x2 2 X}:
9
Any compact metric space X can be approximated arbitrarily well (in the Gromov–
Hausdorff metric) by finite metric spaces: if X is compact, then it is totally bounded; so for
appendix b 153
Recall that associated with any metric space (X, d ) is a metric space
(X, d), the completion of X, that can be thought of as the result of adding just
enough points to X to ensure that all Cauchy sequences in X have limits.10
The completion of a metric space is complete. Isometric metric spaces
have isometric completions. And any complete metric space is isometric to
its completion. The space of compact Riemannian manifolds, the space of
two-dimensional compact Riemannian manidolds, and the space of finite
graphs are each incomplete relative to the Gromov–Hausdorff metric. And
each has the same completion: the space of all compact path metric spaces.
If the principle of plenitude of possibilities is understood as requiring
that the space of metric geometries that represent possible spatial geom-
etries must itself be complete, then we must conclude that if every two-
dimensional compact Riemannian manifold corresponds to a possible
geometry of space, then so does every compact path metric space. If we
have all of the compact two-dimensional Riemannian manifolds and
everything that can be viewed as a limit of them, then we have all compact
path metric spaces.
Remark B.1 (Alternative Notions of Convergence.). The space of (isom-
etry classes of ) compact Riemannian manifolds, like any large space, can
be equipped with many distinct notions of convergence. Why think that
Gromov–Hausdorff convergence is the one relevant to the principle of
plenitude of possibilities? Why not rely instead on a notion according to
which the limit of any convergent sequence of Riemannian manifolds is
always a well-behaved space?
Gromov–Hausdorff convergence is indeed only one of several widely
used notions of convergence applicable to sequences of Riemannian
each > 0 one can find a finite set X ¼ {x1, . . . , xn} X such that every x 2 X is within of
one of an element of X; equipping the X with the metrics they inherit from X gives us a
family of finite metric spaces that converge to X in the Gromov–Hausdorff sense. If X is also a
path metric space, it can be approximated by finite graphs—in order to find a sequence
of finite graphs that converge to X, one just has to introduce edges that join the points in each
of the finite spaces X in a cunning fashion; one finds the desired family of Riemannian
surfaces by, in effect, embedding the resulting graphs in R3, then thickening their edges up
into hollow tubes. See Burago et al., Course, example 7.4.9, proposition 7.5.5., and exercises
7.5.6 and 7.5.15; and Cassorla, “Approximating Compact Inner Metric Spaces by Surfaces.”
10
Let X̂ be the set of Cauchy sequences of points in X. For any {xn}, {yn} 2 X̂, define
is the space that results when one identifies sequences in X̂
d̂({xn}, {yn} ¼ limk ! 1d(xk, yk). X
that are at zero d̂-distance from one another. d̂ induces a metric on X . The map that sends a
point x 2 X to the constant sequence (i.e., sequence with xk ¼ x for all k) gives us a natural
way of identifying X with a subset of X .
154 appendix b
manifolds.11 These alternative notions are strictly stronger than the notion
of Gromov–Hausdorff convergence—i.e., they consider divergent many
of the sequences that have Gromov–Hausdorff limits.12 Indeed, they do
not allow dimension or topology to change as one passes from a sequence of
Riemannian manifolds to its limit.13 This may sound attractive: it means
that if we cash out the principle of the plenitude of possibilities in terms of
one of these notions rather than in terms of Gromov–Hausdorff conver-
gence, including the Riemannian manifolds among the possible geometries
of space will not immediately force us to also include finite graphs and other
non-smooth spaces, since relative to these notions of convergence the limit
of a sequence of manifolds is always a manifold. But this comes at a very
high price: according to these notions it is impossible to approximate a
Riemannian manifold of a given dimension by a sequence of manifolds
of another dimension. It follows that if we were to cash out the principle of
the plenitude of possibilities in terms of one of these strong notions
of convergence, then the principle would no longer be incompatible
with taking the possible geometries of space to be given by all Riemannian
manifolds except those of dimension seventeen—but I take it that this is just
the sort of exclusion that should count as a gap in logical space. &
Remark B.2 (The Weyl Tile Argument). Some might be tempted to object
that something is already wrong as soon as we speak of a finite set of points
as approximating a continuous Riemannian manifold (as in footnote 9
above). There is a famous line of thought known as the Weyl tile argument
that aims to show that the Euclidean plane cannot be approximated by
discrete spaces of a certain sort.14 Consider the lattice of points in R2 with
integer coordinates and count distances by counting the numbers of nodes
one must pass through to get from one point to the other under the
supposition that one is only allowed to move horizontally and vertically.
11
See e.g. the discussions of uniform convergence and Lipschitz convergence in Burago
et al., Course, ch. 7; and the discussion of Cm,Æ-convergence in Petersen, Riemannian Geometry,
§10.3.2. See also Sormani, “How Riemannian Manifolds Converge.”
12
See Burago et al., Course, p. 260 and Petersen, Riemannian Geometry, 309.
13
i.e., if {Xk} is a sequence of Riemannian manifolds that converges in one of these senses
to a space X, then there is some N such that the Xn are homeomorphic to X for all n > N.
14
See Weyl, Philosophy of Mathematics and Natural Science, p. 43. The kernel of the
argument can be found already in Maimonides (Guide for the Perplexed, §i.73). For discussion
and references, see McDaniel, “Discrete Space and Distance” and Forrest, “Is Space-Time
Discrete or Continuous?”
appendix b 155
Pythagoras’ theorem is violated, and the size of the violation remains the
same even if the interval between nodes is shrunk. The same consider-
ations show that no Riemannian geometry can be locally approximated by
discrete spaces of this kind. What this shows, of course, is that one should
use discrete spaces of a quite different kind if one wants to approximate
locally or infinitesimally Euclidean geometries.15 &
The Non-Compact Case. For non-compact X, one can still consider the
space of all (isometry classes of ) metric spaces within finite Gromov–
Hausdorff distance of X. At the heuristic level, at least, this space is a
well-behaved complete metric space.16
And in this setting one still has the notion of Gromov-Hausdorff
convergence of sequences of metric spaces. One can of course find
examples of sequences of Riemannian manifolds that converge to non-
Riemannian metric spaces.17 For example, let A be the subset of the
Euclidean plane E2 consisting of the coordinate axes; recall that for any
r > 0, the r-penumbra of A in E2 , Ur (A), is the set of points in the plane
within distance r of a point in A; the sequence {U1n (A)} converges to A in
the Gromov–Hausdorff sense.18
However, in general the notion of Gromov–Hausdorff convergence
seems too restrictive in the non-compact setting. For example, it seems
natural to think that a sequence of open intervals like {( n, n)} should have
a non-compact limit—but as noted above, a sequence of bounded spaces can
never converge in the Gromov–Hausdorff sense to an unbounded space.
There is a widely used notion of convergence for non-compact metric
spaces which handles such examples well. Roughly and intuitively speak-
ing, the idea is that X is a limit of {Xn } in this new sense if balls of arbitrary
radius around some given point in X are approximated by balls in the Xn .
Definition B.4 (Pointed Metric Space). A pointed metric space is a pair (X, a)
where X is a metric space and a 2 X. The point a is called the distinguished
point of (X, a).
15
See Forrest, “Discrete or Continuous?”
16
See Gromov, Metric Structures, §3:11 12 þ . For a more cautious treatment, see Eichhorn,
Global Analysis on Open Manifolds, §5.1. Complications arise here due to the fact that if X and
Y are complete but non-compact, one can have dGH(X, Y ) ¼ 0 even when X and Y are non-
isometric (thanks here to Peter Petersen).
17
But a complete space that is a limit of a convergent sequence of path metric spaces is
always a path metric space.
18
This follows from the result quoted in fn. 8 above.
156 appendix b
Definition B.5 (Pointed Gromov–Hausdorff Convergence). A sequence

{(Xn , an )} of pointed metric spaces converges in the pointed Gromov–
Hausdorff topology to a pointed metric space (X, a) if for every r > 0 and
> 0 there exists an N 2 N such for for each n > N there is a map
f : Br (an ) ! X such that: (i) f (an ) ¼ a; (ii) the distortion of f is less than ;
and (iii) every point in Br (a) X lies within of a point in f (Br (an )).19
Such limits are essentially unique for well-behaved spaces: if a given
sequence of pointed metric spaces converges in the pointed Gromov–
Hausdorff topology to each of two metric spaces (X, a) and (Y, b), then
either X and Y are isometric (as pointed spaces), or at least one is not
complete, or neither is boundedly compact.20
When all the spaces involved are compact, pointed Gromov–Hausdorff
convergence reduces to the usual Gromov–Hausdorff notion of conver-
gence—in particular the choice of distinguished points is immaterial. But
when the limit is non-compact, the choice of distinguished points can be
important.
Example B.1 (Convergence of Unbounded Intervals). Consider the se-
quence of intervals {[0,2n]}. It is not hard to see that this sequence
converges either to the real numbers or to the non-negative real numbers,
depending on how one chooses the distinguished points: the sequence of
pointed spaces {([0,2n],0)} converges to the pointed space (R$0 , 0) while
the sequence {([0,2n],n)} converges to the pointed space (R,0). &
Remark B.3 (Metric Characterization of Riemannian Manifolds). Let (X, d)
be a metric space. Then for any º > 0, the function ºd: (x, y) 2 X X 7! º
d(x, y) is also a metric on X. Let us call a metric space (X,d ) infinitesimally
Euclidean if at every point x 2 X the pointed metric spaces (X, ºd, x)
converge in the Gromov–Hausdorff sense to (Rn , d, 0) as º ! 1 (here d
is the Euclidean metric on Rn and 0 is the origin). A metric space is a
Riemannian manifold if and only if it is a locally compact infinitesimally
Euclidean path metric space.21 &
19
For the notion of the distortion of a map, see fn. 8 above.
20
See Burago et al., Course, theorem 8.1.7. A metric space is boundedly compact if all of its
closed bounded subsets are compact.
21
See Gromov, Metric Structures, pp. 85 f. A metric space X is locally compact if sufficiently
small open balls have compact closures. For path metric spaces, local compactness is equiva-
lent to being boundedly compact; see Plaut, “Metric Spaces of Curvature $ k,” theorem 8.
Appendix C
Some Background to the
Absolute-Relational Debate
It is traditional among philosophers of space and time to approach the

absolute-relational debate about the nature of space and motion via the
views of Descartes, Newton, and Leibniz. In briefest outline, the story runs
as follows.1 (1) Descartes defined the place of a body via its relations to its
immediate neighbours and motion as change of place so conceived. He
attempted to base upon this notion a mechanistic physics governed by
rules of impact and the principle of inertia. The result was a disappointing
mess. (2) Newton defined the state of motion of a body as its motion
relative to absolute space—the latter he thought of as a non-material
existent, neither substance nor accident, consisting of parts that maintain
their identity and geometric relations to one another over time. This
provided the conceptual underpinnings for his laws of motion. The result
was a resounding success. (3) Leibniz’s criticisms of Descartes’s and New-
ton’s accounts of space and motion have exercised a considerable influence
on natural philosophical discussions down to the present day. But his
attempts at constructing a systematic and credible competitor to the
physics of his rivals never came to fruition—in part, it seems, because
Leibniz was pulled in several incompatible directions by his critical intu-
itions. In particular, it is hard to see how to fit together his relational
account of space (which would seem to undercut any notion of absolute
motion) with his views about force (which would appear to ground
absolute notions of motion).2
1
The canonical sources for this tale are Stein, “Newtonian Space-Time” and “Some
Philosophical Prehistory of General Relativity;” and Earman, World Enough and Space–Time.
2
For an interesting attempt to resolve this tension, see Jauernig, “Leibniz on Motion and
the Equivalence of Hypotheses.”
158 APPENDIX C
Of course, philosophical debates about the nature of motion and space

began long before the seventeenth century. Indeed, the views of Descartes
and Newton are closely related to certain ancient views. And in antiquity,
through the middle ages, and down to the seventeenth century, one finds
many arguments for and against these views, including precursors of some
of those arguments from Newton’s writings and from the Leibniz–Clarke
correspondence that continue to drive much of the philosophy of space
and time. All of this is of course well known to historians of these periods.
The following is a sort of bibliographical essay that concludes with a few
remarks about what seems to have been genuinely novel to the seven-
teenth century context.
***
A natural place to begin is with Aristotle’s account of the cosmos, place,
and motion. The finite material world is organized into a spherical cosmos.
The Earth sits at rest at the centre of this cosmos; above the atmosphere are
several nested spheres rotating about the Earth; the Moon, Sun, planets,
and stars are fixed to these spheres; and the composition of the circular
motions of these spheres gives rise to the motions of the heavenly bodies
through the sky. The matter of the Aristotelian cosmos forms a plenum.
Thus every body is surrounded by matter; this allows Aristotle to take the
place of a body to be “the boundary of the containing body at which it is
in contact with the contained body.”3 Or at least, this works for bodies in
the interior of the cosmos: “the heaven . . . is not anywhere as a whole, nor
in any place, if at least, as we must suppose, no body contains it.”4 We
nonetheless say that the outermost sphere of the heavens—the sphere of
the fixed stars—completes one revolution each day.5
Aristotle had to contend with atomists, who claimed that movement
would be impossible in a plenum and posited a cosmology involving infin-
itely many atoms moving through an infinite void.6 To the atomists’ claim
about the impossibility of motion in a plenum, Aristotle retorted, reasonably
enough, that “not even movement in respect of place involves a void; for
3
Physics, IV.4 212a5–7. Translation of McKeon (ed.), Aristotle.
4
Ibid. 212b8–10. Translation of McKeon (ed.), Aristotle.
5
I believe that we are supposed to reach this conclusion by regarding the Earth as fixed,
then examining the relative motion between each of the surrounding spheres (see below). For
a survey of ancient reactions, see Sorabji, Matter, Space and Motion, pp. 193–6. For further
discussion, see Morison, On Location, pp. 166–9.
6
On the evolution of the atomist’ notion of void, see Sedley, “Two Conceptions of
Vacuum.”
APPENDIX C 159
bodies may simultaneously make room for one another, though there is no
interval separate and apart from the bodies that are in movement. And this is
plain even in the rotation of continuous things, as in that of liquids.”7
Aristotle also provides positive arguments against the possibility of a
void—not only is there no space empty of matter within the cosmos, but
the cosmos itself is not to be thought of as immersed in a larger void space.
(1) In De Caelo, void is characterized as “that in which the presence of
body, though not actual, is possible . . .”8 But it is not possible for there to
be matter beyond the cosmos: such matter could not be there naturally, for
the natural place of earth, water, air, fire, and the heavenly material is
within the cosmos; nor could it have gotten there by violence, for in that
case it would have to be located in the natural place of some other
matter—and there is none such. So an extra-cosmic void is impossible.9
The same argument is supposed to show that there can be no other cosmoi
located outside of our own.
(2) In the Physics, Aristotle tells us that the partisans of the void regard
“it as a sort of place or vessel which is supposed to be ‘full’ when it holds
the bulk which it is capable of containing, ‘void’ when it is deprived of
that—as if ‘void’ and ‘full’ and ‘place’ denoted the same thing, though the
essence of the three is different.”10 Aristotle offers a series of objections to
the void in Book IV, chapter 8, showing that a body immersed in a void
would be both motionless and move with an infinite velocity, etc. These
arguments turn upon the details of the Aristotelian account of natural
place, motion through resisting media, etc., and they exercised a consid-
erable influence on medieval discussions of the possibility and nature of
motion in a void.11 To these we can add the following remark that occurs
in the preamble to Aristotle’s discussion of place: “place cannot be body;
for if it were there would be two bodies in the same place.”12 Many
of Aristotle’s medieval successors saw here a powerful consideration against
the possibility of the void. For if the void is conceived of as a sort of
7
8
Physics, I.9 279a14–15. Translation of McKeon (ed.), Aristotle.
9
There is some reason to think that that Aristotle here assumes that something is possible
only if it happens at some time or other; see Hahm, The Origins of Stoic Cosmology, p. 103
esp. fn. 32.
10
11
See Grant, Much Ado About Nothing, ch. 3.
12
Physics, IV.1 209a6–7; translation of McKeon (ed.), Aristotle. See also IV.8 216a34–b10.
160 APPENDIX C
three-dimensional entity capable of being filled by body, then we must

accept that when it is so filled, we have two things existing in the same
place—an absurdity.13
***
These views of Aristotle were subject to sustained criticism throughout
antiquity, among the Scholastics, and in the early modern period. Let me
begin by noting three types of anti-Aristotelian argument that are of
special interest for present purposes. (The following treatment is of course
a highly selective one.)
(1) Paradoxes of Aristotelian Motion. If, as is natural, ( local ) movement
is understood as change of place, then the Aristotelian definition
of place leads to counter-intuitive consequences: a body such as a
tower moves (because air surrounding it constantly circulates);
similarly, bodies can approach one another even if neither moves.
These observations provide an argument against the Aristotelian
accounts of place and local motion.
(2) Arguments from Cosmic Size and Shape. Many find it plausible that
the cosmos does or could change shape, or that it could have had
a different size or shape from its actual one—and this seems to
suggest that there must be void outside of the cosmos.
(3) Arguments from Possible Motions. Our intuitions recognize the
possible states of motions which must be understood as motion
relative to the parts of a separately existing void—no account of
motion in terms of the relations between material parts will
suffice. This provides another sort of argument in favour of the
void.
***
I will make some remarks about the history of each of these families of
objections, beginning with the Paradoxes of Aristotelian Motion.
Under this heading we find arguments directed against the Aristotelian
definition of place. They appear to have first emerged in the writings of
Aristotle’s immediate successor, Theophrastus.14 They seem to have
13
For the medieval influence of this argument, see Grant, Much Ado, pp. 32 ff.
14
See Sorabji, Matter, Space and Motion, ch. 11.
APPENDIX C 161
played some role in the rejection by the Aristotelian majority in antiquity

of Aristotle’s conception of the place of a body as the boundary of the
surrounding bodies.15
During the middle ages, Aristotle’s account of place was again widely
accepted.16 And the paradoxes of motion were then rediscovered and
widely discussed.17 In the seventeenth century, they were available even
to non-Scholastic philosophers in, e.g., the Physiologia Epicuro-Gassendo-
Charltoniana of Walter Charleton.18
It is clear that these arguments create difficulties for Aristotle. He
affirms in his Physics that: “It is always with respect to substance or to
quantity or to quality or to place that what changes changes.”19 Later he
is quite specific in identifying locomotion with change of place.20 And
so it seems clear that he regards a body as moving (in our sense) if and
only if there is a change along its immediate boundary. And so, prima
facie, it seems that a boat moored in a strong current will count as
moving, while one drifting downstream along with the current may
count as at rest (depending on whether we require the individual
parts of water along its surface to be at relative rest, and whether they
in fact are).
Now, it isn’t clear whether the charge in this form will stick. At one
point Aristotle departs from his original characterization of the motion of a
body in terms of what is happening at the immediate boundary of the
body and maintains instead that:
when what is within a thing which is moved, is moved and changes its place, as a
boat on a river, what contains plays the part of a vessel rather than that of place.
Place on the other hand is rather what is motionless: so it is rather the whole river
that is place, because as a whole it is motionless. Hence we conclude that the
innermost motionless boundary of what contains is place.21
15
On this, see Sorabji, Matter Spaceam Motica pp. 199–201.
16
Grant, “Place and Space in Medieval Physical Thought,” p. 154.
17
See Grant, “The Medieval Doctrine of Place,” §2 and Much Ado, p. 125.
18
See p. 69. This work is an eccentrically augmented free translation of a work by
Gassendi—and it played a pivotal role in making available in English Gassendi’s attempts to
Christianize and modernize atomism. Newton is known to have read this work carefully as an
undergraduate; see Westfall “The Foundations of Newton’s Philosophy of Nature,” p. 172
esp. fn. 5.
19
Physics, III.1 200b33–4. Translation of McKeon (ed.), Aristotle.
20
Ibid. VIII.6 260a27–8.
21
Ibid. IV.4 212a15–19.
162 APPENDIX C
Here is Myles Burnyeat’s influential reading of this passage:

