Categorical Generalization and Physical
Structuralism

Raymond Lal∗1 and Nicholas J. Teh†2

1Quantum Group, Department of Computer Science
University of Oxford

2DAMTP and Department of History and Philosophy of Science
University of Cambridge

April 15, 2015

Abstract

Category theory has become central to certain aspects of theoretical physics.
Bain (2013) has recently argued that this has significance for ontic structural real-
ism. We argue against this claim. In so doing, we uncover two pervasive forms
of category-theoretic generalization. We call these ‘generalization by duality’ and
‘generalization by categorifying physical processes’. We describe in detail how
these arise, and explain their significance using detailed examples. We show that
their significance is two-fold: the articulation of high-level physical concepts, and
the generation of new models.

Contents
1 Introduction 2

2 Categories and structuralism 3
2.1 Categories: abstract and concrete . . . . . . . . . . . . . . . . . . . . 3
2.2 Structuralism: simple and ontic . . . . . . . . . . . . . . . . . . . . . 5

3 Bain’s two strategies 6
3.1 A first strategy for defending (Objectless) . . . . . . . . . . . . . . . 7

3.1.1 The general argument . . . . . . . . . . . . . . . . . . . . . 7
3.1.2 The specific argument . . . . . . . . . . . . . . . . . . . . . 10

3.2 A second strategy for defending (Objectless) . . . . . . . . . . . . . . 13
3.2.1 Purely category-theoretic physics? . . . . . . . . . . . . . . . 14

∗raymond.lal@cs.ox.ac.uk
†njywt2@cam.ac.uk

1



3.2.2 O-objects in category-theoretic physics? . . . . . . . . . . . . 15

4 Two forms of categorical generalization 17
4.1 Generalization by duality . . . . . . . . . . . . . . . . . . . . . . . . 17

4.1.1 The framework for (GenDual) . . . . . . . . . . . . . . . . . 17
4.1.2 Algebraic-geometric duality and Einstein algebras . . . . . . 20

4.2 Generalization by categorification . . . . . . . . . . . . . . . . . . . 25
4.2.1 The framework for (GenCat) . . . . . . . . . . . . . . . . . . 25
4.2.2 A closer look at TQFTs . . . . . . . . . . . . . . . . . . . . . 27

4.3 The rôle of category theory in physics . . . . . . . . . . . . . . . . . 30

5 Conclusion 31

1 Introduction
Recently, Bain (2013) has offered an argument in defense of radical ontic structural
realism (ROSR) which draws on category-theoretic resources. His argument is largely
based on the idea that category theory offers the possibility of formulating physical
theories at an especially high level of generality; this level of generality is in turn
connected to the elimination of objects from physical theories.

In this paper, we will first argue that Bain’s argument fails.1 Our analysis of Bain’s
argument will show that it employs two distinct strategies, each of which implicitly
rests on a form of generalization that involves category theory. Indeed, it turns out
that these are the two main forms of category-theoretic generalization that have been
hitherto invoked in applications of category theory to physics.

This in turn paves the way for the more positive task of our paper. We shall extract
the two forms of generalization from Bain’s argument and develop each in turn, thus
providing a conceptual articulation of the two forms of generalization, and an analysis
of the role that each plays in the physical examples invoked by Bain.

This task is easy to motivate, as the broad question of how one can generalize phys-
ical theories is of immense interest to the philosophy of physics; indeed, it would be
a great discovery if it turned out that some very comprehensive and systematic means
existed. In our investigation, we shall be concerned specifically with textitcategory-
theoretic types of generalization within physics. An advantage of our approach is that,
rather than approaching the question of generalizing physical theories with a precon-
ceived and overly abstract idea of what ‘generalization’ amounts to, our approach has
the benefit of beginning with present applications of category theory to physics (in-
deed, the ones suggested by Bain’s argument) and extracting from these two forms of
generalization that are grounded in the actual practice of physics. Furthermore, our
results provide a partial explanation of how category theory can be useful for physics.

1Contemporaneously with our paper, Lam and Wuthrich (2013) have offered a discussion of Bain’s argu-
ment. We agree with many of their criticisms, but for the most part, they leave the deeply category-theoretic
aspects of his argument—especially the connection with generalization—undiscussed. This is our main
topic, and as such, there is little overlap between our work and theirs.

2



We now proceed to give a prospectus of the paper. Section 2 covers background
material that will be assumed in the rest of the paper: Section 2.1 reviews several key
notions from category theory, and Section 2.2 highlights the way in which category the-
ory is uncontroversially taken to be ‘structuralist’ by its practitioners, independently of
its potential connection with any philosophical doctrines concerning ‘structural real-
ism’.

Section 3 points out that there are really two different argumentative strategies at
work in (Bain, 2013) and analyses each in turn. Section 3.1 discusses the first strategy,
which involves what we call ‘generalization by duality’ (GenDual), and Section 3.2
discusses the second strategy, which involves what we call ‘generalization by categori-
fication’ (GenCat).

By our lights, both of these strategies are unsuccessful. However, their structure is
of independent interest, because they rest on important intuitions about how category
theory helps us to generalize physical theories. Section 4 develops these intuitions and
clarifies exactly how these two forms of generalization are supposed to work. Sec-
tion 4.1.1 develops the framework for generalization by duality (GenDual) and Section
4.1.2 proceeds to consider the particular form of (GenDual) that is invoked in Bain’s
Einstein Algebra example. We shall argue that the details of how (GenDual) is ap-
plied do not work in Bain’s favor. Similarly, Section 4.2.1 develops the framework
for generalization by categorification (GenCat) and Section 4.2.2 discusses its putative
application in TQFTs. Finally, Section 4.3 discusses the differences between (GenD-
ual) and (GenCat), and the role that category theory plays in each of these forms of
generalization.

2 Categories and structuralism
This section connects category theory with various themes that we will touch upon
in the paper. Section 2.1 introduces the basic notions of category theory that will be
required in the rest of the paper. Section 2.2 then proceeds to sketch the simple ‘struc-
turalist’ idea that is implicit (and indeed uncontroversial) in the mathematical use of
category theory. We contrast this with the metaphysical doctrine of ontic structural
realism, since subsequent sections will discuss whether category theory can furnish a
defense of ontic structural realism.

2.1 Categories: abstract and concrete
For the reader’s convenience, we begin by providing the definition of a category.

A category consists of: a class of objects, denoted A,B,. . . , and a class of mor-
phisms, denoted f,g, . . . , which satisfy the following conditions. A morphism f is
written as f : A → B, meaning that it has the object A as its domain, and the object B
as its codomain. There is a composition law −◦− for morphisms, which means that for
any two morphisms f : A → B and g : C → D, there is a morphism g ◦ f whenever
B = C (which is referred to as ‘type-matching’). This composition law is associative,
so that h ◦ (g ◦ f) = (h ◦ g) ◦ f for all morphisms with appropriate type-matching.

3



For every object A, there is always an identity morphism 1A, so that for any morphism
f : A → B, the equation 1B ◦ f = f = f ◦ 1A is satisfied.

Notice that in this definition, one does not need to further specify the details of the
objects and morphisms of a category. For all we know, the objects might not be sets,
and the morphisms might not be functions. If, however, the objects of the category are
sets that are equipped with some additional structure, then category theorists say that
the category is concrete.2

Here are some examples of familiar concrete categories (some of which will figure
in the discussion below):

• The category Set, whose objects are sets and whose morphisms are functions
between sets.

• The category Grp, whose objects are groups and whose morphisms are homo-
morphisms between groups.

• The category Top, whose objects are topological spaces and whose morphisms
are continuous maps between the spaces.

• The category Man, whose objects are smooth manifolds and whose morphisms
are smooth maps between the manifolds.

Any category comes equipped with a way of saying that two objects of the cat-
egory are ‘structurally the same’, viz. the concept of an iso-morphism between the
two objects. This is in fact a generalization of the usual set-theoretic definition of iso-
morphism, i.e. a bijective structure-preserving map between two structured sets. By
contrast, category theory offers a definition purely in terms of morphisms, and does not
explicitly quantify over the elements of the objects:

Two objects A, B in a category C are isomorphic iff there exist morphisms
f : A → B and g : B → A such that g ◦ f = 1A and f ◦ g = 1B.

For example, an isomorphism in Top is a homeomorphism, and an isomorphism in
Man is a diffeomorphism.

What are categories good for? Well, (elementary) category theory is mostly con-
cerned with universal properties. These define certain patterns of morphisms that
uniquely characterize (up to isomorphism) a certain mathematical structure. An ex-
ample that we will be concerned with is the notion of a ‘terminal object’. Given a
category C, a terminal object is an object I such that, for any object A in C, there is a
unique morphism of type f : A → I. So for instance, on Set the singleton {∗} is the
terminal object, and so we obtain a characterization of the singleton set in terms of the
morphisms in Set. Other standard constructions, e.g. the cartesian product, disjoint

2More formally, a concrete category is defined as a pair (C, F) where C is a category and F is a faithful
functor from C to Set. Note that there is a different use of ‘concrete category’ in some of the philosophy
of mathematics literature (Landry, 2013), where it means ‘a category that provides an interpretation of the
Eilenberg-MacLane axioms’. This is not the sense that we have in mind. Instead our usage follows that of
the major category theory textbooks, see e.g. (Mac Lane, 2000, p. 26) or (Lambek and Scott, 1988, p. 4).

4



union etc. can be characterized as universal. The power of universal properties is illus-
trated by their generality; e.g. Top also has a terminal object, which is the one-point
topological space.

Just as morphisms in a category preserve the structure of the objects, we can also
define maps between categories that preserve the composition law. Let C and D be
categories. A functor F : C → D is a mapping that: (i) assigns an object F(A) in
D to each object A in C; and (ii) assigns a morphism F(f) : F(A) → F(B) to each
morphism f : A → B in C, subject to the conditions F(g ◦ f) = F(g) ◦ F(f) and
F(1A) = 1F (A) for all A in C. Examples abound: we can define a powerset functor
P : Set → Set that assigns the powerset P(X) to each set X, and assigns the function
P(f) :: X 7→ f[X] to each function f : X → Y , where f[X] ⊆ Y is the image of f.

Functors ‘compare’ categories, and we can once again increase the level of ab-
straction: we can compare functors as follows. Let F : C → D and G : C → D
be a pair of functors. A natural transformation η : F ⇒ G is a family of functions
{ηA : F(A) → G(A)}A∈|C| indexed by the objects in C, such that for all morphisms
f : A → B in C we have ηB ◦ F(f) = G(f) ◦ ηA.

The notions of ‘functor’ and ‘natural transformation’ will be important in both
Sections 3 and 4, when we discuss the idea of an equivalence between physical theories
(characterized by means of functors), and the idea of a physical theory as a functor
respectively.

2.2 Structuralism: simple and ontic
As we have seen above, category theory can be thought of as ‘structuralist’ in the fol-
lowing simple sense: it de-emphasizes the role played by the objects of a category, and
tries to spell out as many statements as possible (about the mathematical properties of
the objects, e.g. the terminal-object example given above) in terms of the morphisms
between those objects. This much is clear from the formal properties of category the-
ory, as evidenced by the discussion of isomorphism and universal properties above.

These formal features of category theory have developed into a vision of how to
do mathematics. This has for instance been explicitly articulated by Awodey, who
says that a category-theoretic ‘structuralist’ perspective of mathematics, is based on
specifying:

[F]or a given theorem or theory only the required or relevant degree of
information or structure, the essential features of a given situation, for the
purpose at hand, without assuming some ultimate knowledge, specifica-
tion, or ‘determination’ of the objects involved. (Awodey, 2004, p. 3)

Awodey presents one reasonable methodological sense of ‘category-theoretic struc-
turalism’: a view about how to do mathematics that is guided by the features of cate-
gory theory.

