Similarity, Topology, and Physical Significance in

Relativity Theory
Samuel C. Fletcher∗
Logic & Philosophy of Science, UC Irvine
scfletch@uci.edu
June 27, 2013

Abstract
Stephen Hawking, among others, has proposed that the topological stability of
a property of spacetime is a necessary condition for it to be physically significant.
What counts as stable, however, depends crucially on the choice of topology. Some
physicists have thus suggested that one should find a canonical topology, a single “right”
topology for every inquiry. While certain such choices might be initially motivated,
some little-discussed examples of Geroch and some propositions of my own show that
the main candidates—and each possible choice, to some extent—faces the horns of a
no-go result. I suggest that instead of trying to decide what the “right” topology is
for all problems, one should let the details of particular types of problems guide the
choice of an appropriate topology.

1 Introduction
There are many reasons to consider notions of similarity in philosophy and, more specifically,
in philosophy of science. They underlie Lewis’s famous system for counterfactual semantics,
modal logic, laws of nature and causation [Lewis, 1973, 1986]. More recently, Halvorson
[2012] has suggested that similarity is salient for characterizing scientific theories formally:
one cannot recover a theory from its models unless one encodes the similarity amongst the
models’ truth-valuations topologically. And in philosophy of physics, Manchak [2012] has
remarked that placing a topology on models of general relativity can describe how different
possible relativistic worlds (i.e., relativistic spacetimes) are “nearby” one another.
This topological approach to similarity has been used by physicists working in mathe-
matical relativity since the 1970s. In this context the notion of stability has been crucial.
∗
Thanks to audiences at Irvine and Pittsburgh for helpful comments, especially Jeff Barrett, Ben Feintzeig,
Dennis Lehmkuhl, David Malament (who inquired about corollary 2), John Norton, and Chris Wüthrich, and
to Jim Weatherall for much guidance, including conjecturing proposition 3 and sketching part of proposition
4. John Manchak spurred me to generalize a special case of proposition 5. Part of the research leading to
this work was completed with the support of a National Science Foundation Graduate Research Fellowship.

1
Roughly, a property of an object O of a specified class (like mathematical models, solutions
to a differential equation, etc.) is stable when all objects in that class sufficiently similar to
O also have that property, the name “stability” coming from the intuitive picture that the
property is preserved under arbitrary (but sufficiently small) perturbations. In this paper I
am interested in exploring how this literature connects similarity—in particular, stability—
with judgments of physical significance. For example, Hawking has asserted that “the only
properties of space-time that are physically significant are those that are stable in some ap-
propriate topology”1 and that “For physical purposes it is sufficient to prove that a theorem
holds generically” [1971, p. 395], that is, for almost all objects of interest.
One attraction of Hawking’s proposal is that it seems to reduce part of a philosophical
question to a technical question: in certain cases, instead of puzzling over whether a property
is “physical significant,” one instead may determine on which sets it is unstable; instead of
assessing the import of apparent isolated counterexamples to a theorem, one simply proves
that the theorem holds generically. Whether this reduction is successful, however, depends
first on justifying why there should be a connection between physical significance and topo-
logical stability at all. Then, given a satisfactory justification, one must explicate what an
“appropriate topology” is supposed to be. After addressing the former query in §2, I focus
on the latter in the remainder. Since a choice of topology on spacetimes encodes particular
ways in which those spacetimes are similar, an appropriate topology is one that gets this
notion of similarity right. While it would greatly simplify matters if there were a canonical
such topology, in §3–4 I consider two classes of topologies often considered in the literature
and find them flawed for this purpose. Against suggestions from some physicists, in §5 I sug-
gest instead that there cannot be a canonical topology, and that an “appropriate” topology
must covary with the context of inquiry. Without a canonical topology, however, stability
itself does not directly settle any conceptual questions about what is physically significant.
Rather, whether a property counts as physically significant in a model depends upon the
choice of topology—how one considers models to be relevantly similar.

2 Similarity, Topology, and Physical Significance

What does it mean to say that a property is physically significant? There is not likely any
straightforward univocal concept that wholly underlies the broad use of this phrase in the
physics literature, but it will suffice here to draw out some of its connections with epistemic
warrant and stability.2 Physicists use mathematical models to represent physical phenom-
ena, including past observations and potential predictions. But they must be judicious in
using models to make inferences about phenomena, as models are often idealized or only
approximate, for example. “Physical significance” expresses an aspect of this partial con-
nection between models and phenomena. One says that a property of a model of physical
phenomena has physically significance to the degree to which one has warrant to infer that
property about the physical phenomena.
The connection with stability becomes evident in considering that the observed data
1
See also Geroch [1971, p. 70] and Hawking and Ellis [1973, p. 197].
2
I take myself to be making some tentative suggestions rather than undertaking any robust project of
conceptual analysis.

2
used to build such models, based on quantities like length, energy and position, are typi-
cally imprecise. Imprecise data usually yield imprecise models, or rather a range of models
compatible with the data’s imprecision. For convenience, however, scientists typically build
a single model and represent the imprecision in other ways. But the inferences warranted
through that model must cohere with the inferences one would have made through the range
of models—that is, these inferences must be compatible not just with the observed data, but
also with the whole set of data values falling within the measurements’ range of imprecision.
Thus any inferences that crucially depend on perfectly precise data will never be warranted.
Requiring the stability of a property as a necessary condition for its physical significance
enforces this compatibility between imprecise data and the inferences one draws: the only
properties about phenomena one can infer from a model are those that all arbitrarily similar
models share.3
It is important to contrast the epistemological character of physical significance as such
with the more metaphysical character of being “physically (un)reasonable” [Smeenk and
Wüthrich, 2011, Manchak, 2011, Earman, 1995, Ch. 3.4] or just plain “(un)physical” [Norton,
2008, §3.2]. Although physicists do not usually explicitly distinguish these, they tend to say
that a model is physically unreasonable when they wish to exclude it as a genuine physical
possibility countenanced by a theory. By contrast, physical significance tends to refer to
the warrant to infer properties of physical phenomena from a model. There are interesting
connections to explore between the physically significant and the physically reasonable,4 but
the following discussion shall focus on the former.
In order to apply the stability criterion for physical significance, however, one should
formalize the notion of similarity it depends upon. Topology is a natural choice.5 A topology
on a class of objects determines notions of convergence for sequences and continuity for
parameterized families, using its system of open neighborhoods to encode a weak sense of
similarity. A property P of an object O (or, more generally, of each of a set of objects
{Oα }) is then stable just in case there is an open set containing O (resp. {Oα }), all of whose
members also have P . Further, property P holds generically on a collection S when it holds
on an open subset of S that is dense in S. Density ensures that elements of S that do not
have P do so unstably, and openness ensures that elements that do have P do so stably.
However, if the collection of objects is infinite, as is the case with relativistic spacetimes,
then there will be infinitely many topologies one can place on the collection, topologies that
can differ regarding whether a property is stable or generic on a subcollection. How can one
decide which topology is appropriate? Perhaps there is in fact a canonical topology: a single
choice of topology over the collection of relativistic spacetimes that should apply whenever
such a topology is needed. Such a position has been suggested by Geroch, who writes, “It
is important, I feel, that one settles on one (or possibly two) topologies in which to work
3
Stability plays an analogous role in securing inference under idealization. Scientists often use idealized
models to make inference more tractable. In these cases, one would like to infer properties of phenomena the
de-idealized model represents from those of the idealized model. Even if the idealization is not too severe,
such an inference will not be warranted unless the property in question is stable.
4
See, for example, Fletcher [2012, §3.1] for such connections in the case of excluding the indeterministic
trajectories of Norton’s dome [2008].
5
Many other choices, like uniform spaces, are strictly stronger than a topology, so one expects at least to
reckon with topological structure.