The point of the refinement is this: the place of X was to be the boundary of Y
enclosing X, but if Y is moving, this specifies a carrier or vessel of X rather than X’s
place . . . The solution is to find Z such that Z is static and Z encloses X at the same
boundary as Y does. Example: X¼a boat, Y¼the body of water flowing in the
Cayster, Z¼the river Cayster as a geographical entity.22
There is a question of coherence here. Our system of judgements about

place and motion will be founded upon an initial choice of a body that
counts as motionless. Considering a different body as motionless at the
beginning would result in different judgements about place and motion.
Now in Aristotle’s scheme, it is clearly safe to count the Earth as motionless,
and to work outwards from there. But then place is specified by position
relative to the surface of the Earth, and motion by change of distance with
respect to reference points on the surface of the Earth. So Aristotle’s
attempt to shore up his definitions of place and motion lead quickly to
their supersession by quite different ones. Indeed, one strand of Scholastic
thought followed this course, referring motion ultimately to change of a
body’s relation to the immobile centre and poles of the cosmic sphere.23
***
Now we turn to the positive arguments offered by proponents of the
void—the arguments from cosmic size and shape and the arguments from
cosmic motion. Here it is convenient to discuss both arguments together,
moving from one group of commentators to the next.
THE ATOMISTS. The atomist cosmology featured an infinite number of

indivisible particles moving in an infinite void. Our cosmos formed by
chance, and will eventually decay—it is one of an infinite number of
cosmoi.24
Against the finite spherical universe of Aristotle, Lucretius deploys an
argument attributed by ancient authors to Archytas (contemporary of
Plato and teacher of Eudoxus): if you are situated at the edge of the
22
“The Sceptic in His Place and Time,” 102 n. 15. For further discussion, see Morison,
On Location, ch. 5.
23
Grant, “Medieval Doctrine of Place,” §3. Presumably it went unnoticed that a body
moving along the equator of the cosmic sphere would count as immobile according to this
criterion.
24
It has been argued that one should not attribute to the early atomists the account of void
of Epicurus and Lucretius discussed below; see Sedley, “Two Conceptions.”
APPENDIX C 163
cosmos, what happens if you extend your staff (or spear, or sword, . . . )
beyond the edge? If there is something there to prevent its extension, then
you are not yet at the edge—there is further matter. On the other hand, if
you are successful, then there must be receptive void. Repeating the
argument whenever a new putative boundary is reached shows that
there is infinite extension—of either matter or void.25
Lucretius also gives two detailed arguments in favour of the void, defined
as “intangible empty space.”26 The first rests upon the traditional atomist
contention that motion would be impossible in a plenum.27 The second,
cleaned up and amplified, proceeds thus: suppose two bodies in contact
along a surface move away from one another; then air must fill the space
between the surfaces initially in contact; but if it moves with only finite
velocity, there will be void immediately after the separation of the bodies.28
So far, the arguments given allow us to think of the void of the atomists
either as being something like the space of modal relationalists or as being
something like Newton’s absolute space. But following the arguments just
discussed, Lucretius remarks that:
If there were no place and space, which we call void,
Bodies could not be situated anywhere
And they would totally lack the power of movement,
As I explained a little while ago.29
Now, Lucretius has earlier told us that bodies could not move if there were
no void. But that they would be situated nowhere appears to be a new
25
De Rerum Natura, I.968–83. For discussion of origins of this argument and of Aristotel-
ian responses in antiquity, see Sorabji, Matter, Space and Motion, pp. 125–8. For Scholastic
responses, see Grant, Much Ado, pp. 106–8. The argument also appears in More and Gassendi;
see Koyré, From the Closed World to the Infinite Universe, p. 123 and Grant, op. cit., 389 n. 168.
Here is another popular atomist argument: that which is limited must be limited by some-
thing. For this latter argument, see Epicurus, Letter to Herodotus, §41 and Lucretius, De Rerum
Natura, I.957–65. For discussion, see Sorabji, op. cit., 136–8.
26
De Rerum Natura, I.334. Translation of Fowler and Fowler (eds.), Lucretius.
27
Ibid. I.335–45 and I.370–83. See also Epicurus, Letter to Herodotus, §40. Note Aristotle’s
response, Physics, IV.7 214a28–32. See Charleton, Physiologia, p. 19. Note: the undergraduate
Newton appears to have accepted this argument; Westfall, “The Foundations of Newton’s
Philosophy of Nature,” p. 174.
28
De Rerum Natura, I.384–9. See Grant, Much Ado, §4.E for Scholastic responses to this
sort of challenge.
29
Ibid. I.426–9. Translation of Fowler and Fowler (eds.), Lucretius. See also Epicurus, Letter
to Herodotus, §40: “And if there did not exist that which we call void and space and intangible
nature, bodies would have no place to be in or move through, as they obviously do move;”
translation of Inwood and Gerson (eds.), Epicurus, p. 6.
164 APPENDIX C
thought—and one on which he never really elaborates. I do not think it

much of a stretch to think of Lucretius as taking for granted the absolutist
conception here: his void, as an infinite three-dimensional non-corporeal
entity provides a standard of place and movement for bodies—a body
changes place if it occupies a new portion of void and the state of motion
of a body is referred to its change of place in the void.30
Indeed, it is not easy to see how we can otherwise make sense of
certain characteristic atomist theses. At least from Epicurus onward,
atomists held that the void has a natural distinguished direction, down-
wards, and that the natural motion of atoms is downwards, with atoms of
all sizes moving at the same speed.31 In Lucretius we find that this natural
motion is sporadically interrupted by the mysterious swerve which puts
atoms on collision courses.32 These collisions are ultimately responsible
for the formation of cosmic vortices. As Lucretius notes, without the
swerve, the atoms
Would fall like drops of rain through the void.
There would be no collisions, no impacts
Of atoms upon atom, so that nature
Would never have created anything.33
The most obvious way for us to make sense of this is to refer the motion of
atoms to the parts of the void, conceived of as retaining their identity and
relations to one another over time. For if in the swerveless atomist universe
we look at the relations just between the atoms, we find them utterly
static—and we would have no reason to maintain that the atoms were
falling down like drops of rain rather than sitting motionless.
THE STOICS. The Stoics, while accepting a spherical and void-free cosmos,
explicitly located it within an infinite void.34 Now, the Stoics more or less
30
One might worry about this interpretation, on the grounds that the atomists spoke of
void as non-being. But then, they also seem to have thought that the existence of void shows
that the non-existent is just as real as the existent. For discussion and references, see Barnes,
The Presocratic Philosophers, §XIX(b).
31
For discussion of pre-Epicurean atomism, see chs. 9 and 10 of Furley, The Greek
Cosmologists, vol. 1.
32
According to ancient authorities, Epicurus also held this view; see Inwood and Gerson
(eds.), Epicurus, p. 47.
33
De Rerum Natura, II.222–5; translation of Fowler and Fowler (eds.), Lucretius.
34
For discussion and references, see e.g. Hahm Origins, ch. IV and Sambursky, Physics of the
Stoics, ch. IV. For deflationary readings of the infinitude of the Stoic void, see Todd,
APPENDIX C 165
accept Aristotle’s terms—void is that which is capable of being occupied

by matter, but is not so occupied. But in favour of the void they offer
arguments of the sort that we are interested in.
(1) The cosmos will be or could have been a different shape—so there must be
receptive void. The Stoic cosmos is subject to periodic destruction by
conflagration, during which the volume of matter is increased manyfold.
Thus, Cleomedes in The Heavens:
If, according to the doctrine of the most accomplished natural philosophers, the
whole substance [of the cosmos] is also reduced to fire, it must occupy an
immensely larger place, as do solid bodies that are vaporized into fumes. Therefore
the place occupied in the conflagration by the substance [of the cosmos] when it
expands is currently void, since no body fills it.35
So there exists at least some void outside of the cosmos. According to

Simplicius, some Stoics employed Archytas’ argument to show that the
void must in fact be infinite.36
(2) The possibility of motion of the entire world shows that there must be an
infinite void. This argument appears in Cleomedes:
We can also conceive of the cosmos itself moving from the place that it currently
happens to occupy, and together with this displacement of it we shall also at the
same time conceive of the place abandoned by the cosmos as void, and the place
into which it is transferred as taken over and occupied by it. The latter [place] must
be filled void.37
“Cleomedes and the Stoic Concept of the Void” and Inwood “Chrysippus on Extension and
the Void.” Posidonius appears to have held a heterodox view, according to which the extra-
cosmic void was only just large enough to hold the cosmos at its time of maximum expansion;
see Algra, “Posidonius’ Conception of the Extra-Cosmic Void.”
35
Bowen and Todd (eds.), Cleomedes’ Lectures, p. 24. For further discussion and references,
see Sorabji, Matter, Space and Motion, p. 129 and Hahm, Origins, p. 106. Note that Cleomedes
continues: “But if anyone claims that a conflagration does not occur, such a claim would not
confute the existence of the void. For even if we merely conceived of the substance [of the
cosmos] expanding, that is, being further extended (granted that there is no possible obstacle
to such extension), then this very thing into which it would be conceived as entering in its
extension would be void.” Here it is important that the Stoics employed a relatively liberal
notion of possibility. See Hahm, Origins, p. 103 for references and discussion.
36
See Sorabji, Matter, Space and Motion, pp. 126 ff. and Hahm, Origins, p. 106.
37
Bowen and Todd (eds.), Cleomedes’ Lectures, p. 24. (1) Cleomedes himself denied that
the cosmos was in fact in motion (see ibid. 26)—but this would seem to be perfectly
consistent with taking the possibility of such motion to establish the infinitude of the void,
in analogy with the way that the mere possibility of conflagration establishes the existence of
the void (see fn. 35 above). (2) Achilles the Grammarian records the following Stoic
166 APPENDIX C
Presumably we should add: But there is no limit to the direction, velocity,

or duration of this movement, so we must conceive of the void as being
infinitely extended in all directions.
In Archytas’ argument and in the argument from the conflagration, the
existence of the void functions only as a sort of place-holder for possible
deformations or expansions of the cosmos—and so is compatible with an
understanding of the void as something like the empty space of a modern
modal relationalist and with a broadly Aristotelian account of motion. But
with Cleomedes’ thought experiment regarding the possible motion of
the cosmos as a whole this is no longer possible—the thought experiment
is only coherent if the void itself plays a role in defining place and motion.
Cleomedes wants us to judge that in the situation described the cosmos is
moving through the void—because it successively occupies different parts
of the void, rather than because of any characteristic relative motion
between its parts. This suggests that, for some Stoics at least, the void
ought to be viewed as an infinite three-dimensional entity, whose parts
maintain their identity over time and provide the ultimate grounding for
the notions of place and motion.38
THE SCHOLASTICS. No brief summary can do justice to the full range of

Scholastic mutations of Aristotelianism. From Edward Grant I take the
following points.
(1) The Aristotelian account of place remained essentially unchallenged
throughout the medieval period.39 There was, however, active discussion
of the paradoxes of motion and problems regarding the motion of cosmic
sphere.
(2) There was a unanimous consensus among medieval Scholastics that
the cosmos could not be thought of as immersed in an extended, three-
dimensional void.40 Grant identifies a theological basis for this consensus,
in Scholastic reluctance to recognize any infinite being in addition to
argument: “If the cosmos were moving down in an infinite void, rain would not overtake the
earth. But it does. Therefore the cosmos does not move but stands still;” quoted at pp. 109 f.
of Hahm, Origins.
38
Chrysippus, on the other hand, seems to have denied, on broadly relationalist grounds,
that the hypothesis of cosmic motion was a sensible one. For discussion and references, see
Hahm, Origins, p. 122.
39
Grant, “Place and Space,” p. 154.
40
Grant, Much Ado, p. 180.
APPENDIX C 167
God.41 In this context, Aristotle’s complaint that if void could be occupied

by body, then two things would be in the same place was widely accepted
as a decisive argument.42
(3) This position seems entirely compatible with Scholastic use of argu-
ments showing that the world could have been larger that it was, or differently
shaped. Archytas’ argument was communicated to the Scholastics in works of
Simplicius, and was afterwards widely discussed.43 It was also widely accepted
that God could have chosen to create a larger world than he had.44 But, of
course, this is consistent with the insistence that the extra-cosmic void is not
an extended entity.
(4) In 1277, Parisian theologians, fighting a rearguard action against
Aristotelians in the faculty of arts, managed to have a number of proposi-
tions condemned by the Bishop of Paris. For a time, the teaching of these
propositions was punishable by excommunication. Even after this penalty
was lifted, the condemnation continued to have an effect: the condemned
propositions continued to be eschewed by conscientious writers. Among
the propositions condemned, we find the following.45
That there is no more excellent state than to study philosophy.
That the only wise men in the world are philosophers.
That one should not hold anything unless it is self-evident or can be
manifested from self-evident principles.
That if the heaven stood still, fire would not burn flax because God
would not exist.
That a sphere is the immediate efficient cause of all forms.
That it pertains to the dignity of the higher cause to be able to
commit errors and produce monsters unintentionally, since nature is
able to do this.
That the intellect of the dead Socrates does not have the science of
those things of which it once had science.
41
This was not, however, viewed as problematic by early Christians; see Sambursky, The
Concept of Place in Late Neoplatonism, pp. 14–17.
42
See Grant, “Medieval and Seventeenth-Century Conceptions of Infinite Void Space
Beyond the Cosmos” and chs. 1 and 2 of Much Ado.
43
Grant, Much Ado, pp. 106 f.
44
Ibid. 137. See also Sorabji Matter, Space and Motion, p. 129.
45
These are propositions 1, 2, 4, 79, 81, 97, 149, 178, 203, 205, and 66 in the numbering
and translation found in Lerner and Mahdi (eds.), Medieval Political Philosophy, selection 18.
168 APPENDIX C
That by certain signs one knows men’s intentions and changes of

intention, and whether these intentions are to be carried out, and that
by means of these prefigurations one knows the arrival of strangers,
the enslavement of men, the release of captives, and whether those
who are coming are acquaintances of thieves.
That one should not confess, except for the sake of appearance.
That simple fornication, namely that of an unmarried man with an
unmarried woman, is not a sin.
That God could not move the heaven in a straight line, the reason
being that He would then leave a vacuum.
It has been argued that the inclusion of this last proposition had
momentous consequences for the development of the concept of
space—for in the fourteenth century one finds a number of Scholastics
happy to say that God could move the cosmos through the void, or that
God was faced with a choice about where in the void to create the
cosmos.46 It is difficult to see how the possibility of the translation of the
world as a whole along a straight line can be underwritten by anything
short of an extended void whose parts maintain their identity through
time and (thus) provide a standard of place and motion independent of
body.
THE EARLY MODERN ATOMISTS. Spurred in part by a flood of ancient texts

previously unavailable in Europe, the sixteenth and seventeenth centuries
saw the discussion of a wide variety of non-Aristotelian accounts of place,
space, void, motion, matter and the structure of the cosmos.47 Here we
note one particular strand of development which prefigured Newton’s
absolutist accounts of space and motion: Gassendi’s attempt to revive and
Christianize ancient atomism.
Gassendi self-consciously mines ancient and Scholastic authors for argu-
ments. His cosmology features a single material world, created by God and
composed of atoms, immersed in an infinite three-dimensional void space,
itself neither substance nor accident but suffused with the omnipresence of
46
For discussion and references, see Grant, “The Condemnation of 1277, God’s Absolute
Power, and Physical Thought in the Late Middle Ages” and Lindberg, The Beginnings of
Western Science, pp. 233–44.
47
For helpful discussion, see Koyré, From the Closed World and Grant, Much Ado, chs. 7 and 8.
APPENDIX C 169
God.48 In Gassendi and/or his English disciple Charleton, we find the

following arguments and claims. (1) The paradoxes of motion cause diffi-
culties for any Aristotelian account of motion.49 (2) The argument of
Archytas for the existence of an infinite void.50 (3) God could have created
the universe larger than it is or could repeatedly annihilate the universe and
create a larger version—so the void must be infinite.51 (4) We can conceive
God moving the material world from one location to another.52
***
Obviously this is only the tip of the iceberg. But I hope to have given some
feeling for the wealth of interesting arguments and theses salient to the
absolute-relational debate that pre-date Descartes, Newton, and Leibniz,
but which reverberate through, e.g., Newton’s De Gravitatione and the
Leibniz–Clarke correspondence. In closing this discussion I would like to
make a few remarks about what does appear to have been new in
discussion of space and motion in the seventeenth century.
The mathematical physics of the seventeenth century took over from
astronomy the practice of representing the motions of bodies by curves in
Euclidean space, parameterized by time.53 The course of the century saw a
progressive widening of the scope and ambitions of this new physics, with
its dynamical treatment of the motion of bodies: from its first specimens in
Galileo’s treatment of free fall and projectile motion near the Earth; to
Descartes’s qualitative modelling of the celestial motions via vortices; to
the competing quantitative accounts of the system of the world offered by
Newton and the later vortex theorists (including Leibniz). The first half of
the seventeenth century also saw the decisive rejection by astronomers and
natural philosophers of Ptolemaic astronomy and the Aristotelian cosmol-
ogy in which it was set. Of course, these two sets of developments were
48
For discussion, see Lolordo, Pierre Gassendi and the Birth of Early Modern Philosophy,
pp. 106–8. For references to others who rejected the substance-attribute dichotomy in the
case of space, see Grant, Much Ado, pp. 187, 199, 204, 217, 240, and 392 nn. 182 and 185.
49
Charleton, Physiologia, p. 69.
50
For references and discussion, see Grant, Much Ado, p. 389 n. 168.
51
See Charleton, Physiologia, p. 11 and Brush (ed.), Gassendi, p. 387. For discussion, see
Lolordo, Gassendi, pp. 109 ff.
52
Brush (ed.), Gassendi, 388; Charleton, Physiologia, 67 f. Gassendi also says that God faces
a choice in deciding where in space to create the world; see the passage quoted on p. 110 of
Lolordo, Gassendi.
53
Leibniz is of course no exception. See e.g. the passages at Loemker (ed.), Leibniz,
pp. 577 and 668.
170 APPENDIX C
related to one another in many ways. I would like to emphasize just one
aspect by claiming that the transition from Aristotelian cosmology to the
new cosmologies of the seventeenth century undermined the most
straightforward route to interpreting curves in Euclidean space as repre-
senting the motions of bodies; and that the competing accounts of the
nature of space, the nature of motion, and the relation between the two
that one finds in Descartes, Newton, and Leibniz can be viewed as aspects
of the process of recognition and resolution of this problem.
From Galileo onwards, the new mechanics was based on one form or
another of the principle of inertia, according to which bodies free from
interference naturally tend to trace out a certain sort of curve in space. The
interpretation of curves in Euclidean space as representing the motion of
bodies is unproblematic in contexts in which the motion of all bodies can
be understood as motions relative to a natural reference body. For then
(to speak anachronistically) one can regard the curves as describing motion
in the space picked out by coordinate axes attached to the reference body.
The location of a moving body relative to the fixed body is determined at
each moment of time by the parameterization of the geometric curve
associated with the moving body.
In the mainstream cosmological tradition deriving from Aristotle and
Ptolemy, the Earth is at rest at the centre of a finite series of rotating
material spheres which exhaust the contents of the universe. In this
context the Earth provides a geometrically privileged, fixed body—the
natural reference body to which the complicated trajectories of Ptolemaic
astronomy can be referred.
For Copernicus and Kepler, the cosmos is still spherical, and both the
central sun and the outer surface which encloses the fixed stars are
immobile, and are suitable to serve as reference bodies.54 According to
Copernicus, the stars are fixed to the surface of the outermost sphere;
according to Kepler they are scattered throughout a shell within the
outermost sphere, with the shell enclosing a void in which the solar system
is located. Copernicus is quite explicit: “the first and supreme of all is the
sphere of the fixed stars which contains everything and itself and which,
therefore, is at rest. Indeed, it is the place of the world to which are
referred the motion and the position of all other stars.”55
54
On this point, see Koyré, From the Closed World, pp. 29–34 and 76–87.
55
Quoted at Koyré, From the Closed World, p. 33. Of course here the planets, including the
earth, are numbered among the “other stars.”
APPENDIX C 171
Galileo, on the other hand, is able to understand the curves that