Let us now contrast Awodey’s sense of structuralism with a position in the philos-
ophy of science known as Ontic Structural Realism (OSR). Roughly speaking, OSR
is the view that the ontology of the theory under consideration is given only by struc-
tures and not by objects (where ‘object’ is here being used in a metaphysical, and not
a purely mathematical sense). Indeed, some OSR-ers would claim that:

5



(Objectless) It is coherent to have an ontology of (physical) relations without admit-
ting an ontology of (physical) relata between which these relations hold.

On the face of it, the ‘simple structuralism’ that is evident in the practice of category
theory is very different from that envisaged by OSR; and in particular, it is hardly
obvious how this simple structuralism could be applied to yield (Objectless). On the
other hand, one might venture that applying (some form of) this simple structuralism
to physical theories will serve the purposes of OSR. Indeed, a recent proposal by Bain
(2013) purports to show that the ‘structural methods’ of category theory can be used
to defend (Objectless).3 Section 3 will develop and discuss Bain’s suggestion, and
Section 4 will clarify the category-theoretic methods on which it is based.

Before going on to evaluate this proposal, let us pause to consider which forms of
OSR have an interest in such an category-theoretic argument for (Objectless). Accord-
ing to Frigg and Votsis’ recent detailed taxonomy of structural realist positions (Frigg
and Votsis, 2011), the most radical form of OSR insists on an extensional (in the logical
sense of being ‘uninterpreted’) treatment of physical relations, i.e. physical relations
are nothing but relations defined as sets of ordered tuples on appropriate formal ob-
jects. This view is faced not only with the problem of defending (Objectless) but with
the further implausibility of implying that the concrete physical world is nothing but a
structured set (Frigg and Votsis, 2011, p. 261).

More plausible is a slightly weaker form of ontic structural realism, which Frigg
calls Eliminative OSR (EOSR). Like OSR, EOSR maintains that relations are onto-
logically fundamental, but unlike OSR, it allows for relations that have intensions.4

Defenders of EOSR have typically responded to the charge of (Objectless)’s incoher-
ence in various ways. For example, some claim that our ontology is ‘structure all the
way down’ without a fundamental level (Ladyman and Ross, 2007), or that the EOSR
position should be interpreted as reconceptualizing objects as bundles of relations, as
argued by Morganti (2004) and Esfeld and Lam (2011). However, Bain attempts to
provide a novel response, i.e. a category-theoretic defense of (Objectless). As we will
see, his defense implicitly draws on two forms of category-theoretic generalization,
viz. (GenDual) and (GenCat).

3 Bain’s two strategies
Bain does not explicitly schematize the premises of his argument; nor does he divide
his argument into two strategies for defending (Objectless). Thus, part of the work of
this paper is to reconstruct the details of his argument with an eye towards maximal
clarity and coherence.

3Although note that Bain himself does not claim to advocate OSR.
4See (Frigg and Votsis, 2011, p. 262) and (Ladyman and Ross, 2007) for more on the distinction between

OSR and EOSR. There are more sophisticated versions of EOSR—sometimes called ‘attenuated OSR’—
which allow for the existence of objects but deny their individuality; we shall not consider such doctrines
here. See e.g. (Frigg and Votsis, 2011, p. 263).

6



Bain (2013) takes as his starting point the assumption that physical objects, i.e. re-
lata, are represented by elements of structured sets5, and proceeds to develop two argu-
mentative strategies. Roughly, pp. 3-10 concern what we shall call the first strategy: it
discusses (i) how set-theoretic descriptions can in general be translated into category-
theoretic descriptions, and after noting the insufficiencies of this approach, it describes
(ii) a different and rather more specific sense of translation (as exemplified by the case
of Einstein algebras), which is claimed to show that relata are not ‘essential’ to the
articulation of physical structure.

On the other hand, pp. 11-12 presents a second strategy that takes it cue from exam-
ples of physical models whose very construction, it is claimed, is based on categories
that do not contain relata. We will discuss each strategy in turn.

3.1 A first strategy for defending (Objectless)
In what follows, it will be convenient to use the term ‘object’ in two senses. First, as
an object of a category, i.e. in a purely mathematical sense. We shall call this a C-
object (‘C’ for category-theoretic). Second, in the sense commonly used in structural
realist debates, and which was already introduced above, viz. an object is a physical
entity which is a relatum in physical relations. We shall call this an O-object (‘O’ for
‘ontological’).

We will also need to clarify our use of the term ‘element’. We use ‘element’ to
mean an element of a set, or as it is also often called, a ‘point’ of a set (indeed it will
be natural for us to switch to the language of points when discussing manifolds, i.e.
spacetimes, in Section 4.1). This familiar use of element should be distinguished from
the category-theoretic concepts of ‘global element’ and ‘generalized element’, which
we introduce below.

Bain’s first strategy for defending (Objectless) draws on the following idea: the
usual set-theoretic representations of O-objects and relations can be translated into
category-theoretic terms, whence these objects can be eliminated. In fact, the argument
can be seen as consisting of two different parts.

In the first part, Bain attempts to give a highly general argument, in the sense that
it turns only on the notion of universal properties and the translatability of statements
about certain mathematical representations (i.e. elements of sets) of O-objects into
statements about morphisms between C-objects. As Bain himself notes, the general
argument fails, and he thus introduces a more specific argument, which is what he
wishes to endorse. The specific argument turns on the idea of obtaining a transla-
tion scheme from a ‘categorical equivalence’ (we explain the precise notion below in
Section 4.1) between a geometric category and an algebraic category, which in turn al-
lows one to generalize the original C-objects. The argument is ‘specific’ because such
equivalences only hold between rather special sorts of categories.

5 In (Bain, 2013, p. 12), Bain notes that if either category theory or set theory were shown to be more
fundamental than the other, then this would bear on his argument for (Objectless). However, this issue is
controversial, and our investigation will concern the extent to which category theory can be used in aid of
(Objectless) without having to broach such issues within the foundations and philosophy of mathematics.
Hence our investigation follows Bain’s lead in terms of its scope.

7



3.1.1 The general argument

The details of Bain’s general argument can be reconstructed as follows:

G1: Physical objects and the structures they bear are typically identified
with the elements of a set X and relations on X respectively.

G2: The set-theoretic entities of G1 are to be represented in category-
theoretic language by considering the category whose objects are the rel-
evant structured sets, and whose morphisms are functions that preserve
‘structure’.6

G3: Set-theoretic statements about an object of a category (of the type in
G2) can often be expressed without making reference to the elements of
that object. For instance:

1. In any category with a terminal object7 any element of an object
X can be expressed as a morphism from the terminal object to X.
(So for instance, since the singleton {∗} is the terminal object in the
category Set, an element of a set X can be described by a morphism
{∗} → X.)

2. In a category with some universal property, this property can be de-
scribed purely in terms of morphisms, i.e. without making any refer-
ence to elements of an object.

To sum up, G1 links O-objects with a standard mathematical representation, viz.
elements of a set. And G2 and G3 are meant to establish the possibility that, in certain
cases, category theory allows us to translate statements about elements of sets into
statements about the structure of morphisms between C-objects.

Thus, Bain takes G1–G3 to suggest that Bain (2013, p. 4):

C: Category theory allows for the possibility of coherently describing
physical structures without making any reference to physical objects.

Indeed, Bain thinks the argument suggests that the mathematical representatives of O-
objects, i.e. the elements of sets, are surplus, and that category theory succeeds in
removing this surplus structure (Bain, 2013, p. 5).8

Bain himself thinks that the inference from G1–G3 to C fails, but he does give it
serious consideration, and it is easy to see why: its premises based on the most natural
and general translation scheme in category theory, viz. redescribing the properties of
C-objects in terms of morphisms, and indeed—if one is lucky—in terms of universal

6Exactly what Bain means by ‘structure’ here will be clarified below. In particular, we will see that it is
not the sense that mathematicians usually have in mind when they say that a map is ‘structure-preserving’.

7See Section 2.1 for the definition of a terminal object.
8Note that even if there is surplus structure here, it is not of the same kind as, e.g. gauge-equivalent

descriptions of fields in Yang-Mills theory. The latter has to do with various equivalent ways in which one
can describe the dynamical objects of a theory, viz. fields. By contrast, Bain’s strategy involves various
equivalent descriptions of the entire theory, as we will see below.

8



properties. It will be instructive for us to now discuss premises G1–G3 (which Bain en-
dorses) in some detail as this will serve not only to highlight some technical problems
with the premises, but also to draw out conceptual issues concerning Bain’s interpre-
tation of category-theoretic methods in physics, not least what he means by ‘structure’
(and its limitations).

First, the premise G1. Structural realist doctrines are typically formalized by mod-
eling O-objects as elements of a set and structures as relations on that set.9 However,
this is seldom the result of reasoned deliberation about whether standard set theory is
the best expressive resource from some class of such resources, but rather the product
of a deeply entrenched set-theoretic viewpoint within philosophy. Were philosophers
familiar with an alternative to set theory that was at least as powerful, e.g. category
theory, then O-objects and structures might well have been modeled directly in the
alternative formalism.10

So we maintain that there is no reason for the defender of O-objects to accept G1.
For instance, he might try to construct a category such that O-objects are modeled by
C-objects and structures are modeled by morphisms.11 Or he might take as his starting
point a non-concrete category, whose objects have no underlying set and thus cannot
be expressed in the terms of G1.

The premise G2, on the other hand, is ambiguous—it is unclear exactly how Bain
wants us to understand ‘structure’ and thus ‘structure-preserving maps’. First, note that
when mathematicians talk about ‘structure-preserving maps’ they usually have in mind
morphisms that do not preserve all the features of a C-object, but rather the character-
istic (albeit partial) features of that C-object. For instance, with respect to a group, a
structure-preserving map is a homomorphism and not an isomorphism. Bain’s example
of the category Set is of this type, because its morphisms are arbitrary functions (and
not bijective functions).

However, Bain wants to introduce a different notion of ‘structure’ that contrasts
with this standard usage, for he says:

(Structure) [T]he intuitions of the ontic structural realist may be preserved
by defining ‘structure’ in this context to be ‘object in a category’. (Bain,
2013, p. 3)

If we take this claim seriously, then a structure-preserving map will turn out to be an
isomorphism in the relevant category—for only isomorphisms preserve the complete
‘structural essence’ of a structured set.12 For instance, Bain’s example of the category
whose objects are smooth manifolds and whose morphisms are diffeomorphisms is of

9There are of course exceptions: e.g. French (2011) is prepared to consider category theory as a way of
formalizing structural realism if it can prove its worth.

10Of course, it is also a reasonable viewpoint to say that it is most ‘natural’ to do the philoso-
phy/foundations of physics in terms of set theory – what is ‘natural’ depends on how one conceives of
such foundational investigations.

11 For example, there are examples of categories whose C-objects might coincide with the mathematical
representatives of O-objects. For instance, in a path homotopy category, the C-objects are just points of the
relevant space (Brown, 2007), and one might in turn take the points of a space to be O-objects, as Bain does
in his example of general relativity and Einstein algebras.

12We thank Jonathan Bain for confirming (in private communication) that this is a fully spelled-out version
of what he takes structure to be in (Bain, 2013, p. 3).

