Applied Cat-Theory

TA I - D A N A E B R A D L E Y
arXiv:1809.05923v1 [math.CT] 16 Sep 2018
W H AT I S A P P L I E D
C AT E G O RY T H E O RY ?
D E PA R T M E N T O F M AT H E M AT I C S
C U N Y G R A D U AT E C E N T E R
NEW YORK, NEW YORK
tbradley@gradcenter.cuny.edu
what is applied category theory? 2
What is applied category theory?

Upon hearing the phrase “applied category theory,” you might be thinking either one of two thoughts:
#1 Applied category theory? Isn’t that an oxymoron?

#2 Applied category theory? What’s the hoopla? Hasn’t category theory always been applied?
For those thinking thought #1, I hope to convince you that the answer is No way! It’s true that category
theory sometimes goes by the name of general abstract nonsense, which might incline you to think that cate-
gory theory is too pie-in-the-sky to have any impact on the “real world.” My hope is that these notes will
convince you that that’s far from the truth!
For those thinking thought #2, yes it’s true that ideas and results from category theory have found ap-
plications in computer science and quantum physics (not to mention pure mathematics itself), but these
are not the only applications to which the word applied in applied category theory is being applied. So what is
applied category theory?
Read on.
A quick note to the reader

Before we get started, I’ll mention that this document is a collection of notes I amassed while participating
in the 2018 Applied Category Theory Adjoint School—a wonderful online seminar that ran from January -
April 2018 and culminated in a two-week workshop at the Lorentz Center in May 2018. I had a blast learn-
ing from the folks there, and I want to share some of the things I learned with anyone who’s interested.
(So thanks for being interested!) Later, I’ll describe a couple of the research projects discussed during
the workshop. Much of the information in this PDF can be found in various journal articles, blog posts,
and videos of conference talks, most of which are freely available online. I’ve provided citations to these
throughout. Here are a few other things to know:
• I’ll assume the reader is comfortable with the basics of cate-

gory theory: categories, functors and natural transformations. For a For a gentle introduction to (pure)
category theory, here are a few places to
friendly introduction to these topics, feel free to browse through
start:
the articles from my blog Math3ma listed in the margin.
– What is Category Theory, Anyway?
• I’ll make heavy use of hyperlinks, as I have already, and I’ll also – What is a Category?
incorporate the occasional use of color throughout the text. For – What is a Functor?
these reasons, it’s probably best to read this PDF on a computer – What is a Natural Transformation?
rather than in print form. At the first link, you’ll find a list of
other recommended resources for
learning about category theory.
• Finally, a fair warning: I use italics a lot (along with frequent par-
enthetical remarks). I also like exclamation points! And many of
my sentences begin with a conjunction.
My gratitude goes to the participants and mentors of the 2018 ACT Workshop from whom I learned a great deal. I also thank John
Baez, Joseph Hirsh, and Maximilien Péroux for providing valuable feedback on a first draft of these notes.
Introduction
One of the great features of category theory, birthed in the 1940s, is that its organizing principles have
been used to reshape and reformulate problems within pure mathematics, including topology, homotopy
theory and algebraic geometry. Category theory has light on those problems, making them easier to solve
and opening doors for new avenues of research. Historically, then, category theory has found immense
application within mathematics. As John Baez recently noted, “[category theory] was meant to be applied.”
More recently, however, category theory has found applications in a wide range of disciplines outside
of pure mathematics—even beyond the closely related fields of computer science and quantum physics.
These disciplines include chemistry, neuroscience, systems biology, natural language processing, causal-
ity, network theory, dynamical systems, and database theory to name a few. And what do they all have in
common? That’s precisely what applied category theory seeks to discover. In other words, the techniques,
tools, and ideas of category theory are being used to identify recurring themes across these various dis-
ciplines with the purpose of making them a little more formal. This is what the phrase applied category
theory (ACT) is meant to describe. As explained on the ACT 2018 workshop webpage,
...we should treat the use of categorical concepts as a natural part of transferring and integrating knowledge
across disciplines. The restructuring employed in applied category theory cuts through jargon, helping to elu-
cidate common themes across disciplines. Indeed, the drive for a common language and comparison of similar
structures in algebra and topology is what led to the development category theory in the first place, and recent
hints show that this approach is not only useful between mathematical disciplines, but between scientific ones as
well.
Of course, one of the challenges of using category theory to transfer and integrate knowledge across
disciplines is making category theory itself accessible to the broader scientific audience. John Baez and
Brendan Fong address this very point in their 2016 paper on electrical circuit diagrams1 :
While diagrams of networks have been independently introduced in