3
rather than discovering a new topology for each new theorem” [1970, p. 269],6 and more
strongly by Lerner [1972, 1973] and Lerner and Porter, who advocate for a particular choice:
“if one regards all Lorentz metrics on M as being on an equal (mathematical [sic] footing,
it appears that the only acceptable choice for a topology is the Whitney fine C k topology”
[1974, p. 1413]. In the remainder I investigate the viability of this canonicalism about
topologies over spacetime, considering Lerner’s choice in §3 and another initially plausible
class of candidates in §4. In examining the problems each faces, the case against canonicalism
will emerge.

3 The Open Topologies

Recall that a relativistic spacetime is an ordered pair (M, gab ), where M is a four-dimensional
smooth manifold7 and gab is a smooth Lorentzian metric on M , whose indices are abstract.8
Then the collection of objects to topologize consists of the Lorentzian metrics on a fixed
manifold M , which I denote L(M ).9 The Whitney fine C k topology, also called the C k open
topology, may then be defined as follows. First, let hab be some (inverse of a) smooth Rie-
mannian metric on M , and define the “distance” function between the k th partial derivatives
0
of two Lorentz metrics gab , gab relative to hab as

ru sv 0 0 1/2
[h h (grs − grs )(guv − guv )] ,
 k = 0,
0 a b a b ru sv
d(g, g ; h, k) = [h 1 1
···h h h
k k
(1)
 0 0 1/2
k > 0,
⊗ ∇a1 · · · ∇ak (grs − grs )∇b1 · · · ∇bk (guv − guv )] ,


where ∇ is the Levi-Civita derivative operator compatible with hab . I have omitted the
abstract indices in the arguments of d since they needlessly clutter the notation, and I will
hereafter continue to drop them when they will never be contracted.
A particular choice of h effectively determines a smoothly varying basis for the tangent
space at each point of M , through which d(g, g 0 ; h, k) compares the components of the k th
order derivatives of g and g 0 . Then the sets of the form
Bk (g, ; h) = {g 0 : sup d(g, g 0 ; h, 0) < , . . . , sup d(g, g 0 ; h, k) < } (2)
M M

constitute a basis for the C k open topology, where g ranges over all Lorentz metrics,  ranges
over all positive constants,10 and h ranges over all Riemannian (inverse) metrics. One can
view these basis elements as generalizations of the -balls familiar to metric spaces.
6
I do not attribute to him outright advocacy, since a careful reading reveals an admixture of methodologi-
cal pragmatism: “I think it is important . . . to eventually settle on one or possibly two topologies with which
to work. Hardly any economy of thought results if there are hundreds of topologies in use” [1971, p. 73].
Moreover, later writings indicate a preference for the methodologically contextualist approach I take in §5:
“The topology one chooses in practice depends on what one wants the topology to do” [1985, p. 175–6].
7
One also requires M to be connected, paracompact, and Hausdorff.
8
That is, the super- and subscripts of tensor fields like gab label copies of vector spaces in which the fields
reside. See, e.g., Malament [2012, §1.4].
9
One might of course also wish to compare spacetimes whose underlying manifolds are not diffeomorphic.
Although Hawking and Ellis [1973, p. 198] state that this can be done, to my knowledge no one has done so
nontrivially for all spacetimes.
10
One can choose a different  for each derivative order, but the resulting basis generates the same topology.

4
 But how does one justify the C k open topology as canonical? For instance, how should
one chose the right value of k? One way is to investigate examples of stability about which
one has a strong intuition, ruling out available topologies that do not meet them. For
example, in discussing a theorem proving the stability of the strong energy condition11 in
the C 2 open topology, Lerner writes,

It should be pointed out that [this theorem] is not true in any of the weaker
topologies frequently used . . . If we agree that any reasonable topology . . . should
allow perturbations preserving the existence of non-zero rest mass, we may take
this as further evidence in favor of the [open] topologies. (Lerner 1973, p. 28)

Indeed, it seems that virtually all of the results regarding stability and genericity of global
properties of spacetimes have used one of the open topologies. For example, the encyclo-
pedic monograph Global Lorentzian Geometry, which has an entire chapter on “stability
of [geodesic] completeness and incompleteness,” defines only the open topologies for these
purposes [Beem et al., 1996, p. 63 & Ch. 7].
However widely accepted, the universal appropriateness of the open topologies has not
gone unquestioned. Geroch [1970, 1971] has provided a pair of examples that illustrate some
surprising features of the C 0 open topology in particular. His first example is a sequence
that seems like it should converge to Minkowski spacetime but in fact does not. Explicitly,
the sequence of metrics
 
m 1
g ab = 1 + 2 (da t)(db t) − (da x)(db x) − (da y)(db y) − (da z)(db z) (3)
m + x2 + y 2 + z 2

on R4 , where t, x, y, z are scalar coordinate fields, does not converge as m → ∞ to the

Minkowski metric

ηab = (da t)(db t) − (da x)(db x) − (da y)(db y) − (da z)(db z), (4)

even though the “bump,” remaining centered at the coordinate origin, decreases in amplitude
m
to zero. This is because g → η in the C 0 open topology if and only if for every neighborhood
m
of the form B0 (η, ; h), we have g ∈ B0 (η, ; h) for m sufficiently large. But, given any
h, one can pick another multiplied by a conformal factor growing sufficiently rapidly that
m
supM d(η, g; h, 0) = ∞.
Geroch’s second example is the one-parameter family

Λ = {λgab : λ > 0}, (5)

with a fixed gab on a non-compact M , which strikingly does not trace out a continuous curve
in the C 0 open topology—indeed, it is everywhere discontinuous. To see this, note that
the family is continuous in the C 0 open topology if and only if for every λ0 > 0 and every
neighborhood of the form B0 (λ0 g, ; h), there is a positive open interval I 3 λ0 such that
This is the condition that for any timelike vector ξ a at any point of M , Tab − 12 T gab ξ a ξ b ≥ 0, where
11

Tab is the stress-energy tensor and T is its trace. See Hawking and Ellis [1973, p. 95] or Malament [2012,
p. 166].