9



this type. If this is really what Bain has in mind, then one inevitably ends up with a very
limited and dull class of categories. But even if one relaxes this notion of ‘structure’ to
mean ‘the structure that is preserved by the morphisms of the category, whatever they
happen to be’, one still runs into trouble with G3, as we will soon see.13

We now turn to the premise G3. First, note that G3(i) is false, as we now explain.
It will be convenient to introduce a piece of standard terminology: a morphism from a
terminal object to some object X is called a global element of X. And the question of
whether an element of X can be expressed as a global element in the relevant category
turns on the structure of the category in question. For instance, in the category Man
with smooth manifolds as objects and smooth maps as morphisms, this question re-
ceives a positive answer: global elements are indeed in bijective correspondence with
elements of a manifold.14 But in many other categories, e.g. the category Grp defined
above, the answer is negative. As an example, consider that Grp has the trivial group 1
as its terminal object and so a morphism from 1 to a group G only picks out its identity
and not its other elements. In order to obtain the other elements, one has to introduce
the notion of a generalized element of X, viz. a morphism from some ‘standard ob-
ject’ U into X. For instance, in Grp, one takes Z as the standard object U, and the
generalized elements Z → G allow us to recover the ordinary elements of a group G.

Second, while G3(ii) is certainly true, i.e. universal properties can be expressed
purely in terms of morphisms, it is a further—and significant—question for the scope
and applicability of this premise whether all (or even most) physical properties can be
articulated as universal properties.

Hence we have seen that the categorically-informed opponent of (Objectless) need
not accept these premises—there is a lot of room for debate about how exactly one
should use category theory to conceptualize the notion of physical structure. But sup-
posing that one does: is there a valid inference from G1–G3 to C? Bain himself notes
that the plausibility of this inference trades on an ambiguity in what one means by ‘ref-
erence’ in C. If one merely means that such constructions eliminate explicit but not
implicit reference to objects, then the argument is indeed valid. On the other hand, a de-
fense of (Objectless) requires the elimination of implicit reference to objects, and this is
what the general argument fails to offer—it merely provides a translation scheme from
statements involving elements (of sets) to statements involving morphisms between C-
objects (see also (Lam and Wuthrich, 2013) for an argument to this effect). So, the
defender of objects can maintain that one is still implicitly quantifying over elements.
It is to overcome this objection that Bain introduces his specific argument.

3.1.2 The specific argument

G3 above yields a special translation scheme that allows one to avoid making explicit
reference to elements. The key insight driving the specific argument is that, if one looks

13Note further that, when conceiving of structured sets as objects of a (partially-specified) category, one is
free to choose the morphisms (and thus fully specify the category) according to one’s purpose. For instance,
the categories Set and Rel have the same objects, but Rel has a much larger class of morphisms, which
contain the morphisms of Set (cf. Section 3.2.2).

14This is because the terminal object is the 0-dimensional manifold {0}, and so an element of a manifold
M is a morphism {0} → M.

10



T1 T2

T2∗

Figure 1: Schematic depiction of theory translation as described in S1–S4.

at a narrower range of cases, a rather different sort of translation scheme is possible:
indeed one that not only avoids making explicit reference to elements, but also allows
one to generalize the C-objects in such a way that these new C-objects can no longer
be considered to have elements (or as many elements) in the sense of the original ob-
jects. According to Bain (2013, p. 11), this shows that the ‘. . . correlates [of elements
of structured sets] are not essential to the articulation of the relevant structure’ (our
emphasis). Implicit reference to elements is thereby claimed to be eliminated.

Bain argues by appealing to a particular instance of how this translation is supposed
to work, viz. the example of Einstein algebras. By contrast, we will first clarify the
structure of the argument by providing its schema, and then proceed to show how
Bain’s example fits into this.

Here is our abstract reconstruction of Bain’s specific argument. The starting point
for the reconstructed argument is a category-theoretic version of the ‘semantic view of
theories’ on which a scientific theory is identified with its category of models—indeed
this will be the setup assumed in the following argument.15 Now, let there be two
theories T1 and T2, each represented by a category of models respectively. T1 is the
original physical theory that makes reference to O-objects.

S1: T1 can be translated into T2. In particular, each T1-model can be
translated into a T2 model and vice versa.

S2: T2 is contained in a strictly larger theory (i.e. a larger category of
models) T∗2 . In particular, T

∗
2 is constructed by generalizing T2-models to

yield models of T∗2 , typically by dropping an algebraic condition from the
T2-models. We will use T ′2 to denote the complement of T2 in T

∗
2 .

S3: T ′2 cannot be translated back into T1 and so its models do not contain
T1-objects.

S4: T ′2 is relevant for modeling some physical scenarios.

15Note the following two points. First, we are not using ‘model’ here in the strict sense of model theory,
but rather to mean a mathematical structure that represents a physical world that is possible according to
the theory. Second, this proposal is not to be confused with Lawvere’s category-theoretic formulation of
algebraic theories (Lawvere, 1963). In the latter, models are functors between categories, whereas in the
former, the models are just objects of some category (not necessarily a functor category) such as Top. Indeed,
Lawvere’s proposal is much more closely related to—though not the same as—the TQFT example that we
discuss in Section 3.2. We thank an anonymous referee for pushing us to clarify this point.

11



When taken together, S1–S4 are supposed to show that:

CS: The T1-object correlates in T2 do not play an essential role in articu-
lating the physical structure (smooth structure, in Bain’s specific case) of
T∗2 .

Bain’s conclusion is then that T1-object correlates can thereby be eliminated. Fig-
ure 1 depicts the schema of these premises. Let us defer for the moment the question
of exactly how the idea of ‘translation’ is supposed to work here (it is not discussed by
Bain but we shall take it up in Section 4.1). The key idea behind S1–S4 is that one can
generalize T2 to obtain a new—more general—theory T∗2 , some of whose models do
not contain T1-objects (i.e. O-objects in T1).

In Bain’s example, T1 is the category of geometric models of general relativity
(GR), and T2 is the category of Einstein algebra (EA) models of GR (Einstein algebras
were first introduced as models of GR in Geroch 1972; we shall provide the details
in Section 4.1).16 Bain is working with the idea that in geometric models of GR, the
relata, or O-objects, of GR are space-time points of the manifold.17

We now discuss the premises of the argument and show that S3 rests on a technical
misunderstanding; however, we will rehabilitate S3 before proceeding to argue that the
argument fails. First, S1: Bain notes that these space-time points are in 1-1 correspon-
dence with ‘maximal ideals’ (an algebraic feature) in the corresponding EA model. We
are thus provided with a translation scheme: points of space in a geometric description
of GR are translated into maximal ideals in an algebraic description of GR. So the idea
is that EA models capture the physical content of GR without making explicit reference
to points.

Now the version of S2 that Bain uses is one in which T2, the category of EAs,
gets generalized to T∗2 , the category of sheaves of EAs over a manifold, which has a
generalized notion of ‘smooth structure’. The former is a proper subcategory of the
latter, because a sheaf of EAs over a point is just equivalent to an EA.18

Bain then tries to obtain S3 by saying that a sheaf of EAs which is inequivalent to
an EA does not necessarily have global elements (i.e. sections of a sheaf) in the sense
previously defined, and so does not have points. Unfortunately, he confuses the notion
of a local section of a sheaf of EAs (which assigns an element of an EA to an open
subset of a manifold) with the notion of a maximal ideal of an EA (i.e. the algebraic
correlate of a spacetime point). And since the two are entirely different, a lack of global
sections does not imply a lack of spacetime points (i.e. O-objects). Therefore S3 needs
to be repaired.

Nonetheless, we can easily rehabilitate S3 is the following manner. The key idea is
that while T1 (a geometric model of GR) and T2 (the equivalent EA model) both make
reference to T1-objects (explicitly and implicitly, respectively), some sheaves of EAs

16As we shall explain later, T1 and T2 are only equivalent here if we restrict the latter to the class of what
we shall call Geroch representations of EAs.

17Note that as before (cf. G1 above), the friend of O-objects can challenge the idea that space-time points
should be taken to be the representatives of O-objects in geometric models of GR.

18As we shall see later, the example of sheaves of algebraic structures is but one of a whole class of
examples that could fit into this argument schema.

12



do not refer to T1-objects because they have no formulation in terms of geometric mod-
els of GR. In other words, the generalized smooth structure of T ′2 cannot be described
in terms of the structured sets used to define ordinary smooth structure in the case of
T1 and T2.

Finally, as regards S4, various authors have taken the utility of T ′2 to be e.g. the in-
clusion of singularities in space-time (Heller, 1992), and as a step towards formulating
quantum gravity (Geroch, 1972). We shall return to this point in Section 4.1.19

We now turn to considering the inference to CS. It is not entirely clear what Bain
means by ‘[the relata] do not play an essential role’ (Bain, 2013, p. 9)—nor does he
expand on this phrase—but the most straightforward reading is that T1-objects are elim-
inated simpliciter from T∗2 .

One might compare this situation to the way that the collection of all groups (anal-
ogous to T2) is contained in the collection of all monoids (analogous to T∗2 ): it might
be claimed that inverses are eliminated from the collection of all monoids. One could
of course speak in this way, but what this would mean is that some monoids (in par-
ticular, groups) have inverses, and some do not—a ‘monoid’ is just a general term that
covers both cases. Similarly, we can see that CS does not follow from S1–S3, since
T∗2 contains some models that (implicitly) quantify over T1-objects, viz. the models of
T2, and some that do not, viz. the models of T ′2.

20

We have seen that the specific argument will not work if one is concerned with
eliminating reference to T1-objects from the new and more general theory T∗2 . How-
ever, what if one is concerned not with eliminating reference, but rather with down-
grading the role that T1-objects play in T∗2 , e.g. by claiming that the models of T

′
2

have a conceptual or metaphysical priority?21 And what would such a ‘downgrading’
even amount to? Section 4.1 will return to these questions after a discussion of what
‘translation’ really means in the context of S1.

3.2 A second strategy for defending (Objectless)
The second strategy, as laid out in Section 4 of Bain (2013, pp. 10–11), draws on ex-
amples of physical theories whose formulation is supposed to be essentially category-
theoretic, viz. the toy models provided by nCob and Hilb, and the use of these two
categories to construct a Topological Quantum Field Theory (TQFT).

As before, the aim of this argument is to show that, by using category theory, physi-
cal theories can dispense with reference to physical objects, i.e. relata (which we called
O-objects at the start of Section 3.1). But whereas the previous strategy is supposed
to show how one can discard O-objects from the definition of a physical theory that is
putatively formulated in terms of O-objects, this argument is more direct—it tells us
that some category-theoretic formulations of physics (viz. nCob, Hilb and TQFTs)
are so general that they make no reference to O-objects to begin with, and this in turn
yields (Objectless).

19Bain (2013) does not discuss S4, but he does take up this matter in (Bain, 2003).
20Actually, the more precise statement is that the algebraic models of T ′2 will not in general contain

‘enough’ points to correspond to the geometric models of T1.
21Indeed, Bain has (in private communication) clarified that this is how he would develop his argument in

response to the above analysis.

13



Here, as before, one does not need to agree with Bain’s definition of an O-object.
But even granting this definition, we shall see that the strategy is problematic. Our
discussion will begin by recalling the key ideas underlying nCob, Hilb and TQFTs
in Section 3.2.1. In Section 3.2.2 we distinguish the two different senses in which Bain
takes nCob, Hilb, and TQFTs to be more general than physics based on Set, after
which we proceed to his claim about (Objectless).