terrestrial bodies trace out in his mechanics as curves relative to the
Earth, treated as fixed. But of course, he is also a partisan of the Copernican
system, and maintains against Tycho Brahe and Ptolemy that the Earth
rotates daily and moves through the heavens annually. And he can make
sense of these claims, if he wishes—for like Copernicus and Kepler he
maintains that “the fixed stars (which are so many suns) agree with our sun
in enjoying perpetual rest.”56
But as the century progressed, new cosmologies emerged in which
the Earth orbits the Sun along with the other planets, the Sun itself is
just another star, and the stars are scattered haphazardly through
space, each being constantly jostled by the fluid or ether in which it is
immersed.
In this new context, neither the Earth, nor the Sun, nor the set of
“fixed” stars any longer provides a natural reference body for the inter-
pretation of the motion of bodies in terms of geometric curves—the
cosmos has no centre, and there is no body that could naturally be taken
to be at rest. What is needed is an account of motion that refers motions to
something other than body, or one which grapples directly with the fact
that only some reference bodies are suitable to refer motions to (in the
sense that the law of inertia does not hold if all motion is referred to a body
in an arbitrary state of motion)—and that such bodies, if they exist, need
not be of any particular astronomical interest.
It was of course Newton who first saw clearly the difficulties involved.
He showed that Descartes’s analysis of motion in terms of the separation
of contiguous bodies was unable to provide the conceptual scaffolding
required to make sense of the principle of inertia and concluded that
absolute space provided the best foundation for the new mathematical
natural philosophy.57 But Kepler seems to have already sensed the diffi-
56
Drake (ed.), Galileo, p. 327. Galileo, while content to grant for the sake of argument that
the cosmos is spherical in shape, makes a point of noting that there is little evidence that the
material universe is finite in extent; ibid. 319 f. And indeed, he appears to have been
genuinely undecided on—and quite likely, not especially interested in—questions concern-
ing the finitude, infinitude, or deployment of the stars; see Koyré, From the Closed World,
pp. 95–9.
57
Newton’s treatment of Descartes’s analysis of motion can be found in De Gravitatione;
see esp. the passage at pp. 19–21 of Janiak (ed.), Newton. Spinoza appears to make a similar
point in corollary 3 to proposition 22 in pt. 2 of The Principles of Cartesian Philosophy.
172 APPENDIX C
culties that lay ahead—in rejecting the notion of an infinite material

universe, he remarks that that notion “carries with it I don’t know what
secret, hidden horror; indeed, one finds oneself wandering in this im-
mensity, to which are denied limits and center and therefore also all
determinate places.”58
58
Quoted at Koyré, From the Closed World, p. 61.
Appendix D
Leibniz and Modal
Relationalism
In so far as Leibniz is willing to grant a sort of reality to the phenomenal

world, he takes space to have a determinate structure (more on this in a
moment). And he emphatically denies that space can profitably be thought
of as composed of geometrically related parts.1 This makes him some sort
of relationalist.
But what sort? Should we view Leibniz as a conservative relationalist,
aiming to explicate claims concerning the structure of space in terms
closely parallel to those offered by substantivalists ( but with material points
playing the role of points of space)? Or is Leibniz a sort of modal relation-
alist, employing a notion of geometric possibility in giving content to
claims about the structure of space?
The orthodox view has it that Leibniz is a modal relationalist. Indeed,
modal relationalism is traditionally introduced and motivated through the
quotation of some suggestive and cryptic remarks that Leibniz makes in his
correspondence with Clarke. For example:
As for my own opinion, I have said more than once that I hold space to be
something purely relative, as time is—that I hold it to be an order of coexistences,
as time is an order of successions. For space denotes, in terms of possibility, an order
of things that exist at the same time, considered as existing together, without
entering into their particular manners of existing.2
Leibniz makes similar remarks elsewhere—e.g., in his correspondence

with Des Bosses:
1
For texts and commentary, see e.g. Lamarra, “Leibniz on Locke on Infinity.”
2
§4 of Leibniz’s third letter to Clarke. Translation of Ariew (ed.), Leibniz and Clarke.
For similar language, see §41 of Leibniz’s fourth letter and §§47, 104, and 106 of his fifth
letter.
174 appendix d
space, like time, is a certain order, namely (in the case of space) that of coexisting,
which includes not only actual things but also possibles. It follows that it is
something indefinite, like every continuum whose parts are not actual but can be
taken at will, just like the parts or fractions of a unity.3
Now, such remarks on their own do not establish that Leibniz is a modal
relationalist. After all, Leibniz, like Descartes, holds that matter forms a
plenum with the structure of Euclidean three-space. So we could under-
stand his remarks about possibilia not as intended to suggest that possibilia
be employed to probe the structure of empty parts of space, but as
intended merely to draw our attention to the fact that there are many
ways that Euclidean space can be filled with matter.4
I believe, however, that a pretty good case can be made for the
orthodox reading of Leibniz as being a modal rather than a conservative
relationalist. My case comes in two parts. In the first I argue that there are
pretty conclusive reasons for denying that Leibniz is a conservative rela-
tionalist. In the second I argue that the texts strongly suggest that he is
indeed a modal relationalist. I conclude by considering some worries one
might have about these arguments.
Leibniz not a Conservative Relationalist

A good place to begin is with Leibniz’s views about void space. On
Descartes’s view, void space is thoroughly impossible: speaking of exten-
sion void of matter involves a conceptual incoherence in much the same
way as would speaking of a chain of mountains without any valleys;
indeed, if God were to annihilate the contents of a full vessel, the result
would be that the walls of the vessel would then be in contact.5 Leibniz
explicitly rejects the Cartesian view:
although I deny that there is any vacuum, I distinguish matter from extension, and
I grant that if there were a vacuum inside a sphere the opposite poles within the
hollow would still not touch. But I believe that divine perfection does not permit
such a situation to occur.6
3
Leibniz to Des Bosses, 31 July 1709. Translation of Look and Rutherford (eds.), Leibniz–
Des Bosses, p. 141.
4
For suggestions along these lines, see Earman, World Enough and Space-Time, §6.12 and
Futch, Leibniz’s Metaphysics of Time and Space, ch. 2.
5
Principles of Philosophy, §ii.18. For discussion and references, see Garber, Descartes’
Metaphysical Physics, pp. 127–55.
6
New Essays on Human Understanding, §ii.xiii.21. Translation of Remnant and Bennett
(eds.), Leibniz.
appendix d 175
In another passage, Leibniz indicates his reasons for denying that the poles
inside a void sphere would touch.
If there were a vacuum in space (for instance, if a sphere were empty inside), one
could establish its size. . . . It follows from this that we can refute someone who says
that if there is a vacuum between two bodies then they touch, since two opposite
poles within an empty sphere cannot touch—geometry forbids it.7
Clearly there is no conceptual incoherence involved in the notion of void

space, since we are able to reason about situations involving vacuum. And
if there is no conceptual incoherence in the notion of the situation
considered, then, one would think, it follows that there exist possible
worlds in which this situation occurs.8
So Leibniz certainly allows the possibility of worlds in which matter
does not (always) fill all of Euclidean space because there are (at least
sometimes) bubbles of void within material extension. In his correspond-
ence with Clarke, he also allows that there are possible worlds in which
material extension is of finite extent: “[a]bsolutely speaking, it appears that
God can make the material universe finite in extension . . .”9
Now, the recognition of the possibility of bubbles of void within matter
and of the possibility of a finite material universe can be consistently
combined with conservative relationalism.10 But anyone who goes in
for such a combination must deny that space is three-dimensional and
Euclidean at every world—e.g., no conservative relationalist could con-
sistently maintain that space was infinite if the material world permanently
had the structure of Aristotle’s spherical cosmos.
However, as Bertrand Russell notes in passing, Leibniz did take the
structure of space and time to be the same in every possible world.11
Consider, for instance, Leibniz’s assertion that “space and time taken
together constitute the order of possibilities of the one entire universe, so
7
Ibid. §ii.xv.11. Translation of Remnant and Bennett (eds.), Leibniz.
8
In fact, is not clear that Leibniz endorses this line of thought (see the discussion of Worry
I below). But note that Leibniz states unambiguously that a vacuum is possible in his letter to
Johann Bernoulli of 13 Jan. 1699; pp. 170 f. in Ariew and Garber (eds.), Leibniz.
9
§30 of Leibniz’s fifth letter to Clarke (see also §73 of the same letter). Translation of
Ariew (ed.), Leibniz and Clarke.
10
Difficulties arise when one considers worlds more exotic than those Leibniz likely had
in mind. See the discussion of §2 of Ch. II above.
11
Russell, A Critical Exposition of the Philosophy of Leibniz, §§71 and 74. The passages that
Russell had in mind are cited in fnn. 3 and 15 of the present discussion. For a contrasting
reading of Leibniz, see Rescher, “The Plurality of Space-Time Frameworks,” esp. §9.
176 appendix d
that these orders—space and time, that is—relate not only to what actually
is but also to anything that could be put in its place . . .”12 In light of the
discussion above, it is natural to read this as telling us that the fact that space
is Euclidean in structure at our world determines what sort of configur-
ations of matter are possible. Thus, there is some possible world whose
matter forms a finite spherical cosmos of Aristotelian type—but no world
whose matter forms a Klein bottle, or any other configuration that could
not be embedded in a Euclidean space of three dimensions. But what
should we think about the structure of space at an Aristotelian world—is
space infinite there (with the Euclidean geometry of our own world
providing the order of possibilities) or is it finite there (with the limited
extent of matter at that world determining that relative to that world only
finite cosmoi are possible)? Leibniz’s stance is, I think, unequivocal. He tells
us that “time and space indicate possibilities beyond any that might be
supposed to be actual. Time and space are of the nature of eternal truths,
which equally concern the possible and the actual.”13 And for Leibniz the
eternal truths are of course genuinely necessary.14 So it would appear that,
for Leibniz, from the fact that the structure of space at our world is
Euclidean, it follows that every world has Euclidean spatial geometry.
And from that it follows in turn that Leibniz was not a conservative
relationalist, since he is committed to taking space to be infinite even at
worlds of Aristotelian structure.
It would seem that the only way to evade this conclusion would be
to show that distinct notions of possibility are in play in the passages in
which Leibniz allows that worlds of finite material extent are possible
and in the passages in which he seems to imply that space and time have a
fixed structure across possible worlds—perhaps the infinitude of matter is
functioning as a tacit presupposition in the latter sort of passage. But this
suggestion will not work. Consider an analogy that Leibniz develops in his
discussion in “On the Ultimate Origination of Things” for the optimiza-
tion problem that God faces in creating a world:
in this context, time, place, or in a word, the receptivity or capacity of the world
can be taken for the cost or the plot of ground on which the most pleasing building
12
From “Reply to the Thoughts on the System of Preestablished Harmony Contained in
the Second Edition of Mr Bayle’s Critical Dictionary, Article Rosarius.” Translation of
Loemker (ed.), Leibniz, p. 583.
13
New Essays, §ii.xiv.26. Translation of Remnant and Bennett (eds.), Leibniz.
14
For discussion and references, see e.g. Adams, Leibniz, §7.1.
appendix d 177
possible is to be built, and the variety of shapes corresponds to the pleasingness of

the building and the number and elegance of the rooms. And the situation is like
that in certain games, in which all places on the board are supposed to be filled in
accordance with certain rules, where at the end, blocked by certain spaces, you will
be forced to leave more places empty than you could have wanted to, unless you
used some trick.15
It seems that God is to consider the possible worlds that result from variant
ways of filling in space and time with matter—with space and time
themselves possessing their structure independently of their material con-
tents. From this it follows that “there would be as much as there possibly
can be, given the capacity of time and space (that is, the capacity of the
order of possible existence); in a word, it is just like tiles laid down so as to
contain as many as possible in a given area.”16 It seems clear here that space
has the same structure at worlds in which matter is sparse as it does at
worlds in which it forms a plenum with the structure of Euclidean space—
and that the infinitude of matter is a consequence, rather than a presup-
position, of the thesis that the structure of space is invariant across worlds.
That puts an end to the interpretation of Leibniz as a conservative
relationalist: there would appear to be no evading the conclusion that he
countenances possible worlds in which space is Euclidean even though the
extent of matter is permanently limited to some fixed finite size.
Leibniz a Modal Relationalist

But what does it mean to say that space is infinite in a world in which
material extension is bounded? The answer of modal relationalists turns
on the notion of geometric possibility—space is infinite if and only if a
linear, unbounded array of material points is geometrically possible. Is this
Leibniz’s answer as well?
Certainly, at various points in the New Essays on Human Understanding it
appears that Leibniz is up to something very like this. After rejecting the
view that space is a substance, Leibniz asserts that space is rather: “a
relationship: an order, not only among existents, but also among possibles
as though they existed.”17 In a nearby passage, Leibniz considers the
temporal analogue of a vacuum and remarks that this “vacuum which
15
Translation of Ariew and Garber (eds.), Leibniz, p. 150.
16
Ibid. 151.
17
§ii.xiii.17. Translation of Remnant and Bennett (eds.), Leibniz.
178 appendix d
can be conceived in time indicates, along with that in space, that time and
space pertain as much to possibles as to existents.”18 Why does the possible
existence of void space indicate that space pertains to merely possible
existents as well as to actual ones? Why think of space as a relationship
among possible existents as if they were actual? Here it seems that Leibniz
has motivations for speaking of possibilia in the same breath as space which
far outstrip the tame observation that Euclidean space may be filled by
matter in many ways. These motivations are most explicit in a passage in
which he is commenting on Locke’s insistence that we should distinguish
extension from material extension:
there is no need to postulate two extensions, one abstract (for space) and the
other concrete (for body). For the concrete one is as it is only by virtue of the
abstract one . . . In fact time and space are only kinds of order; and an empty place
within one of these orders (called ‘vacuum’ in the case of space), if it occurred,
would indicate the mere possibility of the missing item and how it relates to the
actual.19
This seems to suggest that in worlds in which matter does not fill all of
Euclidean space, there is nonetheless some sense in which the complete
pattern of Euclidean space exists and makes possible the pattern of exten-
sion instantiated by matter—and that the gap between the full Euclidean
pattern and the pattern materially instantiated somehow directs us towards
possible ways of filling out material extension so that it would instantiate
the full Euclidean pattern of spatial relations.
There is no knockdown argument here. But it does seem to me that in
these passages we have modal relationalism all but made explicit.
Worries
The discussion above invites a number of worries. I discuss five such: three
that concern the thesis that Leibniz recognizes the possibility of void space;
two that concern the thesis that Leibniz takes the structure of space and
time to be the same at every possible world. In each case, the question is
whether the textual support adduced in favour of these theses above is
18
§ii.xiv.25. Translation of Remnant and Bennett (eds.), Leibniz.
19
§ii.iv.5. Translation of Remnant and Bennett (eds.), Leibniz. The suggestion that the
merely ideal in some sense governs the real can also be found in Leibniz’s letter to Varignon of
2 Feb. 1702; see Loemker (ed.), Leibniz, p. 544.
appendix d 179
significantly undermined by Leibnizean considerations that pull in the

opposite direction.
Worry I. Vacuum Inconsistent with Divine Nature
One of Leibniz’s favourite arguments against the existence of void space
appears to be to establish its impossibility: Leibniz argues that space must be
supposed to be full of matter, since to do otherwise would be to detract
from God’s perfection.20 This argument can be found in many forms
throughout Leibniz’s work.21 The central point is made very succinctly
in section 2 of Leibniz’s second letter to Clarke: “the more matter there is,
the more God has occasion to exercise his wisdom and power.”22 But
since God exists necessarily and has his perfections necessarily, it would
seem that anything implied by these perfections must itself be necessary. So
Leibniz has no more room to recognize the possibility of void space than
do Aristotle and Descartes—and hence Leibniz has no motive at all to
embrace modal relationalism.
But of course this argument really only serves to draw our attention to a
very general problem for Leibniz (and Leibniz scholars). Parallel reasoning
would suggest that there are no contingent truths: God actualizes this
world because it is the best of all possible worlds; but surely, whichever
world is best is necessarily best; and, since it follows from divine perfection,
surely it is necessarily true that God actualizes the best world. Leibniz
himself seems to have been tempted at one time to accept the necessitarian
conclusion, but his considered view appears to have been that it should be
vigourously rejected.23 So in attempting to make out Leibniz’s views, it
seems only fair to allow him to take void space to be genuinely possible
despite the fact that its non-actuality follows from divine nature alone.
Worry II. Space and Matter Inseparable for Leibniz
A second objection to the claim that Leibniz allows void space is based on
the following.
20
Leibniz’s other favourite argument against the void is driven by the principle of
sufficient reason. To the extent that that principle is taken to be necessary, the considerations
engaged below arise for that argument as well.
21
For discussion and references, see Garber, “Leibniz: Physics and Philosophy,” §4.1.
22
Translation of Ariew (ed.), Leibniz and Clarke. A slightly more detailed version of the
argument can be found in the passage customarily printed as a postscript to Leibniz’s fourth
letter to Clarke; ibid. pp. 27 f.
23
For discussion and references see Adams, Leibniz, ch. 1.
180 appendix d
I do not say that matter and space are the same thing. I only say that there is no
space where there is no matter and that space in itself is not an absolute reality.
Space and matter differ as time and motion. However, these things, though
different, are inseparable.24
On one reading of this passage, Leibniz is telling us that space and

material extension are necessarily coextensive—that there is no time at
any world at which there is empty space. If this is right, then we have a
powerful counterweight to the texts in which Leibniz appears to allow
void space.
But there is another natural reading, on which Leibniz is telling us here
merely that there is no time or world which is completely devoid of
matter but at which space exists. That this second reading is to be
preferred is strongly suggested by the context of the passage under
consideration. Leibniz is engaging with Clarke’s off-target gibe that
anyone who takes matter and space to be the same must regard the
material world as necessarily infinite in extent and eternal in duration
(because space and time are). The passage above serves to set up Leibniz’s
assertion that:
it does not follow that matter is eternal and necessary, unless we suppose space to be
eternal and necessary—a supposition ill-grounded in all respects.25
In these passages Leibniz is concerned not with whether there might be

empty space within or outside the material world but with the question of
worlds and times devoid of matter.
(Note that it is important for Leibniz to insist that while God’s existence
is necessary, that of space is not.26 To this end, he asserts that there is a
possible situation in which God creates nothing and (hence) in which
nothing other than God exists—not even space and time.27 Since for
Leibniz space has the same structure in every possible world, there is no
sense in allowing that God could create distinct empty worlds, and so
Leibniz has no use for a distinction between a situation in which God
creates an empty world and one in which no world is created.)
24
§62 of Leibniz’s fifth letter to Clarke. Translation of Ariew (ed.), Leibniz and Clarke.
25
§63 of Leibniz’s fifth letter to Clarke; translation, ibid.
26
Theodicy, §6 of the appendix “Observations on the Book Concerning ‘The Origin of
Evil’ Recently Published in London;” p. 410 in Farrer (ed.), Leibniz.
27
See §41 of his fourth letter to Clarke and §106 of his fifth letter. See also his letter to
Bourget of 2 July, 1716; Robinet (ed.), Correspondance, p. 118.
appendix d 181
Worry III. Void Space Imaginary for Leibniz