5
{λgab : λ ∈ I} ⊆ B0 (λ0 g, ; h). But, as with the first example, given any h, one can pick
another multiplied by a conformal factor growing sufficiently rapidly that for any δ 6= 0,
supM d(λ0 g, (λ0 + δ)g; h, 0) = ∞.
This example is particularly surprising because one can interpret the elements of Λ to
be physically equivalent, representing mere changes of units, say.12 In fact, one can prove
quite general results regarding the conditions under which a sequence converges or a family
is continuous in the open topologies. Specifically, the following is sketched by Golubitsky
and Guillemin [1973, p. 43–44]:
n
Proposition 1. Let g, {g}n∈N be Lorentz metrics on a non-compact manifold M . Then
n
g → g in the open C k topology on L(M ) iff there is an open C ⊂ M with compact closure
such that:
n
1. for sufficiently large n, g |M −C = g|M −C ; and
n
2. g |C → g|C in the open C k topology on L(C).

In other words, a sequence converges in the C k open topology just in case its elements
eventually equal the limit point everywhere except at most on (the interior of) a compact
set, a criterion of convergence even stronger than uniform! One can then use this theorem
to prove (see §A.1) a necessary condition for a family of Lorentz metrics to be continuous.
Proposition 2. Suppose that L(M ) is given the C k open topology, with M non-compact. If
f : R → L(M ) is continuous, then for every x0 , x1 ∈ R, there is some compact C ⊂ M such
that f (x0 )|M −C = f (x1 )|M −C .
Thus any pair from a continuous one-parameter family of Lorentz metrics must always
be equal everywhere except at most on a compact set. Intuitively, one might picture the
difference between any such pair as a “bump in a rug” that the function f pushes around.
Although the bump may be bigger or smaller, wider or narrower, it always has compact
support. This is clearly a quite restricted class of continuous families.13
One might object that the discussion of canonicalism was motivated by considering the
connection of stability with physical significance, but Geroch’s examples and propositions 1
and 2 bear on convergence and continuity. Why should problems with the latter bear on
the use of topology for the former? Two responses are on offer. First, if the canonicalist is
not willing to deny that continuity and convergence should ever be considered, she would
already concede part of her position in suggesting that different topologies should be used for
12
It also demonstrates that the open topologies, like the other topologies that I will consider, “over-
represent” the physically possible Lorentz metrics on M since in general they represent isometric spacetimes
through distinct points. One can compensate for this defect somewhat by ensuring one constructs only
invariant topologies [Geroch, 1970, p. 281–2], ones for which the pushforward map induced by any element
of the diffeomorphism group of M acts on L(M ) as a homeomorphism. Indeed, all of the topologies considered
in this paper are invariant in this way.
13
Lerner is aware of Geroch’s examples and propositions 1 and 2 [Lerner, 1972, 1973], but concludes from
them that one must give up talking about continuity and convergence. Considering that topology is typically
introduced (in part) in the first place to treat continuity and convergence, a more measured response would
be to reject the open topologies as canonical.

6
different purposes. Second, convergence, continuity and stability are interdependent because
they are all determined by a topology’s lattice of open sets. Recall, for instance, that the
stability of a property depends on the existence of a certain open set. Thus it is in a sense
easier for a property to be stable in a finer topology, since there are more open sets available.
In particular, if a property is stable on a certain set in a given topology T , it is stable in
every topology finer than T . Recall as well that the convergence of a sequence depends on
certain aspects of every open neighborhood of its purported limit point. Thus it is in a sense
easier for a sequence to converge in a coarser topology, since there are fewer open sets that
must fulfill the proper role. In particular, if a sequence converges in a given topology T , it
converges in every topology coarser than T .
Geroch’s examples and propositions 1 and 2 therefore suggest that there are more open
sets in the open topologies than one might have initially thought, which would perhaps
make stability too easily achieved. For example, if a property is stable for g in the C 0
open topology on L(M ) for non-compact M , then it obtains on some basic neighborhood
B0 (g, ; h). However, for any g 0 ∈ B0 (g, ; h) distinct from g on a non-compact set, by
choosing h0 = Ωh/d(g, g 0 ; h, 0) for some unbounded Ω, there is no 0 for which g 0 ∈ B0 (g, 0 ; h0 ).
But one might think that whether g 0 can be sufficiently close to g, for some standard of
sufficiency, should not depend on the choice of h, the smoothly varying choice of a tangent
basis in which to compare g and g 0 . That the open topologies do have this dependence is
ultimately what is responsible for propositions 1 and 2.

4 Continuity in the Geometric Sense and the Compact-

Open Topologies
Given the problems that the open topologies face, one might very well abandon Lerner’s
suggestion and investigate other possible choices for a canonical topology. One idea comes
from Geroch [1969], who has proposed a way of interpreting certain limiting relations entirely
geometrically through the continuity (smoothness, etc.) of certain fields. Roughly, in the
simplest case of a one-parameter family, one constructs a 5-dimensional manifold from the 4-
dimensional manifolds of the family “stacked” by their identifying parameter. More precisely,
t
suppose that one is given a family of metrics {g ab }t∈R on a fixed manifold M .14 Let M be
a manifold diffeomorphic to M × R and let ψ (t) : M → M be a family of embeddings such
(t)
that the field t̃|(p,t) = (ψp )∗ t̄ is smooth, where t̄ is a constant field on M of value t. Then
one can define a symmetric field Γab on M with signature (+, −, −, −, 0) by stipulating
(t) t
that (Γab )|(p,t) = (ψp )∗ (g ab ), hence Γab ∇a t̃ = 0, where ∇ is a fixed derivative operator
satisfying ∇b ∇a t̃ = 0 and ∇a Γbc = 0.15 Now, for each p ∈ M the points ψ (t) (p) for all
t form a smooth curve and the collection of all such curves for all p form a congruence
on M. Thus there is a vector field τ a on M tangent to the curves of this congruence
14
Geroch does not require that the metrics be defined on diffeomorphic manifolds, but I can confine
attention to that case here.
15
For those familiar with the geometrized formulation of Newtonian gravitation, ∇a t̃ functions much like
the temporal metric and Γab like the spatial metric, except that the latter is Lorentzian instead of Riemannian
and only assumed to be smooth on each t̃ = const. hypersurface. Thanks to Jim Weatherall for emphasizing
this to me.