3.2.1 Purely category-theoretic physics?

Bain’s examples in support of the second strategy are: (i) the category Hilb of com-
plex Hilbert spaces and linear maps; and (ii) the category nCob, which has (n − 1)-
dimensional oriented closed manifolds as objects, and n-dimensional oriented mani-
folds as morphisms. These examples purportedly represent ‘purely’ category-theoretic
physics. This means that formal statements about the physical theory, e.g. quantum
mechanics using Hilb, are derived using the category-theoretic rules of morphisms in
Hilb.

Now, prima facie, both of these examples look like good candidates for doing
purely category-theoretic physics. First, each category is potentially useful for studying
the properties of quantum theory and general relativity respectively (we will elaborate
further on their utility in Section 4.2). Second, each possesses categorical properties
which are promising for describing physical properties. More ambitiously, they sug-
gest that one could use categorical tools to develop an approach for integrating parts of
quantum theory and general relativity.

Let us pause to explain this second point, which rests on the fact that, qua cate-
gories, Hilb and nCob share some important properties. For example, both of these
categories are monoidal, meaning that both categories carry a generalization of the ten-
sor product V ⊗ W of vector spaces V and W .22 In nCob the monoidal structure is
given by the disjoint union of manifolds; whereas in Hilb, the monoidal structure is
given by the usual linear-algebraic tensor product of Hilbert spaces.

A second formal property shared by both categories is that they each possess a
contravariant involutive endofunctor (·)† called the dagger functor.23 (This is also
sometimes called a ‘∗-operation’, as in e.g. Müger 2008.) This means that, given a
cobordism f : A → B in nCob or a linear map L : A → B in Hilb, there exists a
cobordism f† : B → A and a linear adjoint L† : B → A respectively, satisfying the
involution laws f† ◦ f = 1A and f ◦ f† = 1B, and identically for L.

The formal analogy between Hilb and nCob has led to the definition of a type
of quantum field theory, known as a topological quantum field theory (TQFT), first
introduced by (Atiyah, 1988) and (Witten, 1988). A TQFT is a (symmetric monoidal)
functor:

T : nCob −→ Hilb,
and the conditions placed on this functor, e.g. that it preserve monoidal structure, re-
flect that its domain and target categories share formal categorical properties. To fur-

22 See (Mac Lane, 2000, Ch. 7) for an exposition of monoidal categories.
23 Recall that a contravariant functor is a functor F : C → D that reverses the direction of arrows, i.e. a

morphism f : A → B is mapped to a morphism F(f) : F(B) → F(A). Also recall that an endofunctor
on a category C is a functor F : C → C, i.e. the domain and codomain of F are equal.

14



ther flesh out the physical interpretation of TQFTs, we note that the justification for the
term ‘quantum field theory’ arises from the fact that a TQFT assigns a state space (i.e. a
Hilbert space) to each closed manifold in nCob, and it assigns a linear map represent-
ing evolution to each cobordism. This can be thought of as assigning an amplitude to
each cobordism, and hence we obtain something like a quantum field theory.

3.2.2 O-objects in category-theoretic physics?

Recall that the significance of these examples for Bain is their apparent status as purely
category-theoretic formulations of physics which, in virtue of their generality, do not
make any reference to O-objects (represented in the standard way, i.e. as elements of
sets). We now turn to a criticism of this claim.

Bain’s key idea seems to be that this ‘generality’ consists of the fact that nCob
and Hilb (and thus TQFTs) have very different properties (qua categories) from Set.
In fact, he claims that three such differences count in favor of (Objectless):

(i) nCob and Hilb are non-concrete categories, but Set (and other categories
based on it) are concrete.24

(ii) nCob and Hilb are monoidal categories, but Set is not.

(iii) nCob and Hilb have a dagger functor, but Set does not.

We address these points and their implications for (Objectless) in turn.
First, (i). Bain wants to argue that since nCob and Hilb ‘cannot be considered

categories of structured sets’ (Bain, 2013, p. 10), nor can these categories be interpreted
as having O-objects. If one is talking about categorical properties (as in Section 4 of
(Bain, 2013)), this claim is best couched in the standard terminology that we introduced
in Section 2.1, viz. as the claim that these are not concrete categories.

But this inference is faulty for two reasons. First, his point about non-concreteness
is not altogether accurate, i.e. point (i) is false as stated. On the one hand, it is true that
nCob is not a concrete category: in particular, while the objects of nCob are struc-
tured sets, its morphisms are not functions, but manifolds, i.e. sets equipped with the
structure of a manifold. But on the other hand, Hilb is certainly a concrete category,
since the objects are Hilbert spaces, which are sets with extra conditions; and the mor-
phisms are just functions with linearity conditions. In other words, the morphisms are
structure-preserving functions. Thus, Bain’s examples of category-theoretic physics
are based in part on concrete categories.

Second and more importantly, it is doubtful that the standard mathematical no-
tion of concreteness will aid Bain in defending (Objectless). Bain wants to hold that
the non-concreteness of a category is a sufficient condition for its not referring to O-
objects. But nCob is an example of a non-concrete category that apparently contain
O-objects—indeed the same O-objects (viz. space-time points) that Bain takes to be

24Bain does not himself use the language of ‘non-concrete’ category but this is the most reasonable—and
indeed the most precise—interpretation of what he means in (Bain, 2013, p. 10), since he is discussing
categorical properties. At any rate, the point we make below applies equally well to his terminology of
‘non-structured set’ without any further categorical gloss.

15



present in geometric models of GR, (cf. Section 3.1.1). We thus see that, by Bain’s
own lights, non-concreteness cannot be a sufficient condition of evading O-objects.25

So the example of nCob still has C-objects that are based on sets, albeit mor-
phisms which are more general than functions. However, one can go further than this:
the notion of a category is in fact defined in a schematic way (cf. Section 2.1), which
leaves open the question of whether C-objects are sets or whether functions are mor-
phisms. One might thus rhetorically ask whether this could this be the full version of
‘categorical generality’ that Bain needs in order to defend (Objectless). In fact, this is
implausible, because of the way in which such a schematic generality ends up being
deployed in physics. As we shall see in Section 4.2, the point of schematic generality
is not so that we can do physics without sets, but rather so that we can determine which
categories of sets do (or do not) share important physical properties.

On to (ii): unfortunately, this claim is straightforwardly false: the category Set
is certainly monoidal, with the monoidal product being given by the cartesian product
(e.g. see Mac Lane 2000, p. 161).

Finally, (iii). While it is true that Set does not have a dagger functor, and nCob
and Hilb do, it is easy to construct an example of a category with a dagger functor, but
which Bain would presumably agree has O-objects. Consider the category C with one
object, namely a manifold M representing a relativistic spacetime; the morphisms of
C are taken to be the automorphisms of M. As with nCob, this category has natural
candidates for O-objects (as Bain assumes), viz. the points of the manifold. But the
category C also has a dagger functor: given an automorphism f : M → M, the
morphism f† : M → M is given by the inverse automorphism f−1. In contrast, the
category Set does not have a dagger functor: this follows from the observation that for
any set A that is not the singleton set {∗}, there is a unique morphism f : A → {∗},
but the number of morphisms g : {∗} → A is just the cardinality |A| > 1. Hence there
does not exist a bijection between the set of morphisms {f : A → {∗}} and the set of
morphisms {g : {∗} → A}, which implies that there does not exist a dagger functor
on Set. Thus, by Bain’s own criterion, it is reasonable to consider C to be structurally
dissimilar to Set, despite the fact that it has O-objects.

More generally, i.e. putting aside the issue of (Objectless), it is quite unclear how
one should interpret the physical significance of the fact that nCob/Hilb, but not
Set has a dagger functor. For instance, it turns out that by an easy extension of Set,
one can construct a category that does have a dagger functor. This easy extension is
the category Rel, whose objects are sets and whose morphisms are relations between
objects (i.e. subsets of the Cartesian product of a pair of objects). Note first that Set is
a subcategory of Rel because Set and Rel have same objects, and every morphism in
Set is a morphism in Rel. This can be seen by noting that every function f : A → B
can be written as a relation f ⊆ A × B, consisting of the pairs (a,b) defined by
f(a) = b. Second, note that—unlike Set—Rel does have a non-trivial involution
endofunctor, i.e. a dagger functor, since given a relation R : A → B, the relation

25One could of course try to introduce the machinery of the first strategy (discussed in Section 3.1.2) to
evade O-objects in this context, but we have already discussed the problems with this move. In any case,
notice that the first strategy is not concerned with concreteness: translation and generalization by duality will
typically yield yet another concrete category (as it does in the case of Bain’s focal example, i.e. the category
of Einstein algebras, which is concrete since it is defined as a structured set).

16



R† : B → A is just defined by bR†a if and only if aRb.
By extending Set to Rel, we have obtained a category that is akin to nCob/Hilb

in having a dagger functor, but which also appears to be at least as good a candidate
as Set for codifying the standard notion of physical structure, since its morphisms are
n-ary relations and they can be used to encode structure. Evidently, if Bain’s sugges-
tion about the physical difference between these categorical properties is to amount to
anything, then one must seek the right physical interpretation of such categories—a
task that we shall take up in Section 4.2.

We have seen that (i)–(iii) will not help Bain defend (Objectless). But one should
still ask: what is the notion of generality that lies behind these three points? Although
Bain does not state it explicitly, we can extract it from the details of his argument and
summarize it as follows:

The second strategy: a moral. Category theory offers us the resources to predi-
cate properties (‘categorical properties’) of a category of C-objects without further
specifying the properties of the C-objects, or indeed what types of objects these are
(e.g. whether they are groups, manifolds, lattices, and so on). And such predica-
tions/properties are general in the sense that they are instanced by categories of specific
C-objects, which provide instances of such categorical properties.

We call this second form of generalization (GenCat), and we will explore its interpreta-
tion and use in physics (in particular the physical examples invoked by Bain) in Section
4.2.

4 Two forms of categorical generalization

4.1 Generalization by duality
Earlier in Section 3.1.2, we saw that Bain invoked the notion of ‘translation’ from ge-
ometric models of GR to EA models of GR without much explanation of its details.
As we will now see in Section 4.1.1, this idea of translation is based on the notion
of an ‘equivalence’ of categories (in contrast with the example of translation in Sec-
tion 3.1.1, which involved working within a category and re-describing elements of
C-objects in terms of morphisms). We then explain the framework for this form of
generalization—which we shall call generalization by duality (GenDual)—which is
based on such equivalences.

In Section 4.1.2, we discuss the duality between geometry and algebra and then
revisit Bain’s EA example, which is based on such a duality. Among other things,
we shall consider whether (GenDual) provides a metaphysical downgrading of T1-
objects—a possibility that was first raised at the end of Section 3.1.2.

17



4.1.1 The framework for (GenDual)

The relevant notion of translation is given by the notion of an equivalence between two
categories—in particular, a special kind of equivalence called a duality.26 Equivalence
allows us to say that two categories (and thus two theories, if we are considering cat-
egories of models) are ‘structurally the same’—that all the relevant content (objects,
structures, operations, etc.) of each category can be described in terms of the other
category. We will describe this notion of equivalence, explain how it underpins (Gen-
Dual), and then provide a way of spelling out the relationship between T1, T2, and T∗2
in the specific argument, as discussed at the end of Section 3.1.2.