Consider next passages in which Leibniz speaks of void space (whether
outside or within the cosmos) as being imaginary.
The same reason which shows that extramundane space is imaginary proves that all
empty space is an imaginary thing, for they differ only as greater and less.28
Since space itself is an ideal thing like time, space out of the world must necessarily
be imaginary, as the schoolmen themselves have acknowledged. The case is the
same with empty space within the world, which I take to be imaginary . . . 29
What does Leibniz intend to communicate by telling us that void space is

imaginary? One possibility that may come to mind is that he is telling us
that void space is not merely non-existent but impossible—for Leibniz
does in fact sometimes employ the term ‘imaginary’ with something like
this force.
The whole difficulty here has therefore only come from a wrong idea of contin-
gency and of freedom, which was thought to have need of a complete indifference
or equipoise, an imaginary thing, of which neither a notion nor an example exists,
nor ever can exist.30
But caution is required here: it is far from obvious that in speaking of void
space as imaginary in his letters to Clarke Leibniz meant to indicate that it
was impossible. (i) As Leibniz indicates in one of the passages under
discussion, ‘imaginary space’ was a Scholastic term of art—one which
had, and was known to have, a dizzying array of established meanings by
the time Leibniz was writing.31 For example, for some prominent authors,
such as Suárez, the distinction between real and imaginary space was the
distinction between space occupied by body and empty space capable of
being occupied by body.32 (ii) Leibniz himself sometimes uses ‘imaginary’
28
§7 of Leibniz’s fourth letter to Clarke. Translation of Ariew (ed.), Leibniz and Clarke.
29
§33 of Leibniz’s fifth letter to Clarke; translation, ibid. See also Leibniz’s letters to
Rémond of 27 March 1716 (Robinet (ed.), Correspondance, pp. 61 f.) and to Des Bosses of 29
May 1716 (Ariew and Garber (eds.), Leibniz, pp. 201–6).
30
Theodicy, §365. Translation of Farrer (ed.), Leibniz, p. 343.
31
See Grant, Much Ado About Nothing, chs. 6 and 7, esp. pp. 120 f.
32
Grant, Much Ado, §7.2. Bayle appears to have taken this for the standard use: see remark
G of the article on Leucippus in the Historical and Critical Dictionary. Gassendi likewise follows
this use—although that does not stop him from going on to speak of imaginary space as a
chimera (see the passages quoted on pp. 110 and 121 f. in Lolordo, Pierre Gassendi and the Birth
of Early Modern Philosophy). Further, Leibniz himself appears to have followed this use in notes
written in 1676 (Parkinson (ed.), De Summa Rerum, p. 77). I mention all of this in order
182 appendix d
to describe entities that are non-actual but possible.33 (iii) Elsewhere

Leibniz speaks of the imagination as an internal sense concerned with
the objects that are the concern of arithmetic and geometry.34
Worry IV. Dimension of Space Varies Across Worlds
Now we turn to a couple of worries concerning the claim that Leibniz
takes space and time to have the same structure in each possible world.
In a famous passage in the New Essays Concerning Human Understanding,
Leibniz discusses the epistemology of spatial and temporal vacua. In the
portion of this passage dealing with the spatial case (quoted above on
p. 175), Leibniz asserts that bubbles of void within material extension are
not only possible but in principle measurable. Things are quite different in
the temporal case: “if there were a vacuum in time, i.e., a duration without
change, it would be impossible to establish its length.”35 Why is this? The
reason appears to be that in the case of a ball-shaped void, we can
determine that opposite points on the boundary do not touch because
“geometry forbids it.” Whereas in the temporal case, “we could not refute
anyone who said that two successive worlds are contiguous in time, with
no possible interval between them. We could not refute him, I say,
because that interval is indeterminable.” This claim raises some interesting
questions.36 But for present purposes, the important point comes in the
final remark that Leibniz makes in this discussion: “If space were only a
line, and if bodies were immobile, it would also be impossible to establish
the length of the vacuum between two bodies.”
to drive home the point that one should not jump to conclusions about the sense of
‘imaginary space.’ It is of course clear that Leibniz himself was not following this use in the
passages cited in fnn. 28 and 29 above—cf. esp. §29 of his fifth letter to Clarke.
33
See “On the Method of Distinguishing Real from Imaginary Phenomena.” In Loemker
(ed.), Leibniz, pp. 363–6.
34
Letter to Queen Sophie Charlotte of Prussia; Ariew and Garber (eds.), Leibniz, pp. 186–92.
For discussion and references concerning this theme, see McRae, “The Theory of Knowledge,”
pp. 178–86. Note that Leibniz also speaks of space as imaginary in his memorandum on
Copernicanism and relativity of motion (see Ariew and Garber (eds.), Leibniz, p. 91). It seems
plausible that this use has its roots in Leibniz’s account of mathematics.
35
All material quoted in this paragraph is from §ii.xv.11 and follows the translation of
Remnant and Bennett (eds.), Leibniz.
36
(1) How should we understand the claim that geometry forbids us from taking the poles
of the empty sphere to touch? (2) If Leibniz is allowing, as he seems to be, that there could be
worlds that differ only as to the empirically inaccessible length of a certain period of
changelessness, how can this be reconciled with the sort of verificationist sentiment that he
gives vent to in his correspondence with Clarke (see esp. §52 of Leibniz’s fifth letter)?
appendix d 183
Applying the same standards as were used above in arguing that Leibniz
is committed to the possibility of void space, it appears that we must
conclude that Leibniz here commits himself to the possibility of worlds
in which space has only a single dimension. And of course if this is
accepted then we must abandon the thesis that Leibniz takes the structure
of space to be the same across possible worlds.
A couple of possible responses suggest themselves. One option is to read
Leibniz’s remarks concerning the situation in which space is a line and
bodies are immobile as referring to a world in which the matter distribu-
tion is a one-dimensional continuum while space, as usual, has the struc-
ture of Euclidean three-space.
Another possibility is to note that Leibniz is perfectly capable in the
course of technical discussions of getting carried away and saying things
that he ought not to—things that are inconsistent with principles that he
holds dear.37 So perhaps here we have a case in which he says something
carelessly which deeper consideration would lead him to retract—the
remark in question is, after all, a fairly casual one.
This suggestion can be bolstered by considering a passage in which
Leibniz very conveniently addresses the question why our world is three-
dimensional. The context is provided by a puzzle raised by Bayle—why
should matter have three dimensions rather than, say, two or four?38
Leibniz notes that Bayle himself appears to expect that the answer should
lie in God’s will. Leibniz denies this and makes much of his denial. On
his view, that matter should have three dimensions follows not from
considerations of what sort of world is best but rather from a “geometrical
and blind necessity.”39
This provides very strong evidence for the verdict that on Leibniz’s
considered view, the dimension of space is invariant across possible worlds.
Unfortunately, the case is not quite conclusive—there are at least two sorts
of worry that one might consider grounds for appeal.
(1) Consider the reason that Leibniz offers for taking matter to be
necessarily three-dimensional:
37
Example: At one point he suggests that if there were only one material point in the
world, its trajectory through space would be a straight line; “Reply to the Thoughts on the
System of Preestablished Harmony Contained in the Second Edition of Mr Bayle’s Critical
Dictionary, Article Rosarius.” See Loemker (ed.), Leibniz, p. 577.
38
Theodicy, §351.
39
Translation of Farrer (ed.), Leibniz, p. 336.
184 appendix d
the ternary number is determined for it not by the reason of the best, but by a
geometrical necessity, because the geometricians have been able to prove that there
are only three straight lines perpendicular to one another which can intersect at one
and the same point.40
This would appear to be baldly question-begging.41 Worse, for present

purposes, it appears at best to head off the possibility that space could have
more than three dimensions while leaving untouched the possibility that
space could be one- or two-dimensional.
(2) In a discussion of the foundations of the calculus, Leibniz remarks
that “even if someone refuses to admit infinite and infinitesimal lines in a
rigourous and metaphysical sense and as real things, he can still use them
with confidence as ideal concepts which shorten his reasoning.”42 Higher-
dimensional spaces have the same status: “we can also conceive of dimen-
sions beyond three, and even of powers whose exponents are not ordinary
numbers—all in order to establish ideas fitting to shorten our reasoning
and founded on realities.” It may appear that Leibniz here comes peril-
ously close to conceding that higher-dimensional spaces are genuinely
conceivable. If this is so, it may be difficult for him to show that the notion
of a world in which matter has more than three dimensions is incoherent,
as would seem to be required in order to show that in creating the world,
God had no choice but to make it three-dimensional.
These last two points certainly raise real worries about whether Leibniz
ought to take matter to be necessarily three-dimensional. But I do not think
that they manage to raise real worries about the claim that he in fact did so,
given his unequivocal pronouncement on the question in his discussion of
Bayle’s question about the dimensionality of space. (In any case, so long as
it is granted that space is Euclidean at every world, Leibniz’s acceptance of
the possibility of a finite material world forces him towards something like
modal relationalism—whether or not he takes the number of spatial
dimensions to vary from world to world.)
Worry V. Extent of Time Varies Across Worlds
A somewhat similar worry can be raised about the extent of time. Above
we considered a passage in which Leibniz noted that space is not eternal if
40
Translation, of Farrer (ed.), Leibniz, p. 335.
41
For discussion, see Russell, Critical Exposition, §11.
42
Letter to Varignon of 2 Feb. 1702. Translation of Loemker (ed.), Leibniz, p. 543.
appendix d 185
the duration of the created world is finite. In a similar vein, Leibniz

remarks: “If there were no creatures, there would be neither time nor
place, and consequently no actual place.”43 These texts suggest but do not
mandate an interpretation of Leibniz as believing that the extent of time in
a given world is equal to the duration of material extension at that world.
Curiously, there is a text in which Leibniz appears to come down on the
other side on this issue. Locke suggests that having fixed the length of
the year in terms of the motion of the Sun, it makes sense to speak of the
durations of various epochs lying prior the creation of the material uni-
verse.44 In his commentary on this passage, Leibniz does not object that
this notion makes no sense, but appears rather to approve of it (aside from a
quibble about biblical dates).45
So it is far from clear that consideration of this issue should move us
away from the claim that Leibniz takes space and time to have the same
structure in every possible world.46 (In any case, a reasonable fallback
position is available, under which space is Euclidean at every world
while the structure of time differs from world to world in virtue of
differing in extent. This would suffice for the present purpose of arguing
the Leibniz should be classified as a modal relationalist rather than as a
conservative one.)
43
§106 of his fifth letter to Clarke. Translation of Ariew (ed.), Leibniz and Clarke.
44
An Essay Concerning Human Understanding, §ii.xiv.24.
45
New Essays, §ii.xiv.24.
46
For further texts and discussion concerning the question of the extent of time for
Leibniz, see Futch, Leibniz’s Metaphysics, ch. 4.
Appendix E
More on Congruence and
Superposability
In Chapter IV we were interested in those metric spaces with the special

feature that specifying the distance relations between the points of a region
suffices to fix all of the qualitative geometric facts about that region. To put
things a bit more precisely: we were interested in those metric spaces X
with the feature that whenever subspaces X1 , X2 X are congruent (i.e.,
there is an isometry ç: X1 ! X2 ) they are also superposable (i.e., there is a
isometry F : X ! X with X2 ¼ F(X1 )). This property of metric spaces
doesn’t have a standard name (and isn’t often discussed). I will call it lability.
Lability is a very strong condition. This appendix aims to give a feeling
for what the condition amounts to, what goes wrong when it fails, and, in
general, what modal relationalists are up against. Its primary focus is the
Riemannian case (but some remarks will be made about more general
cases).
Lability is a close relative to a condition which has received quite a bit of
attention. We call a metric space X fully homogeneous if any isometry
between subspaces can be extended to an isometry of X (i.e., whenever
subspaces X1 , X2 X are related by an isometry ç: X1 ! X2 then there is
an isometry F : X ! X that extends ç, in the sense that F(x) ¼ ç(x) for all
x 2 X1 ). A related notion: for k ¼ 1, 2, . . . , we say that X is k-point
homogeneous if any isometry between k-point subspaces of X can be
extended to an isometry of X.
Clearly, any fully homogeneous space is also labile and k-point homo-
geneous for any k. And for k > m, any k-point homogeneous space is also
m-point homogeneous (of course, one-point homogeneity is just ordinary
homogeneity—the property of there being, for any two points of the
space, an isometry that maps the first to the second).
appendix e 187
It is known that the only fully homogeneous Riemannian manifolds are

the elementary geometries (the Euclidean, hyperbolic, and spherical
geometries).1 So these spaces are also labile. We will see that they are in
fact the only labile Riemannian manifolds.
Our strategy will be to identify several necessary conditions for a
Riemannian manifold to be labile, then to observe that they are jointly
sufficient, since the elementary geometries are the only Riemannian spaces
that satisfy them.
The Riemannian Case

Necessary Conditions: Homogeneity and Completeness. It is imme-
diate that any labile space is homogeneous (any two one-point subspaces
are congruent, and so superposable). And in the Riemannian setting
homogeneity implies completeness. To see this, note that any Riemannian
manifold X is locally compact.2 So for any x 2 X, we can find r > 0 so that
Br(x) is compact. If X is labile and therefore homogeneous, we can choose r
to be independent of x. But since any Cauchy sequence in X is eventually
restricted to a ball of radius r around some point, any such sequence must
converge (since it is eventually confined to a set that is compact and hence
complete).
As with each of the conditions we will encounter, it is worthwhile
thinking about cases in which these fail—and about what the problem of
specifying the superposability type of a one-point set looks like in such cases.
Beginning with any complete Riemannian manifold, one can create a
whole host of incomplete spaces by excising points. So in some sense
incompleteness is the norm.
Example E.1 (The Punctured Plane). Consider the metric space that results if
we remove a single point from the Euclidean plane. For every point x in this
space, there is one special direction, for which there is an upper bound on
how far one can proceed along a straight line in that direction. Two points in
this space are superposable if and only if they agree about the magnitude of
1
See Birkhoff, “Metric Foundations of Geometry. I.”
2
See Remark B.3 of Appendix B above. Recall that a metric space X is locally compact if
sufficiently small open metric balls have compact closures. For path metric spaces, local
compactness is equivalent to the compactness of all closed bounded subsets; see Plaut, “Metric
Spaces of Curvature $ k,” theorem 8.
188 appendix e
this upper bound. Being told that there is a one-particle world with the
spatial geometry of the punctured plane leaves open infinitely many quali-
tatively distinct possibilities—parameterized by how far the occupied point
is from the “missing point of space.” (Of course, if we had begun instead by
removing an asymmetric subset from the plane, then every remaining point
would have played a distinct geometric role.) &
Even among complete Riemannian manifolds, inhomogeneity is the

norm.
Example E.2 (A Crumpled Sphere). Generic Riemannian manifolds are
not homogeneous.3 A helpful example to picture is any manifold that
results from putting gentle bulges or dents in the surface of a sphere in
Euclidean space: the resulting geometry fails to be homogeneous. In
particular, we can give qualitative characterizations of landmarks on the
surface, and in this way distinguish the geometric role played by any point
by mentioning its distance from these landmarks. &
Example E.3 (The Twisted Cylinder). Take the infinite vertical strip of the
Euclidean plane bounded by the lines x ¼ 1 and x ¼ 1 and identify points
on the boundary via (y, 1) $ ( y, 1) (i.e., give the strip a twist about
the x-axis and glue the edges together). The resulting space is a sort of
Möbius band. It contains many closed curves that count as straight lines,
built out of horizontal line segments in the original strip. All such curves are
of length four, except for one special one of length two (corresponding to
the segment of the x-axis in the original strip). Two points in this space are
superposable if and only if they are the same distance from this distinguished
curve. So if we are told that there is a one-particle world with this structure,
we must also be told how far the occupied point lies from the distinguished
curve in order to fix all of the qualitative geometric facts. &
Necessary Condition: Two-Point Homogeneity. Any labile Rie-

mannian manifold is also two-point homogeneous.4 To see this, we let
(X, d) be a labile Riemannian manifold and suppose that we have four
3
Indeed, generic Riemannian manifolds have no non-trivial symmetries. For references
and discussion, see Blair, “Spaces of Metrics and Curvature Functionals,” §§1.1 f.
4
For Riemannian manifolds, two-point homogeneity is equivalent to isotropy (the
condition that for any point and any two unit tangent vectors at that point, there be an
isometry that fixes the point and whose tangent map sends the first of the given tangent
vectors to the second). See Wolf, Spaces of Constant Curvature, lemma 8.12.1.
appendix e 189
points x1 , x2 , y1 , y2 2 X such that d(x1 , x2 ) ¼ d(y1 , y2 ) and then show that

there must be an isometry of X that maps x1 to y1 and x2 to y2 . As we have
just seen, it follows from the lability of X that X is complete. Now, in any
complete Riemannian manifold, any two points are joined by a geodesic
segment that, considered as a metric subspace of the manifold, is isometric
to a segment of the real line of length equal to the distance between the
points.5 Let ª1 be such a segment that connects x1 to x2 and ª2 be such a
segment that connects y1 to y2 , with each ªi including its initial but not its
final point. Then ª1 and ª2 determine congruent subsets of X: so by the
lability of X, there must be an isometry F : X ! X that maps one on to the
other—and given that each segment includes its initial but not its final
point, F must map x1 to y1 and x2 to y2 .
The class of two-point homogeneous Riemannian manifolds is an
exclusive club comprising just the classical geometries (elliptic, Euclidean,
hyperbolic, spherical) in each dimension and various analogues of elliptic
and hyperbolic geometry based on the complex numbers (in even dimen-
sions greater than or equal to four), the quaternions (in dimensions
divisible by four and greater than or equal to eight), and the octonions
(in dimension sixteen only).6
Example E.4 (The Complex Projective Plane). Recall that one way to
construct the ordinary elliptic plane (alias the real projective plane equipped
with a nice metric) has one begin with the real vector space V ¼ R3
equipped with its Euclidean structure, then identify points of the elliptic
plane with one-dimensional (real) vector subspaces of V, equipped with the
metric structure induced by their intersection with the unit sphere (itself
equipped with the metric structure induced by the ambient Euclidean
geometry). The construction of the complex projective plane follows the
same template, with the complex numbers substituted for real numbers.7
5
This is a standard corollary to the Hopf–Rinow theorem; see e.g. §5.8 of Petersen,
Riemannian Geometry. The result continues to hold in the setting of locally compact path
metric spaces; see Burago et al., A Course in Metric Geometry, theorem 2.5.23.
6
See Wolf, Spaces of Constant Curvature, §8.12.
7
The same holds, mutatis mutandis, for higher-dimensional complex analogues of the
elliptic spaces, and for the corresponding quaternionic and octonionic spaces. The same
basic strategy leads to various analogues of the hyperbolic spaces, modelled on the following
construction of the ordinary hyperbolic plane. Begin with the real vector space W ¼ R3
equipped with the structure of Minkowski space, and identify points of the hyperbolic plane
with the one-dimensional (real) vector subspaces of W that point in timelike directions,
equipped with the metric structure induced by their intersection with the unit spacelike
hyperboloid (itself equipped with the metric structure induced by the ambient Minkowski
geometry). For details, see Busemann, The Geometry of Geodesics, §53.
190 appendix e
Introduce an equivalence relation on C3 by taking points (w1 , w2 , w3 ) and