7
satisfying τ a ∇a t̃ = 1. This allows one at last to define a unique symmetric field Γab such
that Γab τ a = 0 and Γab Γbc = δac − τ c ∇a t̃.16 With this construction in place, we can say that
t
the family g ab on M is continuous in the geometric sense when the corresponding field Γab
is continuous everywhere on M. (Analogous definitions would apply to smoothness, etc.)
One can similarly define the limit of a sequence of metrics by embedding the sequence in
a one-parameter family. A great appeal of this proposal is that it uses the natural, widely
accepted geometrical formulation of a relativistic spacetime to do the work of choosing the
canonical topology. It turns out that the topology determined by all the families continuous
in the geometric sense is well-known:
t
Proposition 3. A family of Lorentz metrics {g}t∈R is continuous in the geometric sense iff
it is continuous in the C 0 compact-open topology.17

For a proof, see §A.2. A basis for the C k compact-open topologies, for any non-negative
integer k, may be written as sets of the form

Bk (g, ; h, C) = {g 0 : sup d(g, g 0 ; h, 0) < , . . . , sup d(g, g 0 ; h, k) < }, (6)

C C

where g ranges over all Lorentz metrics,  ranges over all positive constants, h ranges over
all Riemannian (inverse) metrics, and C ranges over all compact subsets of M . The essential
difference between the open and the compact-open topologies is that the former “control”
behavior everywhere on the manifold whereas the latter do so only on compact subsets.
Notably, one can show that, unlike with the open topologies, the sequence defined by
eq. 3 converges to the Minkowski metric and the family defined by eq. 5 is continuous rel-
ative to the C k compact-open topologies. These topologies are also attractive for having
a number of other interesting features. First, they coincide with the topology of C k com-
n
pact convergence—that is, a sequence of metrics g → g on M just when it and its partial
derivatives to order k (with respect to the Levi-Civita derivative operator compatible with
an arbitrary Riemannian metric on M ) do so uniformly on each compact C ⊆ M [Munkres,
n
2000, p. 283, Theorem 46.2].18 Second, if a sequence of C k metrics g converges to g, then
g is guaranteed to be at least C k as well [Munkres, 2000, p. 284, Corollary 46.6]. Third,
there is a close connection with homotopy. One can show that a family of Lorentz metrics
is continuous in the C k compact-open topology if and only if it traces out a C k path in
L(M ). So, in a way, the C k compact-open topology encodes which Lorentz metrics can be
continuously (to order k) deformed into one another.19
Like with the open topologies, however, Geroch has criticized the general appropriateness
of the compact-open topologies, contending that they rule counterintuitively on the sequence
16
This also parallels the construction of the covariant spatial metric in geometrized Newtonian gravitation.
Cf. fn. 15 and Malament [2012, p. 254, Proposition 4.1.12].
17
In particular, the C 0 compact-open topologies are the final topologies respectively associated with the
families continuous in the geometric sense, the finest topologies on L(M ) that make those families continuous.
18
The compact-open topology coincides with the topology of compact convergence on a function space
when the range of the functions is a metrizable space [Munkres, 2000, p. 285–6], and the bundle of Lorentz
tensors over M , being a finite-dimensional manifold, is metrizable.
19
Equivalently, the family is continuous in the C k compact-open topology just when the k-jets of the
family belong to the same path component.

8
of metrics
 
m m
g ab = 1+ (da t)(db t) − (da x)(db x) − (da y)(db y) − (da z)(db z) (7)
1 + (x − m)2
on R4 , where t, x, y, z are natural scalar coordinate fields.20 “The ‘bump’ in the metrics be-
comes larger as it recedes to infinity,” he writes, but the “sequence does approach Minkowski
space in the [C 0 compact-open] topology (because the metrics become Minkowskian in every
m
compact set)” [1971, p. 71]. In other words, since d(η, g; h, 0) is continuous for any choice
of (smooth) Riemannian h, its supremum is bounded on any compact set and will become
as small as one likes for sufficiently large m. However, “[i]ntuitively, we would not think of
this sequence as approaching Minkowski space” [1971, p. 71] (or presumably any spacetime
at all). Thus he takes the C 0 compact-open topology to be too coarse.
This example is less convincing than his examples for the open topology, however. It is
instructive to compare eq. 7 with the sequence of Maclaurin expansions of a real function
like sin(x). For any particular finite-order expansion, one can find a sufficiently large x
such that the expansion, evaluated at this x, differs from sin(x) by as much as one wishes.
But if one fixes some compact neighborhood of the origin, then the Maclaurin series con-
verges uniformly on that neighborhood. Similarly, the sequence given by eq. 7 converges to
Minkowski spacetime because the C 0 compact-open topology corresponds with the topology
of compact convergence. Just as the compact convergence of Maclaurin expansions seems
perfectly reasonable, it is not clear why the same cannot be said in the case of sequences of
Lorentz metrics.
However, there are other counterintuitive features of the compact-open topologies that
bear even more directly on stability. For example, consider Hawking’s theorem [Hawking
and Ellis, 1973, Prop. 6.4.9, p. 198]:
Theorem 1 (Hawking). The existence of a global time function on a relativistic spacetime
is equivalent to stable causality, an absence of closed causal curves that is stable in the open
C 0 topology.
If one of the compact-open topologies were to be canonical, one would want to know whether
Hawking’s theorem holds in it as well. It turns that that it does not, and spectacularly so. In
fact, according to the compact-open topologies, almost all spacetimes contain closed timelike
curves.
Proposition 4. Chronology violating spacetimes are generic in L(M ) in any of the C k
compact-open topologies.21
Corollary 1. No Lorentz metric is stably causal in any of the C k compact-open topologies
on L(M ).22
20
The formula for the first term is garbled in Geroch [1971, p. 71], but appears without error in Geroch
[1970, p. 280].
21
This slightly improves statements by Hawking [1971, p. 396–7] and Hawking and Ellis [1973, p. 198],
who advert without proof to the density of chronology violating spacetimes in L(M ) in any of the C k
compact-open topologies.
22
Cf. proposition 5.1 of Manchak [2012], which shows that each Lorentz metric is homotopic to one that
violates chronology. As alluded to above, there is a close connection between homotopy and the compact-
open topologies: the C k homotopy classes correspond with the path components of the C k compact-open
topologies.

9
 In particular, according to the compact-open topologies, not only does Minkowski space-
time fail to be stably causal despite its global time function, it turns out that not having
closed timelike curves is not physically significant! In other words, one would never have war-
rant to infer that a model of relativistic spacetime does not permit a form of time travel.23 An
alternative but equivalent definition of stable causality brings out why: a spacetime (M, g)
is stably causal with respect to the C 0 open topology just when there is a metric g 0 for which
there are no closed causal curves and whose light cones everywhere lie outside those of g.
By contrast, when stable causality is defined with respect to the C 0 compact-open topology,
the light cones of g 0 need only lie outside of those of g on a compact subset of M , leaving
the rest unconstrained and ripe for the sprouting of closed causal curves.
By contrast, if (M, g) already contains a closed timelike curve γ : I → M , then one can
pick a local basis element Bk (g, ; h, C) from any compact-open topology so that γ[I] ⊆ C
and  is small enough so that, for any g 0 ∈ Bk (g, ; h, C), γ is still g 0 -timelike.
Proposition 5. Every spacetime containing a closed timelike curve does so stably in any of
the C k compact-open topologies.
Thus one always has warrant to infer from spacetimes with closed timelike curves that they
represent the possibility for a type of time travel. One should also, for some relativistic
spacetime models without closed timelike curves, have warrant to infer that they do not
represent this possibility. That this never occurs under the compact-open topologies militates
against taking any of them as canonical.