What does it mean to say that two mathematical objects are ‘structurally the same’?
Earlier, we discussed the idea of isomorphism, which is a way of saying that two math-
ematical objects in the same category are structurally the same. But what if the mathe-
matical objects that we wish to discuss are themselves categories? (This is the relevant
scenario when we are discussing an equivalence between two theories, whose models
will generally have different types of mathematical structures respectively, and thus do
not live in the same category.) A natural way in which to do so is by viewing categories
themselves as the C-objects of a category: the category of categories, whose objects
are categories and whose morphisms are functors (as defined in Section 2.1).

Hence, one might naively think that the correct way in which to articulate the struc-
tural sameness of two categories/theories is by means of the notion of an isomorphism
in the category of categories.27 However, this definition is too strict, as we can see from
a simple example. Suppose that T1 is a category with one object with only an iden-
tity morphism and no other morphisms, and that T2 is a category with two isomorphic
objects and two identity morphisms, and no other morphisms. Evidently, these two
categories are not isomorphic. However, this conflicts with our intuition that T1 and
T2 should really be counted as structurally ‘the same’, because structurally speaking
(i.e. up to isomorphism within a category) they contain just one object and one identity
morphism.

Fortunately, there is a standard generalization of the notion of ‘isomorphism of
categories’ that respects this intuition:

(Equivalence) Two categories T1 and T2 are equivalent iff there exist
functors F : T1 → T2 and G : T2 → T1 such that G ◦ F ∼= 1T1 and
F ◦ G ∼= 1T2 .

where ∼= denotes a natural isomorphism (i.e. a natural transformation—in the sense of
Section 2.1— that is an isomorphism in the category of functors). The pair of functors
F and G are called ‘quasi-inverses’. (Notice that this gives the correct result for the

26Note that the term ‘duality’ is used in a much looser way in the physics literature, i.e. to encompass a
much more complex and profound range of examples, e.g. AdS/CFT and S-duality (Schwarz, 1997). Al-
though ‘duality’ still refers to theoretical equivalence of some kind in these examples, the nature of these
equivalences is often rich and subtle, and defies formalization. On the other hand, while the category-
theoretic sense of duality that we will consider is somewhat more limited in scope, it offers a precise un-
derstanding of what ‘theoretical equivalence’ amounts to, and may in time suggest a way of understanding
some of the more complex dualities that occur in physics.

27We repeat the definition for the reader’s convenience: two categories T1 and T2 are isomorphic iff there
exist functors F : T1 → T2 and G : T2 → T1 such that G ◦ F = 1T1 and F ◦ G = 1T2 .

18



above ‘simple example’, since the two objects in T2 are isomorphic, but not equal.) A
duality is an equivalence in which F and G reverse the direction of the morphisms in
their respective domain categories.28

Two equivalent categories may look rather similar, as in the simple example above.
But equivalent categories can also contain objects that look very different from each
other. This is often the case when, broadly speaking, one of the categories is algebraic
and the other is geometric.29

We now have sufficient resources to make sense of mathematicians’ and physicists’
informal talk of ‘translation’ in this context. Roughly speaking, if two equivalent theo-
ries are sufficiently different in the structure of their objects, then it is typical to apply
the metaphor of a ‘language’ to each category and say that one is ‘translating’ notions
from one category to the other by means of the quasi-inverse functors F and G. In
other words, what translation amounts to is that up to isomorphism, one can construct
any object in T1 from an object in T2 (and vice versa), and that the relational structure
(i.e. the structure of the morphisms between objects) is the same in each category.30

What is the significance of translation? First, it offers us two different perspectives
on what is in some sense the same theory. For instance, it may be easier to prove a
statement about an algebraic object than it is to prove the equivalent statement about its
geometric correlate. But it also builds on this to allow for a more profound possibility,
viz. a way of generalizing a physical theory, which we call generalization by duality
(GenDual).

In order to discuss this profound form of generalization, we will first have to intro-
duce a mundane form of generalization, familiar from elementary mathematics. Con-
sider the following way of constructing a genus ( in this case, a category) T∗ of which
the category T is a species: we define the objects of T∗ by dropping some condition
from the objects of T . So for instance, if T is a category whose objects are commuta-
tive algebras, then T∗ could be the category of algebras, which of course contains T as
a subcategory.31

The profound generalization (GenDual) is achieved when one applies the following
procedure:

1. Equivalence: Let T1 and T2 be equivalent (perhaps dual) theories.

28The important difference between (Equivalence) and isomorphism is this: the latter uses the notion
of equality to characterize G and F as inverses, whereas the former ‘relaxes’ this to the notion of (natural)
isomorphism. This ‘relaxation’ is an elementary case of what mathematicians call categorification: a process
through which one lifts sets to categories, functions to functors, and equality to isomorphism. We do not have
room to discuss categorification in detail here, but suffice it to say that one can continue playing the same
game for higher-order morphisms (i.e. morphisms between morphisms, and so on).

29It is also true of many cases of theoretical equivalence in physics, where the models of the different
theories often have a very different mathematical structure.

30The sense of translation we have just been discussing is implemented via the quasi-inverse functors, and
so applies to the objects and morphisms of each category in the pair. However, theorists also sometimes
speak of translating between the bits of structure of two objects, e.g. in Bain’s Einstein algebra example
– to be developed in Section 4.1.2 – a space-time point of a manifold gets translated into a maximal ideal
of an Einstein algebra. This second sense of translation depends on the construction of a particular functor
between two categories, as we will see in the below examples.

31Notice the strength of the species-genus metaphor here: the species T can be defined as the conjunction
of the genus T ∗ and some differentia (i.e. the commutativity of the algebra, in this case).

19



2. Mundane Generalization: Let T2 be generalized to a larger theory T∗2 by drop-
ping some differentia from the definition of the objects of T2.

3. Analogy: By combining Equivalence and Mundane Generalization, one finds
that there is a sense in which T ′2-objects (i.e. the complement of T2 in T

∗
2 ) can

be thought of as generalized versions of T1-objects.This sense will not be literal,
since T ′2-objects cannot be translated (via equivalence) into T1-objects. However,
Equivalence allows us to obtain a characterization of T1-objects in terms of T2,
and if the essential features of this characterization are preserved in the objects
of T ′2, then one can call the latter ‘T1-objects’ by analogy.

An example of this procedure (which we discuss at greater length below) is where
T1 is a category of geometries and T2 is a category of algebras. Mathematicians typi-
cally deform T2 to a category T2∗ which relaxes some condition on the objects of T2
(e.g. the existence of sufficient points) but which retains algebraic analogs of geometric
features (e.g. symmetries, invariants, open sets, measures) and operations (integration
and differentiation) of T1. The objects of T ′2 are then called ‘generalized geometries’
by Analogy.32

Returning now to the main topic of this section, viz. (GenDual), we can now use the
idea of core-dependent homonymy (Shields, 1999) to explain the mathematical practice
of describing T ′2-objects as T1-objects even when they are evidently not the same thing:
generalization by duality establishes T1 (which is equivalent to T2) as the primary case
of things called T∗2 , and so T

′
2 is related to T2 as a secondary case is to a primary case.

In other words, T2 is not just a species of T∗2 in this scenario, but rather an ‘essential’
or ‘core’ species of T∗2 .

It follows that the form of generality given by (GenDual) is such that one cannot
define T2 as a genus (e.g. T∗2 ) in conjunction with some differentia—this would leave
out the fact that T2 is the essential species of the genus, and that all the other species of
T∗2 are to be interpreted with respect to T2.

An example will serve to make this reasoning more accessible: it will be the in-
stance of algebraic-geometric duality that is implicitly appealed to in Bain’s specific
argument. We now turn to discussing this duality.

4.1.2 Algebraic-geometric duality and Einstein algebras

The duality most commonly invoked in order to accomplish (GenDual) is called ‘algebraic-
geometric duality’, where ‘algebra’ is understood as algebraic structures induced by
classes of functions on a space, and ‘geometry’ is understood broadly to include topo-
logical spaces and various kinds of geometric structures that can be added to a topo-
logical space (e.g. vector bundles with connection, spin structures, geometric operators
acting on the space, etc.).

32What are the essential (algebraic correlates of geometric) features whose retention in T ′2 suffices for
T ′2-objects to be analogically called ‘geometries’? We do not pretend to give a metaphysical answer to this
question. Rather, our answer comes from the practice of mathematics: mathematicians take certain features
(e.g. open sets satisfying certain axioms) as being definitive of certain objects (e.g. a topology) and if one
can hang on to the algebraic correlates of these notions while eliminating the correlates of other features (e.g.
points) then one has preserved the ‘essential features’.

20



One of the simplest examples of algebraic-geometric duality is the duality between
T1, the (geometric) category of locally compact Hausdorff topological spaces, and T2,
the category of commutative C∗-algebras.33

The quasi-inverse (and contravariant) functors F : T1 → T2 and G : T2 → T1
are easy to describe. Roughly speaking, F takes a topological space to the commuta-
tive C∗-algebra of complex-valued continuous functions on that space. On the other
hand, G (also called the ‘functor of points’) takes each algebra to the set of characters
(i.e. homomorphisms from the algebra to C) of the algebra. These characters are then
thought of as points of a topological space, and when they are suitably topologized, one
recovers the entire structure of the relevant topological space.34 It is then a straightfor-
ward matter to check that T1 is equivalent to T2, i.e. that F and G satisfy the natural
isomorphisms:

GF ∼= 1T1 FG ∼= 1T2. (1)

What are the implications of this duality for translating concepts between T1 and
T2? First, two algebras in T2 are isomorphic (as algebras) if and only if their spaces of
characters (given by action of G) are homeomorphic (which is what the notion of iso-
morphism amounts to in T1), and vice versa.35 Furthermore, there is a correspondence
between the topological properties of a space in T1 and the properties of an algebra in
T2. For instance, compact spaces correspond to unital C∗-algebras, an open subset of
a space corresponds to an ideal of an algebra, a closed subset of a space corresponds
to a quotient algebra, a measure on a space corresponds to a positive functional on an
algebra, and so on. Thus topological and algebraic ‘correlates’ can be translated back
and forth between T1 and T2.

The next step of (GenDual), viz. Mundane Generalization can be implemented by
extending T2 to the category T∗2 of all C

∗-algebras. In particular, the complement T ′2
is the category of non-commutative C∗-algebras.

The last step of (GenDual), viz. Analogy, turns on a happy but contingent fact
about this combination of Equivalence and Mundane Generalization, viz. that the al-
gebraic objects in the complement T ′2 still have most of the algebraic properties that,
when possessed by objects of T2, corresponded to topological properties of spaces
via equivalence. Indeed, there is only one significant property that separates T ′2 from
T2: dropping the commutativity condition from C∗-algebras makes it the case that
these algebras no longer have a rich set of characters (i.e. the correlates of points of a
topological space); in particular they will no longer correspond to objects in the cate-
gory of topological spaces. Nonetheless, because the other algebraic correlates in T2
of topological concepts in T1 still make sense in T ′2, mathematicians exploit this anal-
ogy between T2 and T ′2 and deem the objects of T

′
2 to be non-commutative ‘topological

spaces’. This practice is reminiscent of the phenomenon of core-dependent homonymy
(Shields, 1999), where there can be e.g. two non-univocal terms—indeed homonyms—
one of which provides the core sense of the term, whereas the other term has a meaning

33T1 has continuous proper maps as morphisms, and T2 has proper ∗-homomorphisms as morphisms.
34Here we describe only the action of the functors on objects; both F and G act on morphisms by pull-

backs, see e.g. Section 1 of (Khalkali, 2009).
35Similarly, the group of automorphisms of an algebra in T1 is isomorphic to the group of homeomor-

phisms of the character space of that algebra.