(z1 , z2 , z3 ) to be equivalent if there is a non-zero complex number º such
that wi ¼ º zi and call the resulting space of equivalence classes CP 2 . If we
take C3 to be equipped with its usual inner product, then CP 2 inherits a
metric d given by:
j(x, y)j
cos d([x], [y]) ¼
jxjjyj
(where x, y 2 C3 , [x] and [y] are the corresponding points in CP 2 , and we
use the inner product and associated norm on C3 ). Equipped with this
metric, CP 2 , is a two-point homogeneous Riemannian manifold of four
(real) dimensions.8 &
It is not hard to find spaces that are homogeneous but not two-point
homogeneous.
Example E.5 (The Cylinder). Take the vertical strip of the Euclidean plane
bounded by the lines x ¼ 1 and x ¼ 1 and glue the edges together by
identifying points on the boundary via (y, 1) $ (y, 1). The result is a
cylinder, which we will picture as being embedded in Euclidean three-
space and as having a vertical axis of symmetry. So the horizontal sections
are circles. This space is homogeneous: rotations about the axis and
translations along the axis both count as symmetries, and by composing
these we can map any point to any other point. But it is anisotropic and
hence not two-point homogeneous: let x be any point and y and z both be
separated from x by a unit of distance with x and y lying on a horizontal
line and x and z lying on a vertical line; then no isometry can fix x while
mapping y to z, for any such isometry would have to induce an isometric
mapping from a circle (the geodesic joining x and y) to an infinite line (the
geodesic joining x to z). If we have a two-particle world with cylindrical
spatial geometry, then in order to fix the totality of qualitative geometric
facts about matter we have to be told not only how far apart the particles
are, but also what direction they lie in from one another. &
Necessary Condition: Constant Curvature. Any labile Riemannian

manifold is a space of constant curvature. Before giving the argument, it
will be helpful to recall a few basic facts about Riemannian geometry.
8
See e.g. Busemann, Geometry of Geodesics, p. 380. Note that CP2 arises as the space of
states of the spin degrees of freedom of a spin-one quantum particle.
appendix e 191
Recall first the notion of sectional curvature. Let (M, g) be a Riemannian

manifold (i.e., M is a manifold and g a Riemannian metric tensor). Let x be a
point in M and Õ be a tangent plane at x (i.e., Õ is a two-dimensional linear
subspace of the space of tangent vectors at x). For any tangent vector v at x,
we can consider the geodesic in M that departs from x with velocity given
by v. If we restrict attention to v 2 Õ, then we find that in some sufficiently
small neighbourhood of x, the points lying on such geodesics form a two-
dimensional surface SÕ through x. We can calculate the Gaussian curvature
of SÕ at x: this the real number k ¼ k(x, Õ) such that for small r, the
circumference of any circle of radius r about x in SÕ is given by:
r 3 k
2r þ ...
3
(where the ellipsis indicates higher-order terms in r).9 We call k(x, Õ) the
sectional curvature of Õ. We can define spaces of constant curvature as those
in which k(x, Õ) doesn’t depend on either x or Õ.10 It turns out that any
Riemannian manifold M in which for each x 2 M, k(x, Õ) is independent
of Õ is in fact a space of constant curvature.11
Recall next that if f : M ! M is a smooth map from a manifold to itself
that fixes a given x 2 M, then f induces a linear map f* on the tangent space
at x (think of the way that a rotation that fixes a given point in Euclidean
space nonetheless acts non-trivially on the space of directions at that point).
Since the sectional curvature associated with a tangent plane is definable in
terms of the metric structure of the manifold, if f : M ! M is an isometry that
fixes x 2 M, then f* must map any tangent plane at x to a tangent plane with
the same sectional curvature. Putting this together with the last point of the
preceding paragraph, we see that in order to show that a Riemannian
manifold (M, g) is a space of constant curvature, it suffices to show that for
each point x and each pair of tangent planes Õ1 and Õ2 at x, there is an
isometry f of (M, g) such that: (a) f fixes x; and (b) f* maps Õ1 to Õ2.
Let us now turn to the case of interest. Let (X, d) be a labile Riemannian
manifold and let x be any point in X and Õ1 any tangent plane at x. We can
choose > 0 so that any point x0 within a ball of radius 6 of x is connected
to x by a unique geodesic segment ‘(x, x0 ) that remains within the ball.12
9
See e.g. Bishop and Goldberg, Tensor Analysis on Manifolds, §5.14.
10
This condition is equivalent to being a connected Riemannian locally isometric to one
of the elementary geometries; see e.g. Wolf, Spaces of Constant Curvature, §2.4.
11
See e.g. Wolf, Spaces of Constant Curvature, corollary 2.2.7.
12
It suffices that 6 be less than the injectivity radius of X at x; see e.g. §§5.5 and 5.9 of
Petersen, Riemannian Geometry.
192 appendix e
Claim: we can find y1 , z1 2 X such that d(x, y1 ) ¼ 2, d(x, z1 ) ¼ 3,

and d(y1 , z1 ) ¼ 4; and such that the tangent vectors to ‘(x, y1 ) and
‘(x, z1 ) at x lie in Õ1.
To see this, let v be any unit vector in Õ1 and let y1 be the point that lies 2
units of distance along the geodesic through x with tangent vector v. For
any unit vector w * in Õ1, let z* be the point that lies 3 units of distance
along the geodesic through x with tangent vector w * , and let
g(w * ) ¼ d(y1 , z* ). Note that g(w * ) ¼ 1 if w * ¼ v and that g(w * ) ¼ 5 if
w * ¼ v. So there must be some w 2 Õ linearly independent of v such
that g(w) ¼ 4. So we can take z1 to be the point that lies 3 units of distance
along the geodesic through x with tangent vector w.
Now let Õ2 be any other tangent plane at x. The same argument shows
that we can find y2 , z2 2 X such that d(x, y2 ) ¼ 2, d(x, z2 ) ¼ 3, and
d(y2 , z2 ) ¼ 4; and such that the tangent vectors to ‘(x, y2 ) and ‘(x, z2 ) at
x lie in Õ2. So (x, y1 , z1 ) and (x, y2 , z2 ) are congruent triples of points—so
since X is labile, there must be an isometry F : X ! X that maps {x, y1 , z1 }
to {x, y2 , z2 }. Given the distances involved, it follows that F fixes x and
maps y1 to y2 and z1 to z2 . Further, since the tangent vectors to ‘(x, yi ) and
‘(x, zi ) at x span Õi (i ¼ 1,2) the map F* that F induces on the tangent
space at x maps Õ1 to Õ2. Since x, Õ1, and Õ2 were arbitrary, this suffices
to show that X is a space of constant curvature.
We have of course already seen examples of Riemannian manifolds that
fail to be spaces of constant curvature. For present purposes, it is interesting
to note that being two-point homogeneous and being of constant curva-
ture are logically independent conditions: there exist Riemannian mani-
folds that are both two-point homogeneous and of constant curvature (the
classical geometries); some that are neither (in generic Riemannian geom-
etries, the sectional curvature varies from tangent plane to tangent plane at
a point in a way that varies from point to point); some that are spaces of
constant curvature but not two-point homogeneous (such as the cylinder);
and some that are two-point homogeneous but not spaces of constant
curvature (such as the complex, quaternionic, and octonionic analogues of
projective and hyperbolic spaces).13
Example E.6 (The Complex Projective Plane Once More). The complex
projective plane is homogeneous and two-point homogeneous. So for
13
For the behaviour of the sectional curvatures in the complex projective plane and its ilk,
see e.g. Petersen, Riemannian Geometry, §§3.5.3 and 8.1.1.
appendix e 193
one- and two-point regions, congruence implies superposability. But the

argument above by which we showed that lability implies constant curva-
ture suffices to show that any space that is not a space of constant curvature
is not three-point homogeneous (since it shows us how to find congruent
three-points sets {x, y1 , z1 } and {x, y2 , z2 } that cannot be superposed if the
sectional curvature varies from tangent plane to tangent plane at x).
What sort of additional data, in addition to distances, is required to fix the
superposability type of a three-point set in the complex projective plane?
Generically a single additional number will do (for instance, one of the
angles that would be enclosed by two sides if the three points were joined to
form a geodesic triangle).14 But this is only the beginning of the story: the
superposability type of a generic region of the complex projective plane is
fixed by fixing the superposability type of each four-point region—but not
by fixing the superposability type of each three-point region.15 &
Necessary Condition: Non-Ellipticity. Our necessary conditions for

lability among Riemannian manifolds are not quite jointly sufficient: the
only Riemannian spaces that are complete, homogeneous, two-point
homogeneous, and of constant curvature are the classical geometries
(elliptic, Euclidean, hyperbolic, and spherical) in each dimension. But
one of these is not like the others.
Example E.7 (The Elliptic Plane). Recall once again that the points of the
elliptic plane can be represented by lines through the origin in R3 , with the
distance between points being given by the smaller of the angles that the
corresponding lines make at the origin; a line in the elliptic plane corres-
ponds to a plane through the origin in R3 .
The elliptic plane is both two-point homogeneous and a space of
constant curvature. But it is not a labile, for there exist three-point sets in
the elliptic plane that are congruent but not superposable.16 To see this,
consider, first, the points (0, 1, 1), (1, 0, 1), (1, 1, 0) 2 R3 . These three
14
See Brehm, “The Shape Invariant of Triangles and Trigonometry in Two-Point
Homogeneous Spaces.” The same holds true for the other two-point homogenous spaces
that are not spaces of constant curvature.
15
See Brehm and Et-Taoui, “Congruence Criteria for Finite Subsets of Complex Pro-
jective and Complex Hyperbolic Spaces,” propositions 1 and 2. For results concerning the
quaternionic case, see Brehm and Et-Taoui, “Congruence Criteria for Finite Subsets of
Quaternionic Elliptic and Quaternionic Hyperbolic Spaces.”
16
For this example, see Seidel, “Discrete Non-Euclidean Geometry,” p. 877.
194 appendix e
points are the vertices of an equilateral triangle on the sphere of radius

two centred at the origin in R3 ; they therefore determine three non-
coplanar lines through the origin; the corresponding points in the elliptic
plane are non-collinear and separated from one another by 3. Consider,
next, the points (0, 1, 1), ( 1, 0, 1), (1, 1, 0) 2 R3 . These deter-
mine three coplanar lines through the origin in R3 . So the three corre-
sponding points in the elliptic plane are collinear—and they too have
mutual distance 3 from one another.17 But now we have two congruent
three-point sets that cannot be superposed: isometries map straight lines
to straight lines, and so cannot map a set of non-collinear points to a
collinear set of points (nor vice versa, since the inverse of any isometry is
an isometry).18
Since the elliptic plane is homogeneous and two-point homogeneous,
congruence implies superposability for one- and two-point regions. We
have seen that this fails for three-point regions. What data are required to
fix the superposability type of a region consisting of three points, x, y, and
z in the elliptic plane?
Here is something that will suffice: specify the distance a ¼ d(x, y) and
b ¼ d(x, z) and an angle Ł that can be formed at x by geodesic segments of
length a and b connecting x to y and to z.19 In other words: a form of the
side-angle-side rule for congruence of triangles holds in the elliptic
plane.20 In order to see this, it suffices to note the following. Suppose
17
To picture what is going on here, picture choosing two points x and y on the equator of
the unit sphere separated by an arc of length 3. Now choose a point z on the sphere that is 3
from each of x and y and a point z* that is 2 *
3 from x and 3 from y. Then z lies on the equator
with x and y but z does not. When we pass to the elliptic plane by identifying antipodal points
on the sphere, any distance less than or equal to 2 is unaltered while any greater distance b is
replaced by b; and the resulting points are collinear if and only if the points we started
with on the sphere lie on a great circle. So the points in the elliptic plane corresponding to
x, y, and z have distance 3 from one another and are non-collinear; the points in the elliptic
plane corresponding to x, y, and z* satisfy the same pattern of distance relations but are
collinear. For this example, see e.g. Busemann and Kelly, Projective Geometry and Projective
Metrics, pp. 221 f.
18
Don’t make too much of the role that collinearity plays here: it is also possible to
concoct examples of congruent but non-superpoasble three-point subsets of the elliptic plane,
neither of which is collinear. For examples, see e.g. Gans, An Introduction to Non-Euclidean
Geometry, pp. 240 f.; or Busemann and Kelly, Projective Geometry, p. 222.
19
Note that in the elliptic plane of unit radius, two points x and y are connected by a
unique shortest geodesic segment if d(x, y) < 2, but are connected by two such segments
when d(x, y) ¼ 2.
20
Another form of the side-angle-side rule fails: the data we have specified do not in
general suffice to determine the angles formed if we connect x, y, and z by geodesic segments.
See Gans, Introduction, pp. 239 f.
appendix e 195
that (x1 , y1 , z1 ) and (x2 , y2 , z2 ) are two triples of points that satisfy the given
constraints. Note that for each triple of points, (xi , yi , zi ), there are tangent
vectors vi and wi at xi with lengths a and b (respectively) that make an angle
Ł, and such that one can reach yi or zi by travelling for a unit of time along
the geodesic at xi with tangent vector vi or wi . In order to show that
(x1 , y1 , z1 ) and (x2 , y2 , z2 ) are superposable, it suffices to show that there is
an isometry of the elliptic plane that maps x1 to x2 and whose tangent map
sends v1 to v2 and w1 to w2 . Since the elliptic plane is homogeneous, we
can without loss of generality take x1 ¼ x2 . So our question is whether
given two pairs of tangent vectors at x1 , such that the corresponding
members of each pair have equal length and such that the angles formed
by the members of each pair are equal, there exists an isometry of the
elliptic plane that fixes x1 and whose tangent map sends each element of
the first pair onto the corresponding member of the second pair. At this
point, it is helpful to switch to a different way of thinking of the elliptic
plane: as the upper half of the unit sphere, with antipodal points on the
equator identified. Since the elliptic plane is homogeneous, we can with-
out loss of generality take x1 to be the North Pole. Isometries of the elliptic
plane that fix x1 are given by rotations of the hemisphere about the
North–South axis and by reflections in vertical planes that pass through
the North Pole.21 By employing a map of the first sort, then one of the
second, we can map v1 to v2 and w1 to w2 .
Interestingly, it is almost true that the superposability type of a region in
the elliptic plane is specified by specifying the congruence type of each of
its three-point subregions.22 &
In fact, no elliptic space is labile.23 Specifying the distances between three

points in an elliptic space suffices to determine whether they are collinear
if and only if one of the distances is 2 or the sum of the distances is less
than .24 Further, in any elliptic space there are congruent but non-
superposable regions of cardinality Œ for any cardinal number Œ, 3 < Œ # c.
21
See Busemann and Kelly, Projective Geometry, pp. 214 f.
22
See Blumenthal, Theory and Applications of Distance Geometry, §§81, 92, 93, and 97.
23
For the results mentioned in this paragraph, see Blumenthal, “Congruence and Super-
posability in Elliptic Space” or Theory and Applications, ch. x.
24
So there is a sense in which the problem is a global one: locally, of course, the elliptic
space looks like a spherical space (which is fully homogeneous); for certain ‘large’ configur-
ations, one can exploit the global structure of an elliptic space to find congruent but non-
superposable configurations.
196 appendix e
The Foregoing Conditions are Jointly Sufficient. So far we have

seen that that any labile Riemannian manifold is a non-elliptic two-point
homogeneous space of constant curvature (recall that homogeneity and
completeness are implied by two-point homogeneity). Conversely, since
the only Riemannian manifolds that satisfy these conditions are the elem-
entary geometries (Euclidean, hyperbolic, and spherical) and these
are known to be fully homogeneous (and hence labile), we find that
any non-elliptic two-point homogeneous space of constant curvature is
a labile Riemannian manifold.
Beyond the Riemannian Case

The picture that emerges from the above discussion holds more generally
in the much larger class of locally compact path metric spaces.25
Claim: Let (X, d) be a locally compact path metric space. The

following are equivalent.
(i) X is three-point homogeneous.
(ii) X is labile.
(iii) X is fully homogeneous.
(iv) X is one of the elementary geometries.
To see this, note that independently of the assumption that X is a locally
compact path metric space, we know that (iv) ) (iii) (the result of Birkhoff
cited above), that (iii) ) (ii) (immediate from the definitions), and that (iii)
) (i) (likewise immediate). So we need only satisfy ourselves that (ii) ) (iv)
and that (i) ) (iv). The key here is a result of Jacques Tits and Hsien-Chung
Wang, according to which any complete locally compact two-point homo-
geneous path metric space is a Riemannian manifold.26 Now, the arguments
given above showing that for Riemannian manifolds homogeneity implies
25
How can a path metric space fail to be locally compact? e.g., by having a lot of holes
inconveniently arranged or by being infinite-dimensional. If one deletes from the R2 a
sequence of points that converges to (but does not include) the origin, then the resulting
space is still a path metric space but fails to be locally compact—because no closed metric ball
around the origin is complete. Any Hilbert space h is a path metric space; h is locally
compact if and only if finite-dimensional (Abraham et al., Manifolds, Tensor Analysis, and
Applications, proposition 2.1.11).
26
For discussion and references, see Busemann, Recent Synthetic Differential Geometry, §19;
and Freudenthal, “Lie Groups in the Foundations of Geometry,” §§2.19 ff.
appendix e 197
completeness and that lability implies two-point homogeneity carry over

without modification to the setting of locally compact path metric spaces. So
any labile locally compact path metric space is both two-point homoge-
neous and complete—and thus, by the Tits–Wang result is also a Riemann-
ian manifold. But of course we have already seen that the only labile
Riemannian manifolds are the elementary geometries—so (ii) ) (iv).
Similarly, if (X, d) is a three-point homogeneous locally compact path
metric space, then X is both complete and two-point homogeneous and
hence Riemannian. But then it follows more or less immediately that X is
elementary: we already know that X is two-point homogeneous and the
arguments given above to show that a labile Riemannian manifold must be
of constant curvature and cannot be elliptic in fact also show that a three-
point homogeneous Riemannian manifold must likewise be of constant
curvature and non-elliptic. So (i) ) (iv).
But (at least some of) conditions (i)–(iv) can come apart when we move
beyond locally compact path metric spaces.
There are fully homogeneous spaces other than the elementary geom-
etries: a set X equipped with the discrete metric is fully homogeneous if
and only if X is finite; a subset of the real numbers is fully homogeneous if
and only if it has two or fewer members or is isometric to an additive
subgroup of the real numbers (e.g., the integers, the rational numbers, the
algebraic numbers, the real numbers).27
And there are spaces that are three-point homogeneous without being labile
or fully homogeneous. An infinite-dimensional Hilbert space h is k-point
homogeneous for every finite k but includes infinite sets that are congruent but
not superposable.28 For instance, let S1 be an orthonormal basis for h and
let S2 be the set that results from omitting a single element from S1 . Then S1
and S2 are congruent sets that are not superposable (no isometry of h could
map S1 onto S2 because then its image would not be all of h).
I am unsure whether there are examples of spaces that are labile but not
fully homogeneous.29
27
Ovchinnikov, “Homogeneity Properties of Some ‘1 -Spaces,” theorem 2.2.
28
Birkhoff gives this example, and mentions that the Urysohn space has the same feature;
“Metric Foundations,” §6.
29
Birkhoff asserts that lability is weaker than full homogeneity, but gives no examples;
“Metric Foundations,” §6. I have been unable to find any in the literature, so I leave this as a
challenge to the reader.
References
Abraham, R., J. Marsden, and T. Ratiu, Manifolds, Tensor Analysis, and Applications
(2nd edn.). Springer–Verlag: 1988.
Adams, R., Leibniz: Determinist, Theist, Idealist. Oxford University Press: 1994.
—— “Presumption and the Necessary Existence of God.” Noûs 22: 1988, 19–32.
Algra, K., “Posidonius’ Conception of the Extra-Cosmic Void: The Evidence and
the Arguments.” Mnemosyne 46: 1993, 473–505.
Archbold, J., “A Metric for Plane Affine Geometry Over GF(2n ).” Mathematika 7:
1960, 145–8.
Ariew, R. (ed.), G. W. Leibniz and Samuel Clarke: Correspondence. Hackett: 2000.
—— and D. Garber (eds.), G. W. Leibniz: Philosophical Essays. Hackett: 1989.
Armstrong, D., What is a Law of Nature? Cambridge University Press: 1983.
Bader, R., “Towards a Hyperintensional Theory of Intrinsicality.” Unpublished
manuscript: 2010.
Barnes, J., The Presocratic Philosophers (rev. edn.). Routledge: 1982.
Beebee, H., “The Non-Governing Conception of Laws of Nature.” Philosophy and
Phenomenological Research 61: 2000, 571–94. Reprinted in J. Carroll (ed.),
Readings on Laws of Nature. University of Pittsburgh Press: 2004, 250–76.
Benedetti, R. and C. Petronio, Lectures on Hyperbolic Geometry. Springer–Verlag:
1992.
Berger, M., “Une caractérisation purement métrique des variétés riemanniennes à
courbure constante.” In P. Butzer and F. Fehér (eds.), E. B. Christoffel: The
Influence of His Work on Mathematics and the Physical Sciences. Birkhäuser: 1981,
480–92.
—— Riemannian Geometry During the Second Half of the Twentieth Century. American
Mathematical Society: 2000.
Bessaga, C. and A. Pełczyński, Selected Topics in Infinite-Dimensional Topology.
Pańtstwowe Wydawnictwo Naukowe: 1975.
Bigelow, J., B. Ellis, and C. Lierse, “The World as One of a Kind: Natural
Necessity and Laws of Nature.” British Journal for the Philosophy of Science 43:
1992, 371–88. Reprinted in J. Carroll (ed.), Readings on Laws of Nature.
University of Pittsburgh Press: 2004, 141–60.
Bird, A., Nature’s Metaphysics: Laws and Properties. Oxford University Press: 2007.
Birkhoff, G., “Metric Foundations of Geometry. I.” Transactions of the American
Mathematical Society 55: 1944, 465–92.
Bishop, R. and S. Goldberg, Tensor Analysis on Manifolds. Dover: 1980.
200 references
Blair, D., “Spaces of Metrics and Curvature Functionals.” In F. Dillen and