5 Methodological Contextualism
Any canonical topology on L(M ) should have the ability to properly distinguish which se-
quences converge, which families are continuous, and which properties are stable or generic.
But as the previous two sections laid out, the two main classes of topologies in the litera-
ture fall short of these goals. The open topologies, advocated by Lerner, seem too fine to
treat convergence and continuity. The compact-open topologies, naturally suggested through
geometric continuity, seem too coarse for stable causality because their neighborhoods con-
trol behavior only on compact sets. Of course, that Geroch’s examples do evince genuine
problems for the former can well be challenged, and one may decide, according to one’s
inclinations, to bite the bullets of propositions 1 and 2, or 4, but this does not completely
resolve the issue of how to choose the canonical topology. Any proponent of a canonical
topology must decide without being ad hoc on which counterintuitive results to accept and
are obliged to provide an explanation as to why the intuitive features thereby denied do not
have the significance they seemed to.
The considerations already raised for the canonicalist can be cast in terms of a no-go
result. For instance, the results discussed in the previous section yield the following.
Proposition 6. There is no topology on L(M ) relative to which both of the following hold:
23
This result does not state that one can have no inductive evidence for certain global properties, as one
might interpret a theorem of Manchak [2011, Prop. 2, p. 414]. Rather, it states that even if one had data for
the whole universe, however imprecise, and fit that data to a relativistic spacetime, one could never conclude
that there were no closed timelike curves if one takes any compact-open topology as canonical.

10
 1. All one-parameter families continuous in the geometric sense are continuous; and

2. Hawking’s theorem holds.

One might object that asking for such compatibility is too much; perhaps there is hope for
compatibility with a weakened version of Hawking’s theorem, in which one demands the
existence of a global time function just in case the spacetime is stably causal when restricted
to compact sets (or rather their interiors). But this fails, too.

Corollary 2. Suppose TM is a topology on L(M ), for any M , that makes continuous all the
one-parameter families continuous in the geometric sense. Then, for any g ∈ L(M ) there is
no open C ⊆ M with compact closure such that g|C is stably causal in the topology TC on
L(C).

Because there are no subsets of M with the desired property, the analog of Hawking’s theorem
fails (as there remain many spacetimes with global time functions).24 This is just one of a
possibly large family of “no-go” results that the canonicalist must face.
But reminding oneself of the way these topologies are used suggests that one need not pick
any canonical topology at all. Examining the consequences of adopting one topology over
another is a part of the process of deciding which topology will be relevant for a given type
of problem. Hawking has emphasized as much: “A given property may be stable or generic
in some topologies and not in others. Which of these topologies is of physical interest will
depend on the nature of the property under consideration” [1971, p. 396]. Indeed, Geroch’s
later writings (see fn. 6) have indicated the same. If different topologies correspond to
different ways one can specify how spacetimes are similar, it is not surprising that different
topologies would be natural choices for different kinds of questions if those questions bear
upon different kinds of properties. It thus seems best to accept a kind of methodological
contextualism, where the best choice of topology is the one that captures, as best as one can
manage, at least the properties relevant to the type of question at hand, ones that relevantly
similar spacetimes should share. Thus, in contrast to the canonicalist, I would demand that
particular choices of topology must be justified relative to a context as much as one feasibly
can.
Methodological contextualism about topologies—at least in the sense of allowing oneself
to pick the most appropriate topology for a given application instead of deciding on one
in advance—would make all the above worries associated with picking a canonical topol-
ogy moot. But the contextualist fortunately still has the resources to choose reasonable
topologies—resources not so different from the canonicalists, but without the demand to
select a single topology for all purposes. One will expect that similar questions will tend
to demand similar topologies, so the process of justification need not be started afresh each
time. In particular, one should arrive at a particular choice of topology through reflective
equilibrium, balancing the demands of the current understanding of what different topologies
24
In fact, the stated weakening of stable causality with respect to the C 0 open topology is known to
be equivalent to non-total imprisonment, the condition that no future-inextendible causal curve eventually
enters but does not leave some compact subset of spacetime [Minguzzi, 2009, Theorem 1]. That non-total
imprisonment is much weaker than stable causality reveals that finding a weakening of stable causality that
preserves Hawking’s theorem is a subtle matter.

11
capture physically and what notions of similarity one is trying to capture with the implica-
tions of new mathematical results, as the many examples and propositions of §3–4 did for
the open and compact-open topologies. These new results may then change one’s intuitions,
which in turn may suggest further results to investigate. The more one can accumulate these
kinds of facts, the more there will be relevant data at hand for a particular type of inquiry so
that one can make a sharper, better justified conceptual decision regarding which topology
to use. Sometimes this will lead one to reject initially promising and intuitive choices, and
sometimes it will reinforce them. One need not postulate that this reflective equilibrium
lead to a stable limit; even if one has accumulated many results in favor of using a partic-
ular topology for some narrow type of inquiry, one should still be open to new facts and
connections that will disturb one’s equilibrium.25
Nevertheless, more work needs to be done characterizing how particular choices of topol-
ogy may be appropriate for a given kind of question. For example, per the discussion at
the end of §3, one might expect that some topologies fine enough for the stability of some
properties—like the open topologies—are too fine for certain sequences to converge, or vice
versa—like the compact-open topologies. So the topologies that might be natural candi-
dates for inquiries about stability may be different from those for inquiries about continuity
or convergence. At least in the case of stability, one may be able to characterize classes of
properties to which particular topologies are (in)sensitive, or the range of topologies in which
interesting properties, like stable causality, behave as one might expect. The usual classi-
fication of spacetime properties into local and global [Manchak, 2011, p. 413] is too broad
for these purposes.26 Part of the difficulty in answering this question for stability—indeed,
for any inquiry—stems from the small variety of topologies used in the literature. Theorems
about the stability and genericity of global properties generally use the open topologies (e.g.,
see Hawking and Ellis [1973, p. 198], Lerner [1973], and Beem et al. [1996, Ch. 7]). Theo-
rems about the stability of Cauchy developments use variants on the coarser compact-open
topologies (see Hawking [1971, p. 398–9] and Hawking and Ellis [1973, p. 252–254]). The-
orems concerning the convergence of relativistic spacetimes to Newtonian spacetimes (e.g.,
Malament [1986]) use (implicitly) a point-open topology, which is even coarser.27
It turns out that there is a simple modification to the C 0 open topology that makes the
one-parameter families defined by eq. 5 everywhere continuous while, unlike the compact-
open topology, still preventing the sequence defined by eq. 7 from converging. Take the basis
elements of the C 0 open topology (eq. 2), but restricted only to bounded pairs (g, h), ones for
which supM d(g, λg; h, 0) < ∞ for any positive λ. This prohibits choosing a Riemannian h
scaled by a conformal factor that grows too rapidly, eliminating the open neighborhoods of
each Lorentzian g that forced eq. 5 to be everywhere discontinuous. One can show, moreover,
that this topology lies between the open and compact-open topologies in coarseness. However
25
One can find this dynamic and non-teleological conception of reflective equilibrium in the literature on
moral theorizing as well [Schroeter, 2004].
26
This classification takes a property P of a spacetime (M, g) to be local if and only if all spacetimes locally
isometric to (M, g) also have P , and global otherwise. Thus both the topology of M and the existence of
a closed timelike curve are global properties, whereas only the latter has any hope of having an analog to
proposition 4.
27
The point-open topologies are defined similarly to the compact-open topologies (eq. 6), but require the
suprema be taken over only finitely many points in each basis element instead of over compact sets.