many disciplines, we do not expect formalizing these diagrams to
immediately help the practitioners of these disciplines. At first the flow
of information will mainly go in the other direction: by translating 1
A Compositional Framework
ideas from these disciplines into the language of modern mathematics, for Passive Linear Networks,
https://arxiv.org/pdf/1504.05625.pdf
we can provide mathematicians with food for thought and interesting
new problems to solve. We hope that in the long run mathematicians
can return the favor by bringing new insights to the table.
Although their comments refer to a particular project, they can apply to the field at large, too.
The goal of this document is to give a taste of applied category from a graduate student’s perspective.
In doing so, I’ll share two themes and two constructions that appeared frequently during the ACT 2018
workshop. The math underlying these themes and constructions is not new. The newness, rather, is in how
they are being applied. To illustrate the themes and constructions, I’ll also share two examples—two re-
search projects in the field of ACT. The first project relates to chemistry and the second to natural language
processing, though the expositions are weighted unevenly. I’ll devote considerably more time on the sec-
ond example since that’s where my own research interests lie. And that’s what’s on the carte du jour! Two
themes and two constructions and two examples, along with a few crumbs (i.e. digressions) in between.
Here’s the menu in more detail:
Contents
1 Two Themes 5
1.1 Functorial Semantics 5
1.2 Compositionality 11
1.3 Further Reading 12
2 Two Constructions 13
2.1 Monoidal Categories 13 Although the items are listed linearly,
they are very much intertwined. The
2.2 Decorated Cospans 28 themes motivate the constructions; the
constructions embody the themes, and
2.3 Further Reading 29 both the themes and the constructions
come to life in the examples.
3 Two Examples 30
3.1 Chemical Reaction Networks 30
3.2 Natural Language Processing 35
3.3 Further Reading 48
4 But Wait! There’s More... 50
I like to imagine that category theory is like a cup of black coffee,

while fields outside of pure mathematics are like fresh cream. Both
are lovely on their own, but blending them makes for a beverage par
excellence.
I hope you’ll enjoy it as much as I do!

1 Two Themes
Two themes that appear over and over (and over and over and over) in
applied category theory are functorial semantics and compositional-
ity. Let’s talk about the first one first.
1.1 Functorial Semantics

Functorial semantics relates to the idea that a structure-preserving The phrase “functorial semantics” was
coined by William Lawvere.
functor between categories
CÑD
Ð This is how Lawvere defines the word
can be viewed as an interpretation of C within D. It’s often helpful to “functor” in his book with Stephen
think of C as somehow encoding for syntax while D provides seman- Schanuel, Conceptual Mathematics! It’s
tics. Syntax refers to rules for putting things together and semantics a nice introductory text, by the way.
refers to the meaning of those things. A functor
syntax Ñ semantics
provides a way to bring the syntax to life.
To get a better idea of syntax vs. semantics, think of the En- I’m using English language as an anal-
ogy to illustrate syntax vs. semantics,
glish language where two important features of communication
but it’s more than an analogy! As we’ll
are 1) grammar, which provides rules for combining words to form see in Section 3.2, the pairings
sentences, and 2) the actual meaning conveyed by those words and
“grammar ù syntax”
sentences. Grammar is the syntax, and the meaning is the semantics. “meanings of words ù semantics”
grammar ù syntax meaning ù semantics. become quite literal in applied category

theory!
Of course, neither is useful on their own. For instance, it’s easy to
come up with a sentence that is grammatically correct and yet has
no meaning. That’s the whole point behind MadLibs! As another
example, here’s a sentence attributed to linguist Noam Chomsky:
Colorless green ideas sleep furiously.
It is grammatically correct, yet it has no meaning. The point here is

that syntax vs. semantics is nothing new. So when an applied category
theorist wants to model some phenomena in the “real world,” don’t
be surprised if their model is ultimately a functor from a syntax
category to a semantics category!
A small-ish digression...
Even though the idea goes by the fancy name of functorial semantics, it
is not just a “category theory thing.” Mind if I digress for a while to
I’ll take your silence as a No.
elaborate on this?
If you know a little bit about groups, then you’ve seen functo-
rial semantics in action before! How so? A group is a set endowed
with some extra structure, though that tells us nothing about why
groups are useful. It’s better to think of a group as encoding for some
kind of action or transformation. And this is why group representa- Group elements are like verbs. They
tions are so great! A group representation provides a way to view your DO stuff! For more on this notion from
a categorical perspective, check out the
abstract group elements as concrete linear transformations of some article Group Elements, Categorically
vector space. Explicitly, given a vector space V, a group representa- on Math3ma.
tion is a group homomorphism from G to AutpVq, the group of all
automorphisms of V If we replace AutpVq by AutpXq for
G Ñ AutpVq some set X (i.e. the group of automor-
phisms, i.e. bijections, on X), then a
It assigns to each group element a linear isomorphism V Ñ V. group homomorphism G Ñ AutpXq is
precisely a group action on X.
As a quick example, suppose our group is D3 , the dihedral group
of order 6, which is the group of symmetries of an equilateral trian-
gle. If we were to look at a presentation of the group,
D3 “ xr, s | r3 “ s2 “ rsrs “ 1y
it might not seem to have anything to do with triangles. Fortunately,

a representation of D3 makes the connection clearer by assigning to
each group element r and s a linear transformation of the real plane.
Specifically, the standard representation of D3 Ñ AutpR2 q assigns to
each of r and s an invertible 2 ˆ 2 matrix with real entries:
« ff « ff
cosp2π{3q ´ sinp2π{3q 1 0
r ÞÑ R “ s ÞÑ S “
sinp2π{3q cosp2π{3q 0 ´1
Here R is a rotation by 60˝ while S is reflection across the x-axis.