L. Verstraelen (eds.), Handbook of Differential Geometry, vol. i. Elsevier: 2000, 155–85.
Blumenthal, L., “Congruence and Superposability in Elliptic Space.” Transactions of
the American Mathematical Society 62: 1947, 431–51.
—— “Distance Geometries.” The University of Missouri Studies: A Quarterly of
Research 13: 1938, 1–142.
—— A Modern View of Geometry. Dover: 1980.
—— Theory and Applications of Distance Geometry (2nd edn.). Chelsea: 1970.
—— and K. Menger, Studies in Geometry. W. H. Freeman: 1971.
Bogatyi, S., “Metrically Homogeneous Spaces.” Russian Mathematical Surveys 57:
2002, 221–40.
Bombelli, L. and J. Noldus, “The Moduli Space of Isometry Classes of Globally
Hyperbolic Spacetimes.” Classical and Quantum Gravity 21: 2004, 4429–53.
Bowen, A. and R. Todd (eds.), Cleomedes’ Lectures on Astronomy: A Translation of
The Heavens. University of California Press: 2004.
Brehm, U., “The Shape Invariant of Triangles and Trigonometry in Two-Point
Homogeneous Spaces.” Geometriae Dedicata 33: 1990, 59–76.
—— and B. Et-Taoui, “Congruence Criteria for Finite Subsets of Complex
Projective and Complex Hyperbolic Spaces.” Manuscripta Mathematica 96:
1998, 81–95.
—— —— “Congruence Criteria for Finite Subsets of Quaternionic Elliptic and
Quaternionic Hyperbolic Spaces.” Geometriae Dedicata 84: 2001, 261–9.
Bricker, P., “The Fabric of Space: Intrinsic vs. Extrinsic Distance Relations.” In
P. French, E. Uehling, and H. Wettstein (eds.), Midwest Studies in Philosophy, vol.
xviii. Philosophy of Science. University of Notre Dame Press: 1993, 271–294.
—— “Island Universes and the Analysis of Modality.” In G. Preyer and F. Siebelt
(eds.), Reality and Humean Supervenience: Essays on the Philosophy of David Lewis.
Rowman and Littlefield: 2001, 27–55.
—— “Plenitude of Possible Structures.” Journal of Philosophy 88: 1991, 607–19.
Brighouse, C., “Incongruent Counterparts and Modal Relationalism.” International
Studies in the Philosophy of Science 13: 1999, 53–68.
—— “Spacetime and Holes.” In D. Hull, M. Forbes, and R. Burian (eds.), PSA
1994: Proceedings of the Biennial Meeting of the Philosophy of Science Association, vol.
1. Philosophy of Science Association: 1994, 117–25.
Brown, H., Physical Relativity: Space-time Structure from a Dynamical Perspective.
Oxford University Press: 2005.
Brush, S. (ed.), The Selected Works of Pierre Gassendi. Johnson Reprint Corporation:
1972.
Burago, D., Y. Burago, and S. Ivanov, A Course in Metric Geometry. American
Mathematical Society: 2001.
references 201
Burnyeat, M., “The Sceptic in His Place and Time.” In R. Rorty, J. B. Schnee-
wind, and Q. Skinner (eds.), Philosophy in History. Cambridge University Press:
1984, 225–54. Reprinted in M. Burnyeat and M. Frede (eds.), The Original
Sceptics: A Controversy. Hackett: 1997, 92–126.
Busemann, H., The Geometry of Geodesics. Academic Press: 1955.
—— Recent Synthetic Differential Geometry. Springer–Verlag: 1970.
—— “Timelike Spaces.” Dissertationes Mathematicae 53: 1967, 5–50.
—— and P. Kelly, Projective Geometry and Projective Metrics. Academic Press: 1953.
Butterfield, J., “Against Pointillisme about Geometry.” In F. Stadler and M. Stöltzner
(eds.), Time and History: Proceedings of the 28th. International Ludwig Wittgenstein
Symposium, Kirchberg am Wechsel, Austria 2005. Ontos Verlag: 2006, 181–222.
—— “The Hole Truth.” British Journal for the Philosophy of Science 40: 1989, 1–28.
—— “Relationism and Possible Worlds.” British Journal for the Philosophy of Science
35: 1984, 101–13.
Callender, C., “Answers in Search of a Question: ‘Proofs’ of the Tri-Dimensionality
of Space.” Studies in History and Philosophy of Modern Physics 36: 2005, 113–36.
Carroll, J., Laws of Nature. Cambridge University Press: 1994.
—— “Nailed to Hume’s Cross?” In T. Sider, J. Hawthorne, and D. Zimmerman
(eds.), Contemporary Debates in Metaphysics. Blackwell: 2008.
Cartwright, N., The Dappled World: A Study of the Boundaries of Science. Cambridge
University Press: 1999.
Casati, R. and A. Varzi, Parts and Places: The Structures of Spatial Representation. MIT
Press: 1999.
Cassorla, M., “Approximating Compact Inner Metric Spaces by Surfaces.” Indiana
University Mathematics Journal 41: 1992, 505–13.
Charleton, W., Physiologia Epicuro-Gassendo-Charltoniana. Facsimile edn. Johnson
Reprint Corporation: 1966.
Clifford, W., Lectures and Essays (2nd edn.). MacMillan: 1886.
Cohen, J. and C. Callender, “A Better Best System Account of Lawhood.”
Philosophical Studies 145: 2009, 1–34.
Coxeter, H. S. M., Non-Euclidean Geometry (2nd edn.). University of Toronto
Press: 1947.
—— Projective Geometry. Blaisdell: 1964.
Dafermis, M., “General Relativity and the Einstein Equations.” In T. Gowers,
J. Barrow–Green, and I. Leader (eds.), The Princeton Companion to Mathematics.
Princeton University Press: 2008, 483–93.
Divers, J. and J. Melia, “The Analytic Limit of Genuine Modal Realism.” Mind
111: 2002, 15–36.
Dorr, C., “Finding Ordinary Objects in the World of Quantum Mechanics.”
Unpublished manuscript: 2006.
202 references
Dorr, C.,“Review of Every Thing Must Go: Metaphysics Naturalized by James

Ladyman and Don Ross, with David Spurrett and John Collier.” Notre Dame
Philosophical Reviews: 2010 (online publication).
Drake, S. (ed.), Galileo: Dialogue Concerning the Two Chief World Systems. University
of California Press: 1967.
Dretske, F., “Laws of Nature.” Philosophy of Science 44: 1977, 248–68. Reprinted in
J. Carroll (ed.), Readings on Laws of Nature. University of Pittsburgh Press: 2004,
16–37.
Earman, J., Bangs, Crunches, Whimpers, and Shrieks: Singularities and Acausalities in
Relativistic Spacetimes. Oxford University Press: 1995.
—— Bayes or Bust? A Critical Examination of Bayesian Confirmation Theory. MIT
Press: 1992.
—— “In Defense of Laws: Reflections on Bas van Fraassen’s Laws and Symmetry.”
Philosophy and Phenomenological Research 53: 1993, 413–18.
—— “Laws of Nature: The Empiricist Challenge.” In R. Bogdan (ed.), D. M.
Armstrong. D. Reidel: 1984, 191–223.
—— A Primer on Determinism. Reidel: 1986.
—— World Enough and Space-Time: Absolute versus Relational Theories of Space and
Time. MIT Press: 1989.
—— and J. Roberts, “Contact with the Nomic: A Challenge for Deniers of
Humean Supervenience about Laws of Nature. Part I: Humean Superveni-
ence.” Philosophy and Phenomenological Research 71: 2005, 1–22.
—— —— “Contact with the Nomic: A Challenge for Deniers of Humean
Supervenience about Laws of Nature. Part II: The Epistemological Argument
for Humean Supervenience.” Philosophy and Phenomenological Research 71: 2005,
253–86.
Eichhorn, J., Global Analysis on Open Manifolds. Nova Science Publishers: 2007.
Ellis, B., “Causal Powers and Laws of Nature.” In H. Sankey (ed.), Causation and
Laws of Nature. Kluwer: 1999, 19–34.
—— “Response to David Armstrong.” In H. Sankey (ed.), Causation and Laws of
Nature. Kluwer: 1999, 39–43.
Fales, E., “Are Causal Laws Contingent?” In J. Bacon, K. Campbell, and
L. Reinhardt (eds.), Ontology, Causality, and Mind: Essays in Honour of D. M.
Armstrong. Cambridge University Press: 1993, 122–44.
Farrer, A. (ed.), G. W. Leibniz: Theodicy. Open Court: 1985.
Feferman, S., J. Dawson, S. Kleene, G. Moore, R. Solovay and J. van Heijenoort
(eds.), Kurt Gödel: Collected Works, vol. i. Publications 1929–1936. Oxford Univer-
sity Press: 1986.
Field, H., “Can We Dispense with Space-Time?” In P. Asquith and P. Kitcher
(eds.), PSA 1984: Proceedings of the Biennial Meeting of the Philosophy of Science
Association, vol. 2. Philosophy of Science Association: 1985, 33–90. Reprinted
references 203
(with an additional postcript) in H. Field, Realism, Mathematics and Modality.

Blackwell: 1989, 171–226.
Fine, K., “Essence and Modality.” Philosophical Perspectives 8: 1994, 1–16.
—— “Senses of Essence.” In W. Sinnott-Armstrong, D. Raffman, and N. Asher
(eds.), Modality, Morality, and Belief: Essays in Honor of Ruth Barcan Marcus.
Cambridge University Press: 1995, 53–73.
—— “The Varieties of Necessity.” In T. Gendler and J. Hawthorne (eds.),
Conceivability and Possibility. Oxford University Press: 2002, 253–81. Reprinted
in K. Fine, Modality and Tense: Philosophical Papers. Oxford University Press:
2005, 235–60.
Fisher, J., “Geometry According to Euclid.” American Mathematical Monthly 86:
1979, 260–70.
Folland, G., Real Analysis: Modern Techniques and Their Applications. Wiley–In-
terscience: 1984.
Forrest, P., “Is Space-Time Discrete or Continuous?—An Empirical Question.”
Synthese 103: 1995, 327–54.
Forster, M. and E. Sober, “How to Tell When Simpler, More Unified, or Less Ad
Hoc Theories will Provide More Accurate Predictions.” British Journal for the
Philosophy of Science 45: 1994, 1–35.
Foster, J., “In Defence of Phenomenalistic Idealism.” Philosophy and Phenomeno-
logical Research 54: 1994, 509–29.
—— “Induction, Explanation, and Natural Necessity.” Proceedings of the Aristotel-
ian Society 83: 1982–3, 87–101. Reprinted. in J. Carroll (ed.), Readings on Laws of
Nature. University of Pittsburgh Press: 2004, 98–111.
Fowler, D. and P. Fowler (eds.), Lucretius: On the Nature of the Universe. Oxford
Freudenthal, H., “Lie Groups in the Foundations of Geometry.” Advances in
Mathematics 1: 1964, 145–90.
Friedman, M., Dynamics of Reason. CSLI Publications: 2001.
—— Foundations of Space-Time Theories: Relativistic Physics and Philosophy of Science.
Princeton University Press: 1983.
Furley, D., The Greek Cosmologists, vol. i. The Formation of the Atomic Theory and its
Earliest Critics. Cambridge University Press: 1987.
Futch, M., Leibniz’s Metaphysics of Time and Space. Springer–Verlag: 2008.
Gans, D., An Introduction to Non-Euclidean Geometry. Academic Press: 1973.
Garber, D., Descartes’ Metaphysical Physics. University of Chicago Press: 1992.
—— “Leibniz: Physics and Philosophy.” In N. Jolley (ed.), The Cambridge Companion
to Leibniz. Cambridge University Press: 1995, 270–352.
Gerla, G., “Pointless Geometries.” In F. Buekenhout (ed.), Handbook of Incidence
Geometry. Elsevier: 1995, 1015–31.
Glymour, C., Theory and Evidence. Princeton University Press: 1980.
204 references
Grant, E., “The Condemnation of 1277, God’s Absolute Power, and Physical
Thought in the Late Middle Ages.” Viator 10: 1979, 211–49.
—— “Medieval and Seventeenth-Century Conceptions of Infinite Void Space
Beyond the Cosmos.” Isis 60: 1969, 39–60.
—— “The Medieval Doctrine of Place: Some Fundamental Problems and
Solutions.” In A. Maier and A. Paravincini Bagliani (eds.), Studi sul XIV secolo in
memoria de Anneliese Maier. Edizioni di Storia e Letteratura: 1981, 57–79.
—— Much Ado About Nothing: Theories of Space and Vacuum from the Middle Ages to
the Scientific Revolution. Cambridge University Press: 1981.
—— “Place and Space in Medieval Physical Thought.” In P. Machamer and
R. Turnbull (eds.), Motion and Time, Space and Matter. Ohio State University
Press, 1976, 137–67.
Greenstreet, W. (trans.), Henri Poincaré: Science and Hypothesis. Dover: 1952.
Gromov, M., Metric Structures for Riemannian and Non-Riemannian Spaces. Birkhäuser:
2001.
Grove, K., Riemannian Geometry: A Metric Entrance. University of Aarhus: 1999.
Hahm, D., The Origins of Stoic Cosmology. Ohio State University Press: 1977.
Halpin, J., “Scientific Law: A Perspectival Account.” Erkenntnis 58: 2003, 137–68.
Harman, G., “The Inference to the Best Explanation.” Philosophical Review 74:
1965, 88–95.
Hawthorne, J., “Causal Structuralism.” Philosophical Perspectives 15: 2001, 361–78.
Reprinted in J. Hawthorne, Metaphysical Essays. Oxford University Press: 2006,
211–27.
—— “Quantity in Lewisian Metaphysics.” In J. Hawthorne, Metaphysical Essays.
Oxford University Press: 2006, 229–37.
Heil, J., “Dispositions.” Synthese 144: 2005, 343–56.
Heinonen, J., Geometric Embeddings of Metric Spaces. University of Jyväskylä Depart-
ment of Mathematics and Statistics: 2003.
Henkin, L., P. Suppes, and A. Tarski (eds.), The Axiomatic Method: With Special
Reference to Geometry. North–Holland: 1959.
Higginbotham, J., “Truth and Understanding.” Philosophical Studies 65: 1992, 3–16.
Hodges, W., A Shorter Model Theory. Cambridge University Press: 1997.
Howson, C. and P. Urbach, Scientific Reasoning: The Bayesian Approach (2nd edn.).
Open Court: 1989.
Hubička, J. and J. Nešetřil, “A Finite Presentation of the Rational Urysohn
Space.” Topology and its Applications 155: 2008, 1483–92.
Huggett, N., “The Regularity Account of Relational Spacetime.” Mind 115: 2006,
41–73.
—— True Motion. Unpublished manuscript: 2009.
Hušek, M., “Urysohn Universal Space, its Development and Hausdorff ’s
Approach.” Topology and its Applications 155: 2008, 1493–1501.
references 205
Inwood, B., “Chrysippus on Extension and the Void.” Revue internationale de