12
the sequence defined by eq. 3 still does not converge to Minkowski spacetime according to
this topology, so it still would not rule in the intuitively “right” way according to Geroch.
But if further refinements are found that produce a topology satisfying Geroch’s desider-
ata, might that topology end up being satisfactory for all demands? If I allow for the possi-
bility that the methods available for picking an appropriate topology may eventually single
out a unique choice, or perhaps very few, to what extent is methodological contextualism
really distinguished from a slightly liberalized canonicalism? The answer is methodological.
The two positions are not distinct because of differing ends—whether to use one topology
or many—but because of their differing means: what grounds we might have to prefer one
topology over another, and whether those grounds need to be articulated. A canonicalist
holds that because there are definitive reasons always to choose a single topology (or per-
haps very few), there is no reason to say why that choice is appropriate for a given type
of inquiry. By contrast, the contextualist takes the relevant reasons to be provided by the
type of problem at hand, not in advance, and that they should therefore be articulated and
reasonably defended. It bears emphasizing that the latter does not deny that there can be
principled reasons to pick out a certain topology, only that those reasons can ever be given
in enough generality to preclude attention to the details of the type of situation at hand. We
indeed be may be lucky for the sake of our economy of thought if a few topologies are always
appropriate, but we should not obstruct the development of new ones if they fit particular
purposes better.

A Proofs of Propositions
A.1 Continuity and the Open Topologies
Proposition 2. Suppose that L(M ) is given the C k open topology, with M non-compact. If
f : R → L(M ) is continuous, then for every x0 , x1 ∈ R, there is some compact C ⊂ M such
that f (x0 )|M −C = f (x1 )|M −C .28

Proof. The case where f is a constant function is immediate, so suppose otherwise and pick
arbitrary distinct x0 , x1 ∈ R, assuming without loss of generality that x0 < x1 . I claim that,
given any r ∈ [x0 , x1 ], there is an open (relative to [x0 , x1 ]) interval Ir ⊆ [x0 , x1 ], containing
r, such that for any q ∈ Ir , there is some compact C(r, q) ⊂ M for which f (r)|M −C(r,q) =
f (q)|M −C(r,q) . For suppose otherwise, and
T∞ consider any sequence of intervals Ir1 ⊃ Ir2 ⊃ . . .
n n
such that Ir ⊆ [x0 , x1 ] for each n and n=1 Ir = {r}. One can then construct by induction
a sequence of metrics that converges to f (r). For the base step, let Irs1 = Ir1 and note that
there is some q1 ∈ Irs1 distinct from r such that f (r)|M −C 6= f (q1 )|M −C for any compact
C ⊂ M . (Such a q1 6= r exists because f is continuous.) For the inductive step, suppose
Irsn is given so that there is some qn ∈ Irsn distinct from r such that f (r)|M −C 6= f (qn )|M −C
for any compact C ⊂ M . Then pick some Irsn+1 such that sn+1 > sn and qn ∈ / Irsn+1 , noting
that there is some qn+1 ∈ Irsn+1 distinct from r such that f (r)|M −C 6= f (qn+1 )|M −C for any
compact C ⊂ M . The induction is complete, so by construction the sequence qn → r as
28
This proposition may be generalized to families of Lorentz metrics parameterized by any path-connected
space.

13
n → ∞, and for each n, f (r)|M −C 6= f (qn )|M −C for any compact C ⊂ M . But because f
is continuous, it follows that f (qn ) → f (r) as n → ∞ [Munkres, 2000, p. 130, Theorem
21.3], and by proposition 1, this implies in turn that there is a compact C ⊂ M for which
f (r)|M −C = f (qn )|M −C for sufficiently large n, which is a contradiction.
Next, note that the {Ir : r ∈ [x0 , x1 ]} form an open cover of [x0 , x1 ] (relative to [x0 , x1 ]).
The interval is compact, so by definition there is some finite subcover {Iri : i = 1, . . . , m},
each of whose elements has, for all q ∈ Iri , an associated compact C(ri , q) ⊂ M for which
f (ri )|M −C(ri ,q) = f (q)|M −C(ri ,q) . One may assume, without loss of generality, that r1 < . . . <
rm and that, because the interval is one-dimensional, no point of [x0 , x1 ] is included in more
than two of the IriS . Thus pick any qi ∈ Iri ∩ Iri+1 for i = 1, . . . , m − 1 and put q0 = x0 and
qm = x1 . Let C = i=1,...,m C(ri , qi−1 ) ∪ C(ri , qi ) and observe that

f (x0 )|M −C = f (q0 )|M −C = f (r1 )|M −C = f (q1 )|M −C = · · ·

= f (qm−1 )|M −C = f (rm )|M −C = f (qm )|M −C = f (x1 )|M −C .

Since C is compact and x0 , x1 were arbitrary, the proof is complete.

A.2 Equivalence of Geometric Continuity and Compact-Open Con-

tinuity
To prove proposition 3, it will be helpful to use the bundles of Lorentz tensors over M
and degenerate metrics over M × R, denoted L̂(M ) and Γ̂(M, R), respectively. The desired
equivalence is essentially a corollary of the following (adapted from Munkres [2000, p. 287,
Theorem 46.11]).

Lemma 1. Let X and Y be topological spaces, and give the set of continuous functions from
X to Y , denoted C(X, Y ), the (C 0 ) compact-open topology. If f : X × Z → Y is continuous,
then so is the induced function F : Z → C(X, Y ) defined by the equation (F (z))(x) = f (x, z).
The converse holds if X is locally compact29 and Hausdorff.
t
Proposition 3. A family of Lorentz metrics {g}t∈R on M is continuous in the geometric
sense iff it is continuous in the C 0 compact-open topology.30
(t)
Proof. Let ψp : M → M ×R be the embeddings that define the the 5-dimensional metric Γab ,
which corresponds to a cross-section Γ̃ of a bundle Γ̂(M, R) of 5-dimensional metrics. The
(t)
bundle homomorphism φ : Γ̂(M, R) → L̂(M ) defined by φ(Γ̃0|(p,t) ) = (ψp )∗ (Γ0ab ), for any local
(t)
section Γ̃0 at (p, t) ∈ M × R, is smooth because by definition the ψp are jointly smooth in p
t
and t. Thus, if the family g ab is continuous in the geometric sense, f = φ◦ Γ̃ : M ×R → L̂(M )
t
is continuous. Lemma 1 then entails that the map F : R → C(M, L̂(M )) defined by F : t 7→ ĝ,
t t
where ĝ is the cross-section of L̂(M ) corresponding to g ab , is continuous when its range is
given the C 0 compact-open topology.
29
A topological space is locally compact when each point has a compact neighborhood.
30
Using jet bundles [Golubitsky and Guillemin, 1973, Ch. 2.2–2.3], this proposition may be generalized to
any C k compact-open topology and families of Lorentz metrics parameterized by any smooth manifold.