21



that is derived from the core sense. So, in the present example of (GenDual), the core
sense of (locally compact Hausdorff) topological spaces is provided by the category
T1 and its equivalent category of algebras T2, and the non-commutative ‘topological
spaces’ of T ′2 derive their meaning from this equivalence.

36

Now that we have a better understanding of the foundation of Bain’s first strategy,
viz. (GenDual), we are ready to revisit his Einstein algebra (EA) example from the
perspective of algebraic-geometric duality. This will also provide us with the occasion
to analyse the utility of (GenDual) within a physical context.

In his paper, Bain refers primarily to Heller’s (1992) formulation of EAs. But in
order to understand the physical interpretation of (GenDual), it will be helpful to start
by considering Geroch’s motivations when he first introduced EAs in (Geroch, 1972).
Geroch’s original motivation was to find a description of GR that might be useful for
extending it to describe quantum gravity. In particular, Geroch wanted to accommodate
the following plausible intuition about a quantum theory of gravity:

[I]t is perhaps reasonable to expect that, in a quantum theory of gravitation,
the mathematical formalism will, at some point, suggest a ‘smearing out
of events’. (Geroch, 1972, p. 1)

Thus, he sought a description of models of GR that would eliminate explicit reference
to space-time events, i.e. points of the space-time manifold. In order to do so, he
exploited a familiar fact from differential geometry: the vector fields on a manifold can
be characterized in purely algebraic terms, i.e. as the set of derivations acting on the
algebra of smooth real-valued functions.

Geroch developed this familiar fact into a purely algebraic analogue of the theory
of tensors, and proceeded to introduce algebraic versions of the metric, the covariant
derivative, and various tensor fields—thus arriving at an algebraic model of Einstein’s
equations and its solutions, which he called Einstein algebras (EAs).37

Geroch showed that every geometric model of GR gives rise to an EA. In other
words, he defined a functor F from T1, the category of geometric models of GR, to a
subcategory of the category of all EA models. However, he did not try to formulate an
equivalence between T1 and the category of EA models.38 Indeed, Geroch (1972) does
not use the language of category theory at all.

A more sophisticated development of EAs was then accomplished by Heller (1992),
who showed that T1 is equivalent, in fact dual, to T2, the category of (what Heller
called) ‘Geroch representations’ of EAs. Similarly to our previous example of the
duality between topological spaces and commutative C∗-algebras, Heller introduced a
functor of points G in order to build a space from an Einstein algebra and to play the
role of a quasi-inverse with respect to the functor F . Indeed, his functor provides even

36More sophisticated equivalences of algebraic-geometric duality abound, and they extend various geo-
metrical concepts in a similar way, e.g. integration, vector fields, metrics, topological invariants, and sym-
metries of such spaces, to name just a few examples. Many of these dualities also have physical applications,
e.g. Alain Connes’ use of non-commutative geometry to attempt to describe various quantum field theories
and the Higgs mechanism (Connes and Marcolli, 2008).

37Roughly speaking, Geroch defined an EA as an algebra F , with a subring R that is isomorphic to the
real numbers, such that we can define an F -module (corresponding to the set of vector fields) which satisfies
all the necessary algebraic conditions for it to count as a model of GR.

38We thank Robert Geroch for clarifying this with us in private communication.

22



more information because it associates to such a space a class of functions by means
of a ‘functional representation’ of the EA.

It will now be helpful to elucidate Heller’s work by means of our explanation of
(GenDual) above. First, he showed that there was an Equivalence (indeed a duality)
between T1, the category of geometric models of GR, and T2, the category of Geroch
representations of EAs, whose class of functions is isomorphic to C∞(M), i.e. the set
of smooth functions on a smooth manifold M. Second, by Mundane Generalization,
he extended T2 to the category T∗2 of general EA representations. The important thing
about this extension is that T ′2 contains objects that need not be smooth manifolds,
and whose class of functions is not isomorphic to C∞(M), but are nonetheless still
equipped with an algebraic notion of ‘global smooth structure’ which only coincides
with geometric notion of smoothness (associated with C∞(M)) when one restricts to
the sub-category of Geroch representations, i.e. T2. Thus, the objects of T ′2 cannot be
translated into the geometric objects of T1, whose notion of smoothness is inherently
bound up with the idea of local functions. Nonetheless, since the objects of T ′2 still have
an algebraic notion of smoothness and all the algebraic operations on ‘algebraically
smooth’ objects that are necessary for formulating GR, we can—along with Heller—
use Analogy to consider these objects to be ‘generalized models of GR’.

Note that Heller’s motivation for extending T2 to T∗2 was not in order to describe
quantum gravity (at least in the first instance) but rather in order to describe singu-
larities as part of the structure of EA models of GR. By contrast, in traditional ge-
ometric GR, singularities are defined by the inextendibility of causal curves in the
space-time, and as such are not part of the intrinsic structure of space-time. Ordinary
smooth n-manifolds have trouble accommodating singularities precisely because of
the point-like nature of how their ‘smoothness’ is defined. Heller’s extended category
T∗2 overcomes this problem because the objects of T

′
2 have no such local smoothness

condition. (Heller, 1992) then went on to generalize this construction further, by in-
troducing sheaves of Einstein algebras over a differential space, which allows for the
inclusion of an even wider class of singularities than that described by T ′2—we shall
however leave this aside, as the present sketch is sufficient both to illustrate a physical
application of (GenDual), as well as to further discuss Bain’s argument in light of it.

Now that we have gained more perspective on the application of (GenDual) to
Bain’s EA example, we can revisit the issue that we left open at the end of Section 3.1,
viz. whether—instead of claiming that (GenDual) for EAs eliminates T1-objects—one
can claim that it merely downgrades the metaphysical status of T1-objects, because it
gives metaphysical priority to the models of T ′2.

Note that, at least from the purely mathematical point of view, the above analysis
suggests that it is the models of T1/T2 that have conceptual priority (and the models
of T ′2 that are derivatively called ‘generalized models of T1). Nonetheless, one might
try to deploy the following analogy in a physical context. Perhaps one should think of
the new (generalized) models in T ′2 as standing in relation to T2 in the way that quan-
tum mechanics—when formalized in terms of non-commutative C∗-algebras—stands
in relation to classical mechanics, which can be formalized in terms of commutative
C∗-algebras.39 In this analogy, T1 is the classical phase space and T2 is its dual al-

39See e.g. Rieffel’s (1994) seminar paper on deformation quantization for a review of how we can con-

23



gebra of functions on that phase space, i.e. a commutative C∗-algebra. One can then
extend T2 to T∗2 via deformation quantization (Rieffel, 1994), and this new category
will include the non-commutative C∗-algebras which are required for quantum me-
chanics. The complement of T2, again denoted T ′2, can be thought of as playing the
role of a generalized ‘non-commutative’ phase space. Thus, one might try to say that
since quantum mechanics is physically prior to classical mechanics, the O-objects of
classical phase space (whatever they are) are ‘ontologically downgraded’ in T∗2 , even
if they are not eliminated.

Unfortunately, this analogy is unsatisfactory for at least two reasons. First, while
the non-commutative deformation of classical phase space constitutes a complete and
adequate successor theory, T ′2 for EAs does not. Recalling Geroch and Heller’s moti-
vations for constructing an EA description of GR (which we have just discussed) will
highlight this point. From Geroch’s perspective, viz. trying to construct a theory of
quantum gravity, T ′2 is merely a suggestive template (in the sense that it does include
point-events) with which one might begin to construct a theory of quantum gravity.
From Heller’s perspective, T ′2 is just an extension of the models of T2 that is per-
haps useful in gaining physical insight into singularities (it is at the very least unclear
whether it really provides additional insight, since there are standard ways of dealing
with singularities even with respect to geometric models of GR). Thus, it seems clear
that (GenDual) leaves open the question of how the models of T ′2 are to be interpreted
with respect to those of T1/T2.

Second, in the case of EAs, it is implausible to view T2 as being recovered from
T∗2 in any physical sense. This is because the models of T2 are ‘recovered’ from T

∗
2 by

simply adding some algebraic condition to T∗2 models (e.g. the defining condition of
the Geroch representation of EAs). By contrast, the commutative algebras of classical
mechanics are recovered from the non-commutative algebras of quantum mechanics
by taking the limit of a physical parameter ℏ → 0. The point here is that, even if
one could attribute physical significance to the defining algebraic conditions of T ′2 and
T2 (e.g. spacetimes that do and do not include an intrinsic description of singularities
respectively) one still needs to be able to give a robust account of why T ′2 should be
viewed as the physical core of T∗2 , and how T2 can be recovered from T

∗
2 by means of

a physical limit.
In any case, these considerations show that even if one manages to obtain a theory

T∗2 in which the models of T1 (and thus T1-objects) are metaphysically downgraded,
this downgrading will not be accomplished by (GenDual), but rather by the physical
interpretation of the models of T ′2—which turn on details which are manifestly inde-
pendent of (GenDual).40

Let us now summarize this Section and generalize from the example of EAs to
draw a positive moral for the physical application of (GenDual). Our discussion has
shown that Bain’s first strategy to defend (Objectless) fails. Our discussion of his gen-
eral argument showed the limitations of trying to translate statements about elements
(of set-theoretic objects) into statements that involve only morphisms of a category.
Our discussion of his specific argument showed that it is underwritten by (GenDual).

ceptualize classical and quantum mechanics in this way.
40A fortiori, it is not category theory which does the work of accomplishing this downgrading, since

category theory is crucial only in establishing the equivalence step of (GenDual).

24



Furthermore, we argued that while (GenDual) – and the associated use of category the-
ory – does not necessarily have any bearing on ontology, it is a useful methodological
tool for generating new physical theories. Indeed, one might say that it provides the
following recipe for constructing or partially constructing novel theories (or extensions
of theories):

1. Collect some initial class of models into a category T1.

2. Identify some feature f that you would like the new models to exhibit, but which
is inconsistent with T1.

3. Implement (GenDual), i.e. find a category T2 that is equivalent to T1, and extend
it to T∗2 by deforming T2 to possess (the translated version of) f. By core-
dependent homonymy, T∗2 can be called the ‘generalized version’ of T1.

Let us again emphasize that this recipe cannot on its own determine the interpretive
content of the novel models, e.g. whether they should be thought of as ‘fringe cases’
of an extended version of T2, or as the core content of a new and more fundamental
theory, from which T2 is to be recovered under some appropriate physical conditions.
And as we saw in the example of Geroch’s motivation for developing EAs (i.e. in
order to describe quantum gravity), the recipe will in general suggest nothing more
than a template for the latter purpose. One will often have to do further, conceptually
independent work in order to fully specify such novel models and explain how they
effectively yield the models of T2 under the appropriate conditions.

4.2 Generalization by categorification
We now turn to the task of developing Bain’s observations about structural differences
between categories. In particular, we shall discuss (i) how Bain’s examples of struc-
tural differences between categories leads us to another form of generalization, viz.
(GenCat); and (ii) the example of TQFTs, which is complex and subtle – in particular,
it is not clear what work (if any) (GenCat) is doing in TQFTs.

4.2.1 The framework for (GenCat)

In Section 3.2.2, we argued that formulating physics in categories which are struc-
turally different from Set would not help Bain’s defense of (Objectless). However,
it is apparent that Bain’s second strategy has uncovered a phenomenon of great inter-
est, viz. that one can generalize the properties of physical theories from the level of
individual sets to the level of abstract categorical properties, and that such properties
often have interesting physical interpretations and implications. The abstract categor-
ical properties, e.g. that of a category’s being symmetric monoidal, find instances in
particular physical categories, e.g. Hilb or nCob. We shall call this generalization
by categorification, or (GenCat) for short.