philosophie 45: 1991, 245–66.
—— and L. Gerson (eds.), The Epicurus Reader. Hackett: 1994.
Janiak, A. (ed.), Isaac Newton: Philosophical Writings. Cambridge University Press:
2004.
Jauernig, A., “Leibniz on Motion and the Equivalence of Hypotheses.” Leibniz
Review 18: 2008, 1–40.
—— “Must Empiricism be a Stance, and Could it Be One? How to be an
Empiricist and a Philosopher at the Same Time.” In B. Monton (ed.), Images
of Empiricism: Essays on Science and Stances, with a Reply from Bas C. van Fraassen.
Oxford University Press: 2007, 271–318.
Kapovich, M., Hyperbolic Manifolds and Discrete Groups. Birkhäuser: 2001.
Kechris, A., V. Pestov, and S. Todorčević, “Fraïssé Limits, Ramsey Theory, and
Topological Dynamics of Automorphism Groups.” Geometric and Functional
Analysis 15: 2005, 106–89.
Kelly, K., “Ockham’s Razor, Hume’s Problem, Ellsberg’s Paradox, Dilation, and
Optimal Truth Conduciveness.” Unpublished manuscript: 2008.
Klein, F., Elementary Mathematics from an Advanced Standpoint: Geometry. Dover:
1939.
Kment, B., “Counterfactuals and the Analysis of Necessity.” Philosophical Perspec-
tives 20: 2006, 237–302.
Kneale, W., Probability and Induction. Oxford University Press: 1949.
Koyré, A., From the Closed World to the Infinite Universe. The Johns Hopkins
Krantz, D., R. Luce, P. Suppes, and A. Tversky, Foundations of Measurement, vol. i.
Additive and Polynomial Representations. Academic Press: 1971.
Kustaanheimo, P., “A Note on a Finite Approximation of the Euclidean Plane
Geometry.” Societas Scientiarum Fennica. Commentationes Physico–Mathematicae 15:
1950, 1–11.
—— “On the Relation of Congruence in Finite Geometries.” Mathematica
Scandinavica 5: 1957, 197–201.
Lamarra, A., “Leibniz on Locke on Infinity.” In A. Lamarra (ed.), L’infinito in
Leibniz: problemi e terminologia. Edizioni dell’Ateneo: 173–91.
Lang, S., Fundamentals of Differential Geometry. Springer–Verlag: 1999.
Lange, M., Laws and Lawmakers: Science, Metaphysics, and the Laws of Nature. Oxford
Langton, R. and D. Lewis, “Defining ‘Intrinsic’.” Philosophy and Phenomenological
Research 58: 1998, 333–45. Reprinted in D. Lewis, Papers in Metaphysics and
Epistemology. Cambridge University Press: 1999, 116–32.
Leeds, S., “Physical and Metaphysical Necessity.” Pacific Philosophical Quarterly 88:
2007, 458–85.
206 references
Lerner, R. and M. Mahdi (eds.), Medieval Political Philosophy: A Source Book. Cornell
Lewis, D., Counterfactuals. Blackwell: 1973.
—— “Humean Supervenience Debugged.” Mind 103: 1994, 473–90. Reprinted
in D. Lewis, Papers in Metaphysics and Epistemology. Cambridge University Press:
1999, 224–7.
—— “New Work for a Theory of Universals.” Australasian Journal of Philosophy 61:
1983, 343–77. Reprinted in D. Lewis, Papers in Metaphysics and Epistemology.
Cambridge University Press: 1999, 8–55.
—— On the Plurality of Worlds. Blackwell: 1986.
—— Parts of Classes. Blackwell: 1991.
—— Philosophical Papers, vol. ii. Oxford University Press: 1986.
—— “Ramseyan Humility.” In D. Braddon–Mitchell and R. Nola (eds.),
Conceptual Analysis and Philosophical Naturalism. MIT Press: 2009, 203–22.
—— “A Subjectivist’s Guide to Objective Chance.” In R. Jeffrey (ed.), Studies in
Inductive Logic and Probability, vol. ii. University of California Press: 1980,
263–94. Reprinted with additional postscripts in D. Lewis, Philosophical Papers,
vol. ii. Oxford University Press: 1986, 83–132.
—— “Tensing the Copula.” Mind 111: 2002, 1–13.
Lindberg, D., The Beginnings of Western Science. University of Chicago Press: 1992.
Lipton, P., Inference to the Best Explanation (2nd edn.). Routledge: 2004.
—— “Is Explanation a Guide to Inference? A Reply to Wesley C. Salmon.” In
G. Hon and S. Rakover (eds.), Explanation: Theoretical Approaches and Applica-
tions. Kluwer: 2001, 93–120.
Loemker, L. (ed.), Gottfried Wilhelm Leibniz: Philosophical Papers and Letters
(2nd edn.). D. Reidel: 1969.
Loewer, B., “Humean Supervenience.” Philosophical Topics 24: 1996, 101–27.
Reprinted in J. Carroll (ed.), Readings on Laws of Nature. University of Pittsburgh
Press: 2004, 176–206.
—— “Laws and Induction.” Unpublished manuscript: 2000.
—— “Time and Law.” Unpublished manuscript: 2004.
Logothetti, D., “An Interview with H. S. M. Coxeter, the King of Geometry.”
The Two-Year College Mathematics Journal 11: 1980, 2–19.
Lolordo, A., Pierre Gassendi and the Birth of Early Modern Philosophy. Cambridge
Look, B. and D. Rutherford (eds.), The Leibniz–Des Bosses Correspondence. Yale
Lowe, E., The Four-Category Ontology: A Metaphysical Foundation for Natural Science.
Oxford University Press: 2006.
McDaniel, K., “Distance and Discrete Space.” Synthese 155: 2007, 157–62.
references 207
McDaniel, K., “Modal Realism with Overlap.” Australasian Journal of Philosophy 82:
2004, 137–52. Reprinted in F. Jackson and G. Priest (eds.), Lewisian Themes: The
Philosophy of David K. Lewis. Oxford University Press: 2004, 140–55.
McKeon, R. (ed.), The Basic Works of Aristotle. Random House: 1941.
McRae, R., “The Theory of Knowledge.” In N. Jolley (ed.), The Cambridge
Companion to Leibniz. Cambridge University Press: 1995, 176–98.
Maudlin, T., “Buckets of Water and Waves of Space: Why Space-Time is Probably
a Substance.” Philosophy of Science 60: 1993, 183–203.
—— The Metaphysics within Physics. Oxford University Press: 2007.
—— “Time, Topology, and Physical Geometry.” Proceedings of the Aristotelian
Society suppl. vol. 84: 2010, 63–78.
Melia, J., “Weaseling Away the Indispensability Argument.” Mind 109: 2000, 455–79.
Menger, K., “The New Foundations of Hyperbolic Geometry.” In J. Butcher (ed.),
A Spectrum of Mathematics: Essays Presented to H. G. Forder. Auckland University
Press: 1971, 86–97. Reprinted in B. Schwiezer, A. Sklar, K. Sigmund, P. Gruber,
E. Hlawka, L. Reich, and L. Schmetterer (eds.), Karl Menger: Selecta Mathematica,
vol. i. Springer: 2002, 495–506.
Monton, B. and B. van Fraassen, “Constructive Empiricism and Modal Nominalism.”
British Journal for the Philosophy of Science 54: 2003, 405–22.
Morison, B., On Location: Aristotle’s Concept of Place. Oxford University Press: 2002.
Mulder, H. and B. van de Velde-Schlick (eds.), Moritz Schlick: Philosophical Papers,
vol. i. Kluwer: 1979.
Mundy, B., “On Quantitative Relationist Theories.” Philosophy of Science 56: 1989,
582–600.
—— “Space-Time and Isomorphism.” In A. Fine, M. Forbes, and L. Wessels
(eds.), PSA 1990: Proceedings of the Biennial Meeting of the Philosophy of Science
Association, vol. 1. Philosophy of Science Association: 1992, 515–27.
Nerlich, G., The Shape of Space (2nd edn.). Cambridge University Press: 1994.
—— “Space-Time Substantivalism.” In M. Loux and D. Zimmerman (eds.), The
Oxford Handbook of Metaphysics. Oxford University Press: 2003, 281–314.
Nolan, D., “Recombination Unbound.” Philosophical Studies 84: 1996, 239–62.
Ott, W., Causation and Laws of Nature in Early Modern Philosophy. Oxford University
Press: 2009.
Ovchinnikov, S., “Homogeneity Properties of Some ‘1 -Spaces.” Discrete and
Computational Geometry 35: 2006, 301–10.
Pambuccian, V., “Axiomatizations of Hyperbolic Geometry: A Comparison Based
on Language and Quantifier Type Complexity.” Synthese 133: 2002, 331–41.
Erratum: “Correction to ‘Axiomatizations of Hyperbolic Geometry.” Synthese
145: 2005, 497.
—— “Simplicity.” Notre Dame Journal of Formal Logic 29: 1988, 396–411.
Pargetter, R., “Laws and Modal Realism.” Philosophical Studies 46: 1984, 335–47.
208 references
Parkinson, G. (ed.), De Summa Rerum: Metaphysical Papers, 1675–1676. Yale

Parsons, J., “Is Everything a World?” Philosophical Studies 134: 2007, 165–81.
Petersen, P., Riemannian Geometry (2nd edn.). Springer–Verlag: 2006.
Plaut, C., “Metric Spaces of Curvature $ k.” In R. Daverman and R. Sher (eds.),
Handbook of Geometric Topology. Elsevier: 2002, 819–98.
Poincaré, H., “On the Foundations of Geometry.” The Monist 9: 1898, 1–43.
Reprinted in P. Pesic (ed.), Beyond Geometry: Classic Papers from Riemann to
Einstein. Dover: 2007, 117–46.
—— La Science et l’Hypothèse. Flammarion: 1938.
Quine, W., “Natural Kinds.” In W. Quine, Ontological Relativity and Other Essays.
Columbia University Press: 1969, 114–38.
—— “Necessary Truth.” In W. Quine, The Ways of Paradox and Other Essays
(rev. edn). Harvard University Press: 1976, 68–76.
—— “Reply to Parsons.” In L. Hahn and P. Schilpp (eds.), The Philosophy of
W. V. Quine. Open Court: 1986, 396–403.
—— “On Simple Theories of a Complex World.” Synthese 15: 1963, 103–6.
Reprinted in R. Gibson (ed.), Quintessence: Basic Readings from the Philosophy of
W. V. Quine. Harvard University Press: 2004, 210–13.
—— “On What There Is.” In W. Quine, From a Logical Point of View (2nd edn.).
Harvard University Press: 1953, 1–19. Reprinted in R. Gibson (ed.), Quintes-
sence: Basic Readings from the Philosophy of W. V. Quine. Harvard University Press:
2004, 177–92.
—— “Posits and Reality.” In W. Quine, The Ways of Paradox and Other Essays
(rev. edn.). Harvard University Press: 1976, 246–54.
—— “The Scope and Language of Science.” British Journal for the Philosophy of
Science 8: 1957, 1–17. Reprinted in R. Gibson (ed.), Quintessence: Basic Readings
from the Philosophy of W. V. Quine. Harvard University Press: 2004, 193–209.
—— “Things and Their Place in Theories,” In Quine, Theories and Things.
Harvard University Press: 1981, 1–23. Reprinted in R. Gibson (ed.), Quintes-
sence: Basic Readings from the Philosophy of W. V. Quine. Harvard University Press:
2004, 229–48.
—— Word and Object. MIT Press: 1960.
Reichenbach, H., Experience and Prediction. University of Chicago Press: 1938.
—— The Philosophy of Space and Time. Dover: 1957.
Remnant, P. and J. Bennett (eds.), Leibniz: New Essays on Human Understanding.
Cambridge University Press: 1996.
Rescher, N., “Leibniz and The Plurality of Space-Time Frameworks.” In
N. Rescher, Leibniz’s Metaphysics of Nature: A Group of Essays. Reidel: 1981,
84–100.
Roberts, J., The Law-Governed Universe. Oxford University Press: 2009.
references 209
Roberts, J., “ ‘Laws of Nature’ as an Indexical Term: A Reinterpretation of Lewis’s

Best-System Analysis.” Philosophy of Science 66: 1999, S502–S511.
—— “Reply to Skow.” Philosophy and Phenomenological Research 75: 2007, 163–7.
Roberts, S., King of Infinite Space: Donald Coxeter, the Man Who Saved Geometry.
Walker & Company: 2006.
Robinet, A. (ed.), Correspondance Leibniz–Clarke. Presses universitaires de France:
1957.
Robinson, R., “Binary Relations as Primitive Notions in Elementary Geometry.”
In L. Henkin, P. Suppes, and A. Tarski (eds.), The Axiomatic Method: With Special
Reference to Geometry. North–Holland: 1959, 68–85.
Roush, S., Tracking Truth: Knowledge, Evidence, and Science. Oxford University
Press: 2005.
Royden, H., “Remarks on Primitive Notions for Elementary Euclidean and Non-
Euclidean Plane Geometry.” In L. Henkin, P. Suppes, and A. Tarski (eds.), The
Axiomatic Method: With Special Reference to Geometry. North–Holland: 1959,
86–96.
Russell, B., A Critical Exposition of the Philosophy of Leibniz (2nd edn.). Routledge:
1992.
Salmon, W., Reality and Rationality. Oxford University Press: 2005.
Sambursky, S., The Concept of Place in Late Neoplatonism. The Israel Academy of
Sciences and Humanities: 1982.
—— Physics of the Stoics. Macmillan: 1959.
Schaffer, J., “Causation and Laws of Nature: Reductionism.” In T. Sider,
J. Hawthorne, and D. Zimmerman (eds.) Contemporary Debates in Metaphysics.
Blackwell: 2008, 82–107.
Sedley, D., “Two Conceptions of Vacuum.” Phronesis 27: 1982, 175–93.
Seidel, J., “Discrete Non-Euclidean Geometry.” In F. Buekenhout (ed.), Handbook
of Incidence Geometry. Elsevier: 1995, 843–920.
Sellars, W., “Concepts as Involving Laws and Inconceivable Without Them.”
Philosophy of Science 15: 1948, 287–315.
Shoemaker, S., “Causal and Metaphysical Necessity.” Pacific Philosophical Quarterly
79: 1998, 59–77. Reprinted in S. Shoemaker, Identity, Cause, and Mind
(expanded edn.). Oxford University Press: 2003, 407–26.
—— “Causality and Properties.” In P. van Inwagen (ed.), Time and Cause. D. Reidel:
1980, 109–35. Reprinted. in S. Shoemaker, Identity, Cause, and Mind (expanded
edn.). Oxford University Press: 2003, 206–33.
Sidelle, A., “On the Metaphysical Contingency of Laws of Nature.” In T. Gendler
and J. Hawthorne (eds.), Conceivability and Possibility. Oxford University Press:
2002, 309–36.
Skow, B., “Are Shapes Intrinsic?” Philosophical Studies 133: 2007, 111–30.
210 references
Skow, B., “Earman and Roberts on Empiricism about Laws.” Philosophy and
Phenomenological Research 75: 2007, 158–62.
—— Once Upon a Spacetime. Unpublished NYU doctoral dissertation: 2005.
Soames, S., “Truth, Meaning, and Understanding.” Philosophical Studies 65: 1992,
17–35.
Sober, E., “Parsimony and Predictive Equivalence.” Erkenntnis 44: 1996, 167–97.
—— Reconstructing the Past: Parsimony, Evolution, and Inference. MIT Press: 1988.
Sorabji, R., Matter, Space and Motion: Theories in Antiquity and Their Sequel.
Duckworth: 1988.
Sorenson, R., “Nothingness.” Stanford Encyclopedia of Philosophy: 2009 (online
publication).
Sormani, C., “How Riemannian Manifolds Converge: A Survey.” Unpublished
manuscript, 2010.
Stein, H., “Newtonian Space-Time.” Texas Philosophical Quarterly 10: 1967. 174–200.
Reprinted in R. Palter (ed.), The Annus Mirabilis of Isaac Newton. MIT Press: 1970,
253–84.
—— “Some Philosophical Prehistory of General Relativity.” In J. Earman,
C. Glymour, and J. Stachel (eds.), Foundations of Space-Time Theories. University
of Minnesota Press: 1977, 3–49.
Swoyer, C., “The Nature of Natural Laws.” Australasian Journal of Philosophy 60:
1982, 203–23.
Tarski, A., “What is Elementary Geometry?” In L. Henkin, P. Suppes, and
A. Tarski (eds.), The Axiomatic Method: With Special Reference to Geometry.
North–Holland: 1959, 16–29.
Todd, R., “Cleomedes and the Stoic Concept of the Void.” Apeiron 16: 1982,
129–36.
Tooley, M., “A Defense of Absolute Simultaneity.” In W. Craig and Q. Smith
(eds.), Einstein, Relativity, and Absolute Simultaneity. Routledge: 2008, 229–43.
—— “The Nature of Laws.” Canadian Journal of Philosophy 7: 1977, 667–98.
Reprinted in J. Carroll (ed.), Readings on Laws of Nature. University of Pittsburgh
Press: 2004, 38–70.
Torretti, R., Philosophy of Geometry from Riemann to Poincaré. D. Reidel: 1984.
Tweedale, M., “Armstrong on Determinable and Substantial Universals.” In
R. Bogdan (ed.), D. M. Armstrong. D. Reidel: 1984, 171–89.
Uspenskij, V., “The Urysohn Universal Metric Space is Homeomorphic to a
Hilbert Space.” Topology and its Applications 139: 2004, 145–9.
Vallentyne, P., “Explicating Lawhood.” Philosophy of Science 55: 1988, 598–613.
—— “Intrinsic Properties Defined.” Philosophical Studies 88: 1988, 209–19.
van Fraassen, B., Introduction to the Philosophy of Time and Space (2nd edn.). Columbia
—— Laws and Symmetry. Oxford University Press: 1989.
references 211
van Fraassen, B., “A Re-Examination of Aristotle’s Philosophy of Science.”

Dialogue 19: 1980, 20–45.
—— The Scientific Image. Oxford University Press: 1981.
Vershik, A., “Random Metric Spaces and Universality.” Russian Mathematical
Surveys 59: 259–95.
Weatherson, B., “Intrinsic vs. Extrinsic Properties.” Stanford Encyclopedia of
Philosophy: 2006 (online publication).
Westfall, R., “The Foundations of Newton’s Philosophy of Nature”, British Journal
for the History of Science 1: 1962, 171–82.
Weyl, H., Philosophy of Mathematics and Natural Science. Princeton University Press:
1949.
Wolf, J., Spaces of Constant Curvature (5th edn.). Publish or Perish: 1984.
Woodward, J., Making Things Happen: A Theory of Causal Explanation. Oxford
Yablo, S., “Intrinsicness.” Philosophical Topics 26: 1999, 479–505.
Index
Abraham, R. 22 n39, 22 n41, 196 n25 best-system approaches:

absolute-relational debate, history 157–72 generally 54–77
accessibility relations 84, 97 and super-grounding 121–32
accompaniment 114–15 betweenness 10, 11, 32, 97
Achilles the Grammarian 165 n37 Bigelow, J. 106 n13, 127 n61
Adams, R. 9 n5, 176 n14, 179 n23 Bird, A. 9 n3, 106 n13, 107 n16,
admissible metric spaces 84 108 n18, 122 n45, 127 n60
Algra, K. 164 n34 Birkhoff, G. 187 n1, 196, 197 n28,
ambition: 197 n29
generally 52–3, 102, 135 Bishop, R. 191 n9
and necessitarianism 119 Blumenthal, L. 12 n12, 13 n13, 15 n16,
and primitivist approaches 80–1, 85 15 n18, 32 n62, 58 n10, 65 n27,
anti-realism 1–4, 123, 125–7 70 n36, 131 n63, 195 n22
Archbold, J. 30 n57 Bogatyi, S. 49 n29
Archytas 162–3, 166, 169 Bolyai, J. 15
Ariew, R. 175 n8, 175 n9, 177 n15, Bombelli, L. 150 n1
179 n22, 180 n24, 181 n28, bounded spaces 150–6, 177–8
181 n29, 182 n34, 185 n43 bouquets, metric 23
Aristotle 2, 18, 38, 39, 125 n54, 158–62, Bowen, A. 165 n35, 165 n37
166 Brehm, U. 132 n65, 193 n14, 193 n15
Armstrong, D. 79 n1, 112 n28 Bricker, P. 9 n2, 9 n5, 13 n13, 25 n46,
Armstrong–Dretske–Tooley approach 25 n48, 26 n50, 26 n52, 30 n59,
80 n3, 106, 109, 110, 112 101 n28
astronomy 169 Brighouse, C. 50 n31, 82 n7, 100 n26
atomists 158–9, 162–4 Brush, S. 169 n51, 169 n52
attraction 103–4 bubbles of void 175, 182
Axes of Evil 23 Burago, D. 17 n23, 21 n36, 23 n43,
axiomatization: 75 n52, 150 n1, 152 n5, 152 n7,
and Euclidean geometry 10–11, 12, 152 n9, 154 n11, 154 n12,
32, 65 156 n20, 189 n5
of scientific theories 56 Burnyeat, M. 162
axiom of free mobility 16, 186–97 Busemanm, H. 138 n67, 189 n7,
190 n8, 194 n17, 194 n18,
Bader, R. 115 n33 195 n21, 196 n26
balance, and best-system approaches to Butterfield, J. 26 n50, 50 n31, 82 n7,
geometry 67–72 100 n26
Barnes, J. 164 n30
Bayle, P. 1 n1, 183 Callender, C. 15 n17, 72 n41
Beebee, H. 60 n17, 75 n49 Carroll, J. 53, 56 n6, 60 n16,
Benedetti, R. 19 n30 75 n48, 79 n2, 80 n4,
Bennett, J. 174 n6, 175 n7, 176 n13, 100 n25, 122 n46
177 n17, 178 n18, 178 n19, 182 n35 Cartwright, N. 122 n45
Berger, M. 21 n37, 36 n3 Casati, R. 101 n27
Bessaga, C. 47 n21 Cassorla, M. 152 n9
214 index
Cauchy sequences 12, 14, 28, 31, 152, 153 modal relationalism 52
Cayley–Menger determinants 12, 129 necessitarianism 117–21
Charleton, W. 161, 163 n27, 169, non-quantitative relations 95–6
169 n49, 169 n51, 169 n52 substantivalism 41, 84, 85
Chrysippus 166 n38 summary 137
classical geometries: Divers, J. 30 n59
and distance 15–16 Dorr, C. 92 n16, 100 n24, 140 n2
and simplicity 64–6, 70 Drake, S. 171 n56
Cleomedes 165, 166 Dretske, F. 79 n1, 112 n28;
Clifford, W. 65 n28 see also Armstrong–Dretske–Tooley
Cohen, J. 72 n41 approach
collinearity 36, 41, 90, 95–7, 131 duplicate objects 51, 115–21
compact metric spaces 151–5, 196–7
compatibility properties 114–17, Earman, J. 18 n24, 51 n32, 60 n17,
119–21, 129, 132 61 n18, 75 n49, 75 n51, 110 n24,
complete metric spaces 14, 47, 153, 155 112 n28, 112 n29, 122 n46,
complex projective planes 189–90, 123 n47, 144 n11, 147 n16,
192–3 157 n1, 174 n4
Condemnation of 1277 167–8 edges, defined 23–4
congruence 10, 11, 32, 87–91, 97, 99, Eichhorn, J. 22 n40, 155 n16
186–97 Einstein, A. 35
conservative relationalism 37–49, 50, elementary geometries 16, 196
52 n35, 174–7 elliptic geometry:
constant curvature, spaces of 18–19, generally 16
88–9, 190–2 and balance 69–71, 73–4
continuum, space as 40 best-system approaches 65
convergence 12, 14, 23, 24, 150–6 and constant curvature 19
Copernicus 170 and dimension 36 n4
countable metric spaces 46–9 elliptic planes 193–6
Coxeter, H. S. M. 65, 66 and necessitarianism 131
crumpled spheres 88, 188 and primitivist approaches 90
cumulative limit 44 Ellis, B. 53, 106 n13, 107 n16,
curves: 110 n23
spaces of constant curvature 18–19, embeddability 42–4, 45, 47, 75, 131
88–9, 190–2 empty space 38 n8; see also void space
spaces of vanishing curvature 19 empty worlds 82, 100–1, 120 n42
cylinders 17, 89, 92, 120, 190; Epicurus 163 n25, 163 n27, 164 n32
see also twisted cylinders essential properties 126–7
Et-Taoui, B. 132, 193 n15
Dafermis, M. 63 n21 Euclidean geometry:
Descartes, R. 2, 38, 157–8, 169, generally 8–9, 12, 15
171, 174 and balance 69–70, 73–4
dimension 36; see also uni-dimensionality and constant curvature 18–19
discrete metric spaces 24, 69 and dimension 36 n4
distance relations: and distance 10–14
best-system approaches 57–8 generalizations of Euclidean
compatibility properties 131 geometry 10–30
distance and Euclidean and necessitarianism 119–21
geometry 10–14 and simplicity 64–5, 66
distances from discrete fields 29–30 and substantivalism 82
inhomogeneous spaces 88–9 vs taxi-cab metrics 23
metric spaces 31–4 extrinsic properties 115–16
index 215
Fales, E. 106 n14 grounded modal relationalism 80–1