14
 t
Conversely, suppose that the family g ab is continuous in the C 0 compact-open topology,
i.e., that F : R → C(M, L̂(M )) defined above is continuous. Since M is locally compact
t
and Hausdorff, lemma 1 entails that f : M × R → L̂(M ) is continuous. Thus (g ab )|p is
jointly continuous in t and p. Note that Γab is continuous when, for any smooth field αab on
(t) t
M × R, αab Γab is continuous. Now, for any (p, t) ∈ M × R, (αab Γab )|ψ(t) (p) = (ψp )∗ (αab )g ab ;
(t) t
by assumption (ψp )∗ (αab ) is smooth; and g ab is continuous because its inverse is. Thus Γab
is continuous, so Γab must be so as well by construction.

A.3 Stable Causality and the Compact-Open Topologies

The proof of proposition 4 uses proposition 5, but because the latter is a straightforward
computation, the order of presentation follows that of the main text. First, a lemma is
required allowing one to “interpolate” between any spacetime and a chronology-violating
region of Gödel spacetime.31

Lemma 2. Let (M, g) and (R4 , g 0 ) be two spacetimes. For any open S ⊂ M and any open
R ⊂ R4 with compact closure, there is a spacetime (M, g 00 ) such that g|M
00 0
−S = g|M −S and g|R
00
is isometric to g|U for some U ⊂ S.32

Proof. Pick a chart (V, ϕ) of M such that V ⊆ S and ϕ[V ] is an open ball of radius 4, i.e.,
ϕ[V ] = BR4 (~0, 4) = {~x ∈ R4 : k~xk < 4}, where k · k is the Euclidean norm on the coordinates
~x ∈ R4 . For brevity, define Ai = ϕ−1 [BR4 (~0, i)] for i = 1, 2, 3, and let r be a scalar field on V
defined by r|p = kϕ(p)k. Finally, let ψ : R4 → V be a diffeomorphism such that ψ[R] ⊆ A1
and define U = ψ[R].
Because all Lorentz metrics on R4 are homotopic [Finkelstein and Misner, 1959], ψ∗ (g 0 )
is homotopic to g|V , considering V as a submanifold. Thus there is some continuous function
f : [0, 1] → L(V ) such that f (0) = ψ∗ (g 0 ) and f (1) = g|V . One can then define the continuous
Lorentz metric 
g|p ,
 p ∈ M − A3 ,
γ|p = f (r|p − 2)|p , p ∈ A3 − A2 ,

[ψ∗ (g 0 )]|p , p ∈ A2 .


In order to produce the desired smooth metric g 00 , one can convolve γ with an appropriate
positive, symmetric mollifier on the region V − A1 . In more detail, define w : R → R to be
the smooth function ( 2
ce−1/(1−x ) , |x| < 1
w(x) =
0, |x| ≥ 1,
R
where c is a positive constant chosen so that R w(x)dx = 1. Further, define W : Rn ×
[0, ∞) → R as the jointly smooth function W (~x, ) = −n w(k~xk/), where W (~x, 0) =
limδ→0 W (~x, δ) is the Dirac delta, the convergence being understood in the distributional
31
For more on the properties of Gödel spacetime, see Malament [2012, Ch. 3.1].
32
Following Manchak [2012], one can use this lemma to answer affirmatively a question posed by Stein
[1970, p. 594] about whether it is always possible to continuously deform a spacetime into one containing a
closed timelike curve.

15
sense. Now, one can express γ in terms of its matrix components γαβ (~x) determined by the
chart (V, ϕ), allowing one to define on ϕ[V − A1 ] for some fixed  the new components
Z
γ̃αβ (~x) = W (~x − ~y , ec−1 w(k~xk − 5/2))γαβ (~y )d~y , (8)
ϕ[int(V −A1 )]

which are smooth on ϕ[V − A1 ].33 Moreover, for sufficiently small , the γ̃αβ approximate
the γαβ arbitrarily well on V − A1 .34 Therefore such γ̃αβ are the components of a smooth
Lorentz metric γ̃ on V − A1 . Note that, in the integrand of eq. 8 the function W becomes
the Dirac delta for k~xk ≥ 7/2 and k~xk ≤ 3/2, so on the points of V corresponding to these
coordinate regions, γ̃ is equal to g and ψ∗ (g 0 ), respectively. We can define at last

g|p ,
 p ∈ M − V,
00
g|p = γ̃|p , p ∈ V − A1 ,
 0
[ψ∗ (g )]|p , p ∈ A1 .


00 00 0
By construction, g|M −S = g|M −S since M − S ⊆ M − V and g|U is isometric to g|R since
00
g|A 1
= [ψ∗ (g 0 )]|A1 and U = ψ[R] ⊆ A1 .
Proposition 4. Chronology violating spacetimes are generic in L(M ) for every compact-
open topology.
Proof. Every spacetime with a compact M contains timelike curves [Hawking and Ellis,
1973, Prop. 6.4.2, p. 189], so suppose M is non-compact. Select any neighborhood N (g) of
an arbitrary g, which must contain a set of the form Bk (g, ; h, C). Letting S = M − C, by
lemma 2 there is some g 0 ∈ N (g) such that g|U0
is isometric to a chronology violating region
of Gödel spacetime for some U ⊂ M − C. By proposition 5, there is an open neighborhood
of g 0 consisting only of chronology violating metrics. Let Ak (g, N (g), Bk ) be the union of
all such open S neighborhoods
S S determined by the choices of g, N (g), and Bk (g, ; h, C), and
consider A = g N (g) Bk Ak (g, N (g), Bk ). By construction, every neighborhood N (g) of
each g contains an element of A, i.e., A is dense in L(M ); A is open, being the union of open
sets; and A contains only chronology violating spacetimes. So by definition the chronology
violating spacetimes are generic in L(M ).
Proposition 5. If the spacetime (M, g) contains a closed timelike curve, then g is stably
chronology violating in every compact-open topology.
Proof. Fix any Riemannian metric hab and note that one can write gab = ham µm hbn µm − hab
for some smooth vector field µa [Hawking and Ellis, 1973, p. 39]. One can thus express that
γ : I → M is a closed g-timelike curve with unit tangent vector ξ a as the condition that
1/2
|hab ξ a µb ||γ[I] > (hab ξ a ξ b )|γ[I] . Pick
(  2 )
|hab ξ a µb |
 = inf min 1, −1 ,
γ[I] (hab ξ a ξ b )1/2
33
This follows essentially from Oden and Reddy [1976, Theorem 2.6, p. 48–49].
34
Oden and Reddy [1976, Theorem 2.7, p. 49] show that, for the case where the integrand contains
W (~x − ~y , δ) with a fixed δ, the analog of eq. 8 would converge to γαβ as δ → 0 in Lp (ϕ[int(V − A1 )])-norm.
As before (cf. footnote 33), allowing δ to smoothly vary introduces no new complications.