Before providing a precise framework for describing this sequence of steps, let us
explain how it is motivated by the structural differences that Bain discusses. For ex-
ample, consider the categories Set and Hilb. Both categories are monoidal, meaning

25



that, given any two objects A and B in the category, there exists a third object A⊗B.41
If one interprets the objects of the category as abstract state spaces (as is indeed sug-
gested by Hilb), then one can interpret the monoidal product as capturing the concept
of ‘composition of state spaces’.

Although Set and Hilb are both instances of monoidal categories, they each have
a very different kind of monoidal structure, which indicates a structural difference be-
tween these categories. On the one hand, in Set the monoidal structure is given by
the cartesian product, and so an element of a composite ‘system’ A × B is always a
pair (a,b). If we consider Set to represent physical systems, then this corresponds to
the fact that the ‘parts determine the whole’—a feature that we would associate with
classical systems. On the other hand, in Hilb the monoidal structure is given by the
tensor product of Hilbert spaces, for which there are elements of H1 ⊗ H2 that are not
of the form ψ1 ⊗ ψ2, i.e. there exist entangled states.

We can thus identify a physical property, viz. entanglement, that is possessed by
monoidal categories such as Hilb. However, the above way of identifying this property
relies on a distinction that is based on how the objects in each category are internally
defined – in other words, the identification is based on a property of the elements of the
(set-based) objects of the category.

This set-theoretic distinction is hardly transparent in terms of the overall structure
of Set and Hilb. However, it turns out one can achieve such transparency—and indeed
generality—by formulating the distinction in terms of category theory. For a monoidal
category C, the absence of entanglement in the above sense corresponds to C satisfying
the axioms of a cartesian category, of which Set is an instance.42 On the other hand,
a category that is monoidal but not cartesian can be said to possess the relevant notion
of entanglement. We thus see that given an appropriate background interpretation,
the notion of a cartesian category provides a category-theoretic distinction between
theories with a notion of entanglement, and those without.43

Moreover, we have achieved the following generality in our description: since
this notion of entanglement has been described categorically, it can be applied to any
monoidal category (whose objects are interpreted as state spaces and whose morphisms
are interpreted as dynamical processes) and not just Hilb and Set. Take for instance
the category Rel (introduced in Section 3.2.2), which has the same objects as Set, but
which is not a cartesian category. One might thus say that Rel can be considered to
‘have entanglement’—in order for this locution to be meaningful, of course, we must
be able to interpret Rel physically, e.g. as in (Edwards, 2009) and (Coecke et al.,
2009), where the morphisms are interpreted as non-deterministic classical processes
between state spaces.

41 The physical interpretation of the monoidal product requires a general interpretation of the category—
this is indeed clear in the case of Hilb, since it describes some of the mathematics of quantum mechanics.
See (Coecke and Paquette, 2009) for an interpretation of the category Set.

42A cartesian category is a monoidal category supplemented with an extra condition, viz. that for any
object A, there are projection morphisms π1 and π2 that guarantee that every element p : I → A satisfies
p = ⟨π1 ◦ p, π2 ◦ p⟩. The category Set is an example, since for any element (a, b) ∈ X there projection
morphisms π1 ◦ (a, b) = a and π2 ◦ (a, b) = b.

43 The label of ‘categorification’ is also used by Baez (2004, p. 12). Our usage is similar but subtly
different. In particular, the programme of categorification pursued by Baez typically includes further steps
such as replacing identity statements with isomorphisms.

26



Let us take stock: the physical applications that we have described—and on which
Bain draws—turn out to have a significance that is rather different from Bain’s morals.
In particular, the generality implicit in formulating structural differences (say between
some category C and Set) at a categorical level does not suggest (Objectless), but
rather that one can use such categorical properties to qualitatively distinguish between
theories, each of which is defined as a category of structured sets.

The key steps behind this generalization by categorification (GenCat) can be artic-
ulated as follows:

1. Identification: Isolate a physical phenomenon in a specific category C of set-
based objects;

2. Categorical Definition: Define the phenomenon in purely categorical terms, i.e.
without referring to any specific category.

3. Application: Apply this definition to other specific categories, which will instan-
tiate the physical phenomenon.

The utility of (GenCat), is three-fold. Firstly, as we saw above with the example
of Rel and Hilb, (GenCat) can be used to classify theories (thought of as categories)
according to the type of physical resources that they have. Secondly, (GenCat) can
be used to generate new models with the physical properties that are desired. For ex-
ample, given the category-theoretic definiton of entanglement, any monoidal category
that is not cartesian can potentially serve as a representation of a physical theory with
entanglement. Thirdly, this allows for an incisive high-level analysis of the logical
relationship between different physical phenomenona: a striking example is given in
(Coecke et al., 2009) and (Coecke et al., 2011), where it is shown—using categorical
definitions—that there exist categories which have entanglement and complementary
observables but which do not provide non-locality. Thus non-locality is shown to be
independent of entanglement.

Before concluding this section, let us briefly return to Bain’s argument concerning
(Objectless). Recall that Bain’s thinks of physical structure as something that is borne
by a C-object in a category (cf. (Structure) in Section 3.1.1). On the other hand, our
discussion above show that the categorical properties of Hilb can not only be used
to formalize the physics of quantum systems, but also that doing so encodes physical
structure in the algebra of morphisms of the relevant category (which does not corre-
spond to either the structure of the objects, or the specific morphisms which preserve
the object’s structure). This provides a further reason for thinking either that Bain’s
notion of physical structure is too limited, and that his second strategy is not relevant
to evaluating the plausibility of (Objectless).

4.2.2 A closer look at TQFTs

We have used Hilb to illustrate (GenCat), but Bain uses the examples of Hilb and
nCob to discuss topological quantum field theories (TQFTs). This raises the question:
does the use of TQFTs in physics exemplify (GenCat)? As we shall see, the answer is
rather subtle.

27



We will focus on understanding the significance of a result that we shall refer to as
‘Kock’s theorem’ below. This result is of interest for two reasons: first, it is a central
result in the the study of TQFTs; second, it appears to be an instance of (GenCat), at
least prima facie.

Recall from Section 3.2 that a TQFT is a functor:

T : nCob −→ Hilb.

Since T is a functor, it consists of an assignment of objects and morphisms to the
category nCob. First, T assigns a quantum state space, viz. a Hilbert space, to an
object in nCob. And second, T assigns a linear map to a morphism in nCob, i.e. a
cobordism. Since a cobordism is a a kind of ‘spacetime’, the linear map that T assigns
can be thought of as an evolution operation. For example, if n = 2, then the objects
in nCob are disjoint unions of circles, and the morphisms are 2-dimensional oriented
topological spaces, where each of these has a boundary given by a disjoint union of
circles.

The spacetime structure in 2Cob is evidently rather simple. In particular, a TQFT
is a topological theory. This means that quantities that are calculated for a particular
type of evolution are the same for any topological deformation of that evolution. For
example, the probability amplitude for transition from one state to another depends
only on the topology change that a cobordism supplies. So a TQFT will assign the
same linear operator in Hilb to the following two morphisms in nCob:

A particularly interesting feature of TQFTs for our purposes is that 2-dimensional
TQFTs can be classified in a purely category-theoretic way. To explain the classifi-
cation theorem, it will be helpful to introduce the graphical calculus that exists for
describing TQFTs.

For example, consider the cobordisms:

(2)

A TQFT detects only topology change, e.g. the ‘merging’ of two circles to one circle
in Eq. 2. We can depict this as:

u = e =

(3)

28



Similarly, the associativity of the ‘merge’ operation represented by the cobordism in
Eq. 2 is represented as:

=

(4)

Now, Eq. 3 and Eq. 4 can be considered to be graphical laws, in the sense that the
topological deformations can now be seen as rules for rewriting graphs (see Selinger
2011 for a more formal discussion of graphical languages in monoidal categories).
These graphical laws correspond to algebraic laws, and thus: the diagrams in Eq. 3
correspond to morphisms u : A ⊗ A → A and e : A → I, and the graphical rules in
Eq. 4 define a monoid object on the object A.

Now, as we have stated it, Eq. 4 is an equation in nCob, since the morphisms u
and e are cobordisms. But we can define morphisms u and e, and impose Eq. 4 in
any monoidal category C. In particular, we can define a monoid object in the category
Set: in this case a monoid object is just a monoid—i.e. the usual notion of a set with
a binary operation and a unit element. This process of abstraction can be thought of as
categorifying the cobordism laws (or indeed the rules for a monoid), since it amounts
to defining categorical rules for the topological laws of cobordisms. Note also that we
can continue to use the graphical rules in C, since these correspond to the algebraic
rules. The definition of a monoid object here appears to be an instance of (GenCat)—
to what extent is this appearance accurate? We shall return to this question shortly,
but note that (GenCat) requires the identification of a physical phenomenon before the
categorification step.

Given our graphical notation, we can define a similar set of graphical rules to Eq. 4
as in (Coecke and Paquette, 2009, p. 87), but with the diagrams reflected in the x-axis.
For example, given morphisms δ : A → A ⊗ A and ϵ : A → I, we can consider the
condition:

=

(5)

These rules define a comonoid, where the prefix ‘co’ is used to indicate that the direc-
tion of the arrows—or, correspondingly, the orientation of the diagram—is reversed.
If an object A carries both a monoid and a comonoid (subject to compatibility laws,
known as the Frobenius laws), then A is said to carry a Frobenius algebra.

Now, to state the classification theorem we must define the category of two-dimensional
TQFTs, denoted 2TQFT: its objects are TQFTs, i.e. the functors T defined above,
and its morphisms are monoidal natural transformations. There is then an algebraic
classification of TQFTs as follows. Consider the category cFA of commutative Frobe-
nius algebras and Frobenius algebra homomorphisms. As Kock (2003) explains, a

29



TQFT induces a Frobenius algebra: a TQFT F : nCob → Hilb preserves monoidal
structure, and so it preserves the Frobenius algebra laws such as Eq. 4 and Eq. 5. Hence
a TQFT defines a commutative Frobenius algebra in Hilb. Conversely, it can be shown
that a Frobenius algebra defines a TQFT. Thus we obtain the following equivalence,
due in part to Abrams (1996) (and with much useful exposition by Kock [2003]):

2TQFT ≃ cFA.

We refer to this equivalence as Kock’s theorem44.
Now, recall that Bain introduced TQFTs as an example of category-theoretic gen-

eralization that is similar to Hilb. However, having sketched the details of this gener-
alization, viz. (GenCat), we can see that they are in fact rather different. On the one
hand, Hilb is an instance of a symmetric monoidal category, and by interpreting its
morphisms as the dynamical processes of quantum mechanics, we can formulate vari-
ous physical properties (e.g. entanglement) at the level of general categorical proper-
ties. On the other hand, a TQFT is a functor between two specific symmetric monoidal
categories (viz. nCob and Hilb), and it is not obvious to see how one can abstractly
formulate any information of physical relevance at the level of categorical properties.
Kock’s theorem is important precisely because it provides a sense in which such an
abstraction occurs, viz. it says that a TQFT is equivalent to a Frobenius algebra object,
which can be defined in any symmetric monoidal category.