Farrer, A. 180 n26, 181 n30, 183 n39 groundedness 51, 52–3, 54 n2, 102, 119,
Feferman, S. 75 n50 135; see also super-grounding
Field, H. 33 n67, 51 n32, 95 n19, Grove, K. 21 n36
131 n64
fields, finite 29–30 Hahm, D. 159 n9, 164 n34, 165 n35,
Fine, K. 9 n3, 75 n48, 79 n2, 80 n4, 165 n37, 166 n38
108 n17, 110 n24, 126 half-open intervals 45, 94
finite metric spaces 24, 33 Halpin, J. 72 n42
finitude 3–4, 36, 39–41; Harman, G. 145 n12
see also infinitude Hawthorne, J. 30 n59, 105 n10,
Fisher, J. 30 n57 106 n13, 110 n22
flat spaces 16–17, 19, 20 n35, 50 Heil, J. 122 n45
Folland, G. 21 n38, 22 n41 Heinonen, J. 46 n18, 47 n21
Forrest, P. 13 n13, 25 n45, 26 n51, Hempel, C. G. 139
26 n54, 155 n15 Henkin, L. 10 n9
Foster, J. 2 n3, 112 n28 Higginbotham, J. 37 n5
Fowler, D. 163 n26, 163 n29, 164 n33 Hilbert spaces and manifolds 21–2,
Fowler, P. 163 n26, 163 n29, 164 n33 47 n19, 196 n25, 197
Fraïssé limits 48–9 hockey, ice 62
Freudenthal, H. 196 n26 Hodges, W. 48 n25
Friedman, M. 10 n8, 63 n22 homogeneity 88, 143–5, 186, 189–90,
fullness (full worlds) 84, 85–6, 92–3, 99, 192, 196–7
117–21 Hopf–Rinow theorem 189 n5
fundamental properties 55, 103–14 Howson, C. 144 n9
Furley, D. 164 n31 Hubička, J. 47 n23
Futch, M. 174 n4, 185 n46 Huggett, N. 54, 57–60, 69, 73 n44,
75 n49
Galileo 170, 171 Hušek, M. 47 n23
Gans, D. 194 n18, 194 n20 hyperbolic geometry 15, 66
gappiness of metric spaces 30–1 and constant curvature 19
Garber, D. 174 n5, 175 n8, 177 n15, and dimension 36 n4
179 n21, 181 n29, 182 n34 hypothetico-deductive method 143
Gassendi, P. 161 n18, 163 n25, 168–9,
181 n32 idealism 72, 73–4
Gauss, C.F. 15 imaginary space 181–2
Gerla, G. 28 n55 indefiniteness 36, 40–1
Gerson, L. 163 n29, 164 n32 inertia, principle of 170
Glymour, C. 144 n10, 144 n11 infinitude:
Gödel, K. 75 n50 best-system approaches 58
Goldberg, S. 191 n9 Leibniz on 176
Grant, E. 18 n24, 159 n11, 160 n13, relationalism 36, 40–1, 49, 50
161 n16, 161 n17, 162 n23, of the void 165
163 n25, 163 n28, 166–7, 168 n46, see also finitude
169 n48, 169 n50, 181 n31, 181 n32 inhomogeneous spaces 88–9
graphs, metric 23–4, 31 intervals, open and closed 16 n20, 22,
gravity, theory of 62–3 45, 94
Greenstreet, W. 64 n25 intrinsic properties 51, 103 n2, 114–17
Gromov–Hausdorff limits 44, 45, intuitions 74–5, 87
150–6 Inwood, B. 163 n29, 164 n32, 164 n34
Gromov, M. 19 n31, 44 n13, 152 n6,
155 n16, 156 n21 Jauernig, A. 140 n2, 157 n2
216 index
Kant, I. 15 n17, 17, 72 Mahdi, M. 167 n45

Kapovich, M. 44 n13 material geometry 3, 38–9
Kechris, A. 49 n28 matter, location of 131–2, 163
Kelly, K. 144 n9, 147 n16, 194 n17, Maudlin, T. 9 n4, 13 n13, 26 n51,
194 n18, 195 n21 26 n53, 26 n54, 53, 75 n48, 79 n2,
Kepler 170, 171–2 80 n4, 110 n24
Klein bottles 17 McDaniel, K. 114 n31, 123 n48,
Klein, F. 65–6 154 n14
Kment, B. 127 n58 McKeon, R. 158 n3, 158 n4, 159 n7,
Kneale, W. 106 n15, 112 n28 159 n8, 159 n10, 159 n12, 161 n19
Koyré, A. 163 n25, 168 n47, 170 n54, McRae, R. 182 n34
170 n55, 171 n56, 172 n58 mediation, of distance relations 26–7
k-point homogeneity 186 Melia, J. 30 n59, 32 n63, 33 n67
Krantz, D. 33 n65 Menger, K. 65, 66
Kripke, S. 78, 111 mereology 101
Kustaanheimo, P. 30 n57 metric spaces:
generally 13–34
lability 186–97 admissible metric spaces 84
Lamarra, A. 173 n1 compact metric spaces 151–5, 196–7
Lange, M. 75 n48, 79 n2, 80 n4 complete metric spaces 47, 153, 155
Lang, S. 22 n39, 151 n3 countable metric spaces 46–9
Langton, R. 114 n32, 115 n33, 126 n55 discrete metric spaces 24, 69
laws, necessitarianism about 103–14, finite metric spaces 24, 33
116–17 gappiness of metric spaces 30–1
Leeds, S. 9 n4 non-path metric spaces 24–30
Leibniz, G. 2, 3–4, 49 n30, 157, 169, path metric spaces 22–4, 187 n2
169 n53, 173–85 pointed metric spaces 155–6
Lerner, M. 167 n45 pointless metric spaces 28–9
Lewis, D. 9, 25 n49, 30, 53, 55–7, 61, random metric spaces 47
67, 72 n41, 73, 73 n43, 73 n45, separable metric spaces 46–7
73 n46, 80 n3, 101 n27, 101 n28, uncountable metric spaces 46–7
103 n2, 104–6, 114 n32, 115, 124, universal metric spaces 46–7
126 n55 minimality 43–4
Lindberg, D. 168 n46 Minkowski spaces 189 n7
Lipton, P. 145 n12 Mirror Principle 81–2, 92
Lobachevski, N. I. 15, 65 modal relationalism:
Locke, J. 185 generally 37–8, 49–53, 134–8
Loemker, L. 169 n53, 176 n12, and Leibniz 173–85
178 n19, 182 n33, 183 n37, and necessitarianism 121–32
184 n42 and primitivist approaches 84–90
Loewer, B. 60 n16, 60 n17, 80 n4, Monton, B. 123 n51
100 n24, 108 n17, 112 n28, More, T. 163 n25
123 n47, 127 n60 Morison, B. 158 n5, 162 n22
Logothetti, D. 66 n30 motion 136, 160–2, 164, 165–6, 169–72
Lolordo, A. 169 n48, 169 n51, Mulder, H. 64 n24
181 n32 Mundy, B. 11 n10, 32 n63, 32 n64,
loneliness 114–15 33 n67
Look, B. 174 n3
Lorentzian spacetimes 138 Nash embedding theorem 20 n33
Lowe, E. 108 n17 nature, laws of 103–14, 125 n54
Lucretius 18 n24, 162–4 necessitarian approaches 102–33, 135–6
index 217
negative curvature, spaces of constant 19 properties:

Nerlich, G. 26 n50, 26 n53, 26 n54, 29 compatibility properties 114–17,
Nešetřil, J. 47 n23 119–21, 129, 131, 132
Newton, I. 2, 62–3, 79, 103–4, 105, essential properties 126–7
107–8, 136, 157–8, 161 n18, extrinsic properties 115–16
163 n27, 169, 171–2 fundamental properties 55, 103–14
Nolan, D. 9 n7 intrinsic properties 51, 103 n2, 114–17
Noldus, J. 150 n1 Ptolemy 170, 171
non-metricity 95–8 punctured planes 89, 94, 120, 187–8
null individual hypothesis 101 purity 32, 33
one-particle worlds 5, 52–3, 79, 82, 102, Quine, W. V. 102, 123 n49, 123 n51,
119–21 125 n54, 139, 141
ontology 2, 139–49
Ott, W. 106 n15 random metric spaces 47
Ovchinnikov, S. 197 n27 ratios, of distance 11 n11
realism, defined 1–4
Pambuccian, V. 61 n20, 65 n27 recombination, principle of 104, 115,
paradoxes of Aristotelian motion 117, 122 n44, 124–5
160–1, 169 regions:
Pargetter, R. 80 n4 in metric spaces 28–9
parsimony 45–6 and necessitarianism 117–21
Parsons, J. 114 n31 primitivist approaches 84, 87
particle physics: Reichenbach, H. 147 n14
general principles 68, 74 relationalism:
one-particle worlds 5, 52–3, 79, 82, and Aristotle’s cosmos 18
102, 119–21 and necessitarianism 117–21
path metric spaces 22–4, 187 n2 and primitivist approaches 80–1
paths, defined 16 and Riemannian manifolds 20
Pełczynski, A. 47 n21 and spatial structures 35–53
Petersen, P. 21 n36, 23 n42, 152 n6, summary 134
154 n11, 154 n12, 155 n16, vs substantivalism 1–4
191 n11, 192 n13 see also modal relationalism
Petronio, C. 19 n30 reliability, of methods 146–8
planes: Remnant, P. 174 n6, 175 n7, 176 n13,
complex projective planes 189–90, 177 n17, 178 n18, 178 n19, 182 n35
192–3 Riemannian manifolds:
elliptic geometry 193–6 generally 17 n23, 19–24, 191
flat spaces 17 convergence 152–6
punctured planes 89, 94, 120, 187–8 and gappiness 31
taxi-cab metrics 23 and homogeneous spaces 89–90
Plaut, C. 156 n21, 187 n2 and Huggett 58 n12
plenitude, principle of 30–1, 153 and lability 187–96
Poincaré, H. 64–5, 66, 70 Riemannian metric tensors 21
pointed metric spaces 155–6 separability 47
pointless metric spaces 28–9 and total boundedness 151 n3
Pooley, O. 82 n7 Roberts, J. 60 n17, 72 n42, 75 n49,
Posidonius 164 n34 108 n17, 110 n24, 112 n28,
positive constant curvature, spaces of 18 112 n29, 122 n46, 123 n47, 144 n11
primitivist approaches 78–101, 135–6 Roberts, S. 65 n28
projective geometry 11, 66 Robinet, A. 180 n27
218 index
Robinson, R. 32 n62 and spatial structures generally 35–7

Russell, B. 175, 184 n41 summary 134–8
Rutherford, D. 174 n3 super-congruence 99
super-grounding:
Salmon, W. 142 n5 and modal relationalism 52
Sambursky, S. 164 n34, 167 n41 and necessitarianism 121–32
scale 33 superposability 87–9, 94, 118, 136–7,
Schaffer, J. 121 n43 186–97
Schlick, M. 64 n24 supervenience 74–6, 81, 95
Scholastics 166–8 swerving atoms 164
scientific methods 139–50 Swoyer, C. 106 n14, 109 n21
Sedley, D. 158 n6, 162 n24 symmetry 45–6, 48
Seidel, J. 71 n38, 71 n39, 193 n16 symplectic spaces 11
selection problem 42–4, 76
Sellars, W. 106 n15 Tarski, A. 32 n61
separable metric spaces 46–7 taxi-cab metrics 23
Shoemaker, S. 9 n3, 53, 106 n13, temporality 137–8
108 n18, 111 n26, 112 n28 tensors 11, 21
Sidelle, A. 9 n3 test bodies 51 n33
simplicity: Theophrastus 160
best-system approaches 55–6, 58, ties 73–4
59–66 time 137–8, 184–5
and ontology 139–49 time-independence 40–1
Simplicity and balance 67–72 Tits, J. 196
Skow, B. 32 n63, 82 n7, 84 n8, 122 n46 Todd, R. 164 n34, 165 n35, 165 n37
Soames, S. 37 n5 Tooley, M. 51 n32, 79 n1, 112 n28;
Sober, E. 144 n10, 148 n17 see also Armstrong–Dretske–Tooley
sophisticated substantivalists 82 approach
Sorabji, R. 18 n24, 158 n5, 160 n14, Torretti, R. 65 n27
163 n25, 165 n35, 165 n36, torus shapes 17, 20, 50
167 n44 trans-world identity 92 n17, 99, 114 n31
Sorenson, R. 101 n27 truth conditions 37
Sormani, C. 44 n13, 154 n11 Tweedale, M. 106 n14
spacetime 137–8 twisted cylinders 17, 188
species-relativity 67–8, 70, 72, 73–4 two-point homogeneity 189–90, 192,
spheres: 196–7
constant curvature 18
crumpled spheres 88, 188 ultralimits 44 n13
Euclidean metric on the uncountable metric spaces 46–7
Sphere 25 uni-dimensionality 41–5, 183
shapes of balls and spheres 23 universal metric spaces 46–7
spherical geometry 15, 65, 71 Urbach, P. 144 n9
Spinoza 171 n57 Urysohn space 47, 49, 197 n28
static worlds 55 Uspenskij, V. 47 n24
Stein, H. 157 n1
Stevens, Justice 54 Vallentyne, P. 110 n24, 114 n31
Stoics 164–6 van de Velde-Schlick, B. 64 n24
strength 55–6, 58, 59–60, 67–72 van Fraassen, B. 60 n16, 61, 80 n3,
substantivalism: 100 n24, 112 n29, 123 n49,
and primitivist approaches 81–2, 83–95 123 n50, 123 n51, 124 n52,
and relationalism 1–4, 49, 50–2, 139 125 n54, 127 n57
sophisticated substantivalists 82 vanishing curvature, spaces of 19
index 219
Varzi, A. 101 n27 Westfall, R. 161 n18, 163 n27

Vershik, A. 47 n24 Weyl, H. 154 n14
void space 3, 38 n8, 159, 162–8 Weyl Tile argument 154–5
bubbles of void 175, 182 Wolf, J. 17 n23, 19, 189 n6,
and Leibniz 174–5, 179–82 191 n10
see also empty worlds Woodward, J. 75 n48,
79 n2
Wang, H.-C. 196
Weatherson, B. 114 n30 Yablo, S. 114 n31

(Gordon Belot) Geometric Possibility

Uploaded by

Document Information

Original Title

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

(Gordon Belot) Geometric Possibility

Uploaded by

Copyright:

Available Formats

Geometric Possibility

This page intentionally left blank

This book is about one aspect of the substantival-relational debate con-

I Possible Structures of Space 8

II Spatial Structure for Relationalists 35

III Best-System Approaches 54

V Necessitarian Approaches 102

[S]pace is something; but like time . . . [it] is a general order of things.

It is helpful to distinguish between two sorts of philosophical debates

Substantivalists take space to consist of parts standing in geometric

of possibility: we should retool the substantivalist truth conditions so

Overview. Chapter I below addresses a preliminary question—What

a reasonable auxiliary assumption). Roughly speaking, the three condi-

in many ways analogous to the necessitarianism about laws developed by

Our approach to general metric spaces bears the undeniable imprint

of mathematical structures that represent possible spatial geometries cor-

physical theory.8 What does it mean to speak of the distance relations at a

classical non-Euclidean geometries (elliptic, hyperbolic, and spherical

by considering a characterization of Euclidean space in terms of the

Axiom VIII. For any x1 , x2 , x3 , x4 2 X the determinant of the

image under f of some x 2 X and f (x1 ) ¼ f (x2 ) implies x1 ¼ x2 ). We say that

3 Candidates for Spatial Structure

considered—and in some cases I mention examples of structures permitted

Generalization 0: Euclidean Geometry. Our first generalization

Generalization 1a: The Classical Geometries. This class comprises

Hyperbolic geometry is the geometry associated with Gauss, Bolyai, and

Generalization 1b: Flat Spaces. Let us now generalize Euclidean

Euclidean geometry is the only one of the classical geometries that

widespread willingness to consider them as representing possible structures

Generalization 2: Spaces of Constant Curvature. This class arises

geometry spaces of vanishing curvature (this is a variant on ‘flat space’),

Generalization 3: Riemannian Manifolds. The notion of a Riemann-

The resulting metric space is called a Riemannian manifold when X satisfies

a metric space. Clearly, any Riemannian metric tensor on a (connected)

Generalization 4: Generalizations of Riemannian Manifolds. There

of a Riemannian manifold, we take a Hilbert manifold to be a space with the

d( (x1 , y1 ), (x2 , y2 ) ) ¼ jx1 x2 j þ jy1 y2 j:

path metric space constructed by gluing together a collection of edges,

Generalization 6: Metric Spaces. Of course, many metric spaces are

Example 3.7 (Integer Lattices). Let Rn be equipped with its Euclidean

to the way distances are defined in differential geometry.50 (ii) Adopting

in which the distance between points is not determined by the lengths of

could be—but in the following chapters I will be as conservative as

pointless metric space will have points corresponding to each of the

of plenitude as ruling out (at least) the sort of gappiness in question.

geometric relations at that world are distance relations. As observed above,

R(w1 , w2 , x1 , x2 , y1 , y2 , z1 , z2 )()jd(w1 , w2 ) d(x1 , x2 )j$jd(y1 , y2 )

We call (X, d ) replete if whenever R(w1 , w2 , x1 , x2 , y1 , y2 , z1 , z2 ) there

jd(w1 , w2 ) d(u1 , u2 )j ¼ jd(y1 , y2 ) d(z1 , z2 )j:

It turns out that if (X, d ) is replete then R fixes d up to a scale factor.65 So

The more universal a concept is the more frequently it enters into

1 Substantivalism and the Structure of Space

Finitude. Space is finite in extent if and only if there is an upper

Remark 1.1 (Truth Conditions). I take the notion of a truth condition to be

Conservative approaches make do with actual material configurations. In

However, the situation becomes more puzzling when one consider

a such world under this approach can be coherently viewed as characterizing

These considerations motivate the following scheme for adapting sub-

Selection Problem for one-particle relationalist worlds. Less dramatic ver-

Worse, problems arise even for the simple case of one-dimensional

d({xn }, {yn }) ¼ sup jxk yk j:

z7!{dZ (z, z0 ), dZ (z, z1 ) dZ (z1 , z0 ), . . . , dZ (z, zn ) dZ (zn , z0 ), . . . }:

is an isometry. So ‘1 is a universal metric space for the class of countable