16
and consider any g 0 ∈ B0 (g, ; h, C), where γ[I] ⊆ C. Writing gab 0
= ham µ0m hbn µ0m − hab , γ is
1/2
g 0 -timelike just in case |hab ξ a µ0b ||γ[I] > (hab ξ a ξ b )|γ[I] . Now, one can calculate that

0 0
ham hbn (gab −gab )(gmn −gmn ) = ham hbn (µa µb −µ0a µ0b )(µm µn −µ0m µ0n ) = [hab (µa −µ0a )(µb −µ0b )]2 ,

so putting η a = µ0a − µa yields that supC hab η a η b < . (The remaining calculations involve
fields defined on γ[I], so the subscript indicating as much will be omitted.) It follows from
this inequality and the Cauchy-Schwartz inequality that

|hab ξ a η b | ≤ (hab ξ a ξ b )1/2 (hab η a η b )1/2 < (hab ξ a ξ b )1/2 ≤ (hab ξ a ξ b )1/2 , (9)

where the last inequality uses the fact that, by definition,  ≤ 1. Then the reverse triangle
inequality entails that

|hab ξ a µ0b | = |hab ξ a (µb + η b )| ≥ |hab ξ a µb | − |hab ξ a η b | = |hab ξ a µb | − |hab ξ a η b |,

where the last equality follows since |hab ξ a µb | > (hab ξ a ξ b )1/2 > |hab ξ a η b | by the hypothesis
and equation 9. Applying this equation again along with the definition of  yields that

|hab ξ a µb |
|hab ξ a µ0b | > |hab ξ a µb |−(hab ξ a ξ b )1/2 ≥ |hab ξ a µb |−(hab ξ a ξ b )1/2 − 1 = (hab ξ a ξ b )1/2 .
(hab ξ a ξ b )1/2

Thus γ is g 0 -timelike, but g 0 was arbitrary so each element of B0 (g, ; h, C) contains a closed
timelike curve. Since B0 (g, ; h, C) is open in every C k compact-open topology, g must be
stably chronology violating in each.

Corollary 2. Suppose TM is a topology on L(M ), for any M , that makes continuous all the
one-parameter families continuous in the geometric sense. Then, for any g ∈ L(M ) there is
no open C ⊆ M with compact closure such that g|C is stably causal in the topology TC on
L(C).

Proof. Consider an arbitrary g ∈ L(M ) and an open C ⊆ M with compact closure. By

corollary 1, g|C is not stably causal in the C 0 compact-open topology on L(C). But if TM
makes continuous all the one-parameter families continuous in the geometric sense, it follows
from proposition 3 that it is no finer that the C 0 compact-open topology, hence g|C is not
stably causal in the topology TC on L(C) either.

References
Beem, John K., Paul E. Ehrlich and Kevin L. Easley. Global Lorentzian Geometry. 2nd ed.
New York: Marcel Dekker, 1996.

Earman, John. Bangs, Crunches, Whimpers, and Shrieks. New York: Oxford UP, 1995.

Finkelstein, David and Charles W. Misner. “Some New Conservation Laws,” Annals of
Physics 6: 230–243 (1959).

17
Fletcher, Samuel C. “What counts as a Newtonian system? The view from Norton’s dome,”
European Journal for Philosophy of Science 2.3: 275–297 (2012).

Geroch, Robert. “Limits of Spacetimes,” Communications in Mathematical Physics 13: 180–

193 (1969).

Geroch, R. “Singularities,” Relativity. Ed. Moshe Carmeli, Stuart I. Fickler, and Louis Wit-
ten. New York: Plenum, 1970.

Geroch, Robert. “General Relativity in the Large,” General Relativity and Gravitation 2.1:
61–74 (1971).

Geroch, Robert. Mathematical Physics. Chicago: University of Chicago Press, 1985.

Golubitsky, Martin and Victor Guillemin. Stable Mappings and their Singularities. New York:
Springer, 1973.

Halvorson, Hans. “What Scientific Theories Could Not Be,” Philosophy of Science 79.2:
183–206 (2012).

Hawking, S.W. “Stable and Generic Properties in General Relativity,” General Relativity
and Gravitation 1.4: 393–400 (1971).

Hawking, S.W. and G.F.R. Ellis. The large scale structure of space-time. Cambridge: Cam-
bridge University Press, 1973.

Lerner, David E. The Space of Lorentz Metrics on a Non-Compact Manifold. PhD disserta-
tion: University of Pittsburgh, 1972.

Lerner, David E. “The Space of Lorentz Metrics,” Communications in Mathematical Physics

32: 19–38 (1973).

Lerner, D. and J.R. Porter. “Weak gravitational fields,” Journal of Mathematical Physics
15: 1413–1415 (1974).

Lewis, David. Counterfactuals. Oxford: Blackwell, 1973.

Lewis, David. Philosophical Papers. Vol. II. Oxford: Oxford University Press, 1986.

Malament, David B. “Newtonian Gravity, Limits, and the Geometry of Space,” From Quarks
to Quasars. Ed. Robert G. Colodny. Pittsburgh: University of Pittsburgh Press, 1986.

Malament, David B. Topics in the Foundations of General Relativity and Newtonian Gravi-
tation Theory. Chicago: University of Chicago Press, 2012.

Manchak, John. “What Is a Physically Reasonable Spacetime?” Philosophy of Science 78:

410–420 (2011).

Manchak, John. “A Note on ‘Nearby’ Universes,” unpublished manuscript, 2012.

18
Minguzzi, E. “Chronological Spacetimes without Lightlike Lines are Stably Causal,” Com-
munications in Mathematical Physics 288: 801–819 (2009).

Munkres, James R. Topology. 2nd ed. Upper Saddle River, NJ: Pearson Prentice Hall, 2000.

Norton, John. “The Dome: An Unexpectedly Simple Failure of Determinism,” Philosophy

of Science 75: 786–798 (2008).

Oden, J.T. and J.N. Reddy. An Introduction to the Mathematical Theory of Finite Elements.
New York: Wiley, 1976.

Schroeter, François. “Reflective Equilibrium and Antitheory,” Noûs 38.1: 110–134 (2004).

Smeenk, Chris and Christian Wüthrich. “Time Travel and Time Machines,” The Oxford
Handbook of Philosophy of Time. Ed. Craig Callender. Oxford: Oxford University Press,
2011: 577–630.

Stein, Howard. “On the Paradoxical Time-Structures of Gödel,” Philosophy of Science 37.4:
589–601 (1970).

19