Can Kock’s theorem then be described as an example of (GenCat)? Prima facie, the
answer is not clear, because it is unclear which physical properties are being encoded
in the categorical structure of ‘Frobenius algebra object in a category’; and it is also
unclear which specific categories can provide physical instantiations of this structure.

However, a partial answer is forthcoming from recent developments in physics. For
instance, it has been shown that the algebraic sector of some conformal field theories
can be classified by certain Frobenius algebra objects (Runkel et al., 2007). In this
case, the Frobenius algebra encodes the fundamental (algebraic) structure of the theory,
and it is instantiated in the category of representations of different conformal field
theories. Our conceptual work shows one way in which the elaboration of this work
can be construed as an instance of (GenCat): the ‘physical phenomenon’ that is being
identified is the abstract structure of a conformal field theory, which is then described
in ‘purely categorical terms’ as a Frobenius algebra object. By choosing the category
in which this Frobenius algebra object lives, one then arrives at a determinate instance
of this abstract structure, viz. a particular conformal field theory.

4.3 The rôle of category theory in physics
Thus far, we have identified two forms of category-theoretic generalization that are
employed in physics. We will now comment on some differences between the two
strategies, and emphasize the significance of these differences.

In the first type of generalization, viz. (GenDual), the set-theoretic models of a
theory are simply collected into a category—it thus follows that these models (and

44 This name is justified by the fact that, although the roots of this result lie with the other authors that we
have cited, (Kock, 2003) is the main source for the explicitly category-theoretic formulation of this result.

30



thus the theory) are not defined categorically, i.e. their mathematical structure does not
rely only (or even primarily) on category theory. On the other hand, the second form
of generalization, viz. (GenCat), does indeed aspire to define (the abstraction of) a
physical phenomenon in purely categorical terms.

For this reason, one sees that morphisms play a different role in (GenDual) and
(GenCat) respectively. In (GenDual), morphisms are typically taken to be the auto-
morphisms of a model (e.g. the diffeomorphisms from a spacetime to itself, in Bain’s
example) and thus do not connect different objects in a category. But in (GenCat), the
relevant physical phenomena need to be represented in way that is germane to category-
theoretic abstraction, e.g. dynamical processes are typically represented as morphisms
between objects, and composition is represented by a functor. The generalization then
occurs when we ask that these structures retain their meaning in virtue of the abstract
properties of the relevant category, e.g. being symmetric monoidal and having a dagger
functor.

These considerations point to a central difference in the role that category theory
plays for (GenDual) and (GenCat) respectively. In (GenDual), the main role of category
theory is to provide a way of saying when two theories (represented by categories) are
equivalent; the generalization then comes about by extending one of these theories and
using analogy to conceive of the new objects in the extension as ‘generalized models’
of the other theory. Thus, while category theory provides the right language to express
‘equivalence’, which is in turn necessary (but insufficient) for constructing the analogy,
it is clear that it does not itself provide the desired generalization. That is to say, the
generalization does not involve category-theoretic propreties; instead it involves using
sets with structure, of the form of generalizing from groups to monoids.

By contrast, in (GenCat), it is precisely category theory which accomplishes the de-
sired generalization, since this is brought about by reformulating the physically relevant
properties of some specific category such as Hilb as the properties of an abstract cate-
gory, whose objects are not further specified as being of any particular type. This phys-
ical application of category theory is the analog of taking the notion of a set-theoretical
mathematical object, e.g. a monoid, and generalizing it to a purely category-theoretic
concept, e.g. a monoid object, which can be formulated in any monoidal category.

The differences between the two strategies can be summarized as follows. We
might say that in the first strategy, category theory is playing an organizational role, in
the sense that it is employed to collect together the models of a theory in a single math-
ematical structure, i.e. a category. The morphisms in this category are ‘scaffolding’ for
the other mathematics that is employed, and do not play a role in elucidating the physi-
cal concepts. On the other hand, the second strategy exhibits the direct use of category
theory in physics. Here, category theory is used to directly encode physical processes.
Among other things, this means that the choice of underlying category is potentially
not crucial for doing calculations, in the sense that physical consequences are derived
by e.g. computing the composition of morphisms. What’s important is that the category
satisfies certain category-theoretic axioms (e.g. having monoidal structure), which are
then used to calculate physical consequences.

31



5 Conclusion
Our path has taken us from Bain’s argument for OSR to uncovering two paradigmatic
forms of generalization for physical theories. Our work suggests two specific avenues
for further research.

Firstly, it raises the question of the extent to which the forms of reasoning that we
have uncovered rely on category theory per se. That is, could generalization by duality
and generalization by categorification be implemented with something other than the
mathematics of category theory? This might seem rather unlikely in the latter case,
but a priori it could be that, e.g. lattice theory, could play an equally powerful rôle in
formalizing the types of generalization.

Furthermore, one might observe that the use of category theory in physics is a
relatively recent development—and thereby argue that its prominence could be faddish.
However, there are reasons to think otherwise. For example, although the initial work
on TQFTs in (Witten, 1988) was not itself explicitly category-theoretic, much of the
current work in this field now uses category theory in a widespread and deep way, as
shown by the work of Baez and Dolan (1995) and Runkel et al. (2007). This suggests a
kind of evolution in the definition of a physical theory: from a set-theoretic formalism
to a category-theoretic formalism. If such a process holds true in general, then it could
point to the essential utility of category theory in physics.

Secondly, it is interesting to note that Awodey (2004) and Landry (2013) have ar-
gued that category theory provides a formalism for one sense of mathematical struc-
turalism, which Awodey calls ‘schematism’. These arguments concern distinguishing
between the methodological and metaphysical significance of the use of category the-
ory in pure mathematics. Since this has been one of our concerns in this paper—but
with respect to the use of category theory in physics—the work of Awodey and Landry
is ripe for comparison with ours.

Funding. We gratefully acknowledge the support of the John Templeton Foundation.

Acknowledgements. We thank audiences in Oxford, Cambridge, Princeton and Chicago
for their comments. We also thank Jeremy Butterfield for extensive feedback on earlier
drafts, as well as Jonathan Bain and two anonymous referees for comments that helped
to improve this paper. Last but not least, we thank Colin Stephen for highly stimulating
discussion that led to the inception of this work.

References
Abrams, L. (1996). Two-dimensional topological quantum field theories and Frobenius

algebras. Journal of Knot Theory and its Ramifications, 5(05):569–587.

Atiyah, M. (1988). Topological quantum field theories. Publications Mathématiques
de l’IHÉS, 68(1):175–186.

32



Awodey, S. (2004). An answer to Hellman’s question: Does category theory provide a
framework for mathematical structuralism?’. Philosophia Mathematica, 12(1):54–
64.

Baez, J. (2004). Quantum quandaries: a category-theoretic perspective. arXiv:quant-
ph/0404040.

Baez, J. C. and Dolan, J. (1995). Higher-dimensional algebra and topological quantum
field theory. Journal of Mathematical Physics, 36:6073.

Bain, J. (2003). Einstein algebras and the hole argument. Philosophy of Science, 70
(5):1073–1085.

Bain, J. (2013). Category-theoretic structure and radical ontic structural realism. Syn-
these, 190:1621–1635.

Brown, R. (2007). Topology and groupoids. Bull. London Math. Soc, 39:867–872.

Coecke, B., Edwards, B., and Spekkens, R. (2009). The group theoretic origin of
non-locality for qubits. Technical Report RR-09-04, OUCL.

Coecke, B., Edwards, B., and Spekkens, R. W. (2011). Phase groups and the origin of
non-locality for qubits. Electronic Notes in Theoretical Computer Science, 270(2):15
– 36. arXiv:1003.5005.

Coecke, B. and Paquette, E. O. (2009). Categories for the practising physicist.
arXiv:0905.3010.

Connes, A. and Marcolli, M. (2008). Noncommutative geometry, quantum fields and
motives, volume 55. American Mathematical Soc.

Edwards, B. (2009). Non-locality in categorical quantum mechanics. PhD thesis,
University of Oxford.

Esfeld, M. and Lam, V. (2011). Ontic structural realism as a metaphysics of objects.
In Bokulich, A. and Bokulich, P., editors, Scientific structuralism, page 143159.
Dordrecht: Springer.

French, S. (2011). Shifting to structures in physics and biology: A prophylactic for
promiscuous realism. Studies in History and Philosophy of Science Part C: Studies
in History and Philosophy of Biological and Biomedical Sciences, 42(2):164–173.

Frigg, R. and Votsis, I. (2011). Everything you always wanted to know about structural
realism but were afraid to ask. European journal for philosophy of science, 1(2):227–
276.

Geroch, R. (1972). Einstein algebras. Communications in Mathematical Physics,
26(4):271–275.

Heller, M. (1992). Einstein algebras and general relativity. International journal of
theoretical physics, 31(2):277–288.

33



Khalkali, M. (2009). Basic Non-commutative Geometry. EMS.

Kock, J. (2003). Frobenius Algebras and 2-D Topological Quantum Field Theories.
London Mathematical Society Student Texts. Cambridge University Press.

Ladyman, J. and Ross, D. (2007). Every thing must go: Metaphysics naturalized,
volume 61. Oxford University Press Oxford.

Lam, V. and Wuthrich, C. (2013). No categorial support for radical ontic structural
realism. ArXiv e-prints.

Lambek, J. and Scott, P. (1988). Introduction to Higher-Order Categorical Logic.
Cambridge Studies in Advanced Mathematics. Cambridge University Press.

Landry, E. (2013). The genetic versus the axiomatic method: Responding to feferman
1977. The Review of Symbolic Logic, 6:24–51.

Lawvere, F. W. (1963). Functorial semantics of algebraic theories. Proceedings of the
National Academy of Sciences of the United States of America, 50(5):869.

Mac Lane, S. (2000). Categories for the Working Mathematician. Springer-Verlag.

Morganti, M. (2004). On the preferability of epistemic structural realism. Synthese,
142(1):81–107.

Müger, M. (2008). Tensor categories: A selective guided tour. arXiv:0804.3587.

Rieffel, M. (1994). Quantization and c∗ algebras. Contemporary Mathematics, 167:67–
97.

Runkel, I., Fjelstad, J., Fuchs, J., and Schweigert, C. (2007). Topological and confor-
mal field theory as frobenius algebras. Contemporary Mathematics, 431:225.

Schwarz, J. H. (1997). Lectures on superstring and M theory dualities: Given at ICTP
Spring School and at TASI Summer School. Nucl.Phys.Proc.Suppl., 55B:1–32.

Selinger, P. (2011). A survey of graphical languages for monoidal categories. In Co-
ecke, B., editor, New Structures for Physics, volume 813 of Lecture Notes in Physics,
pages 289–355. Springer Berlin / Heidelberg.

Shields, C. (1999). Order in Multiplicity: Homonymy in the Philosophy of Aristotle.
Oxford University Press on Demand.

Witten, E. (1988). Topological quantum field theory. Communications in Mathematical
Physics, 117(3):353–386.

34


	Introduction
	Categories and structuralism
	Categories: abstract and concrete
	Structuralism: simple and ontic

	Bain's two strategies
	A first strategy for defending (Objectless)
	The general argument
	The specific argument

	A second strategy for defending (Objectless)
	Purely category-theoretic physics?
	O-objects in category-theoretic physics?


	Two forms of categorical generalization
	Generalization by duality
	The framework for (GenDual)
	Algebraic-geometric duality and Einstein algebras

	Generalization by categorification
	The framework for (GenCat)
	A closer look at TQFTs

	The rôle of category theory in physics

	Conclusion