Moreover R3 and S2 and RSRS each are equal to the 2 ˆ 2 identity
matrix, which is exactly what we would expect: rotating an equi-
lateral triangle by one full revolution leaves it unchanged, as does
reflecting it twice in a row, and so on.
More generally then, we can think of a group G as providing the
syntax while automorphisms AutpVq provide the semantics
G ù syntax AutpVq ù semantics
So a group representation is like a (structure-preserving) morphism
syntax Ñ semantics
In fact... it’s not like that. It IS that. If we view both the groups G and
AutpVq as one-object2 categories, then a group representation 2
Every group G gives rise to a category
having a single object ‚ (the group
g
G Ñ AutpVq itself) and a morphism ‚ Ñ ‚ for each
group element g P G. Composition is
given by the group operation.
IS a functor from syntax to semantics. That’s because every group
homomorphism is a functor when the groups are viewed as one-
object categories! So although functorial semantics has the word
Here’s another example I can’t resist
“functor” in it, don’t think that the idea behind it is unique to cat- sharing: operads! If you’re not familiar
egory theory. Indeed, representation theory capitalizes on the rela- with operads, just know that this is
a souped-up version of the group
tionship between syntax and semantics: a representation assigns to an
theory example. If you are familiar
abstract algebraic gadget (the syntax) some concrete meaning (the with operads, then you know this is the
semantics). souped-up version of the group theory
example.
An operad is an example of syntax,
I could end our digression here, but I’d like to share one more while an algebra over that operad
provides the semantics. For example,
instance of functorial semantics at work in pure mathematics. The
given a vector space V, an operad
next few examples involve monoids and monoidal categories, so I’ll homomorphism from the [commu-
assume you are familiar with those words. If you are not familiar tative, associative, Lie, Poisson,...]
operad to the endomorphism operad
with those words, don’t fret—you’re in luck! Section 2.1 is all about on V IS a [commutative, associative,
monoids and monoidal categories, so feel free to read that section Lie, Poisson,...]-algebra! That is, the
structure-preserving homomorphism
first then come back here. In either case, let’s proceed with another
provides an interpretation of each ab-
neat example of functorial semantics in action: stract n-ary operation as a actual,
concrete operation V bn Ñ V.
Example: a monoid is the image of a functor from a certain syntax cate-

gory to a certain semantics category. This next comment is really digressing
from the digression, but: I also like to
think of simplicial sets as an instance of
More specifically,3 functorial semantics. A simplicial set X
is a bit like syntax, while a topological
space is like semantics. Geometric
realization X ÞÑ |X| provides a map
from one to the other.
3
A functor F : C Ñ D between
monoidal categories is called lax
monoidal if for every pair of objects
c, c1 in C there is a morphism
Here I’m viewing both 1 and Set as monoidal categories. The symbol Fc b Fc1 Ñ Fpc b c1 q
1 is meant to represent the category with one object and only one (which assembles into a natural trans-
morphism (the identity), which we can view as a monoidal category formation.) It’s called strong monoidal
if Fc b Fc1 – Fpc b c1 q, and it’s called
p1, b, 1q in exactly one way. The category of sets has a monoidal strict monoidal if Fc b Fc1 “ Fpc b c1 q.
structure given by the Cartesian product with the set containing one
element, denoted t˚u, as the monoidal unit. Technically then,
In footnote 3 you’ll notice I dropped

parentheses and wrote Fc rather than
Fpcq. The reason for my preference is
categorical! Let me explain by saying a
few words about sets: Did you know
that an element x in a set X is the same
thing as a function t˚u Ñ X, where
t˚u denotes the one-element set? It’s
true. The function t˚u Ñ X is uniquely
determined by where it sends that one
point ˚. So since an element x is the
same thing as an arrow t˚u Ñ X, we
might as well label that arrow by x,
Why is this true? First observe that a functor F : 1 Ñ Set picks out a
x : t˚u Ñ X
set, F1 :“ M. And the data of a lax monoidal functor F : p1, b, 1q Ñ
pSet, ˆ, t˚uq consists of a morphism Now if we have another function
f : X Ñ Y, then an element f pxq P Y is
precisely the composition
‚ : F1 ˆ F1 Ñ Fp1 b 1q i.e. ‚: M ˆ M Ñ M
x f
t˚u Ñ X Ñ Y
along with a morphism That is,
1 : t˚u Ñ F1 i.e. 1 : t˚u Ñ M f pxq “ f ˝ x “ f x
where on the right hand side, I’ve

both of which are required to fit into some commuting diagrams. I omitted the composition symbol ˝
because it’s cleaner. So there you have
won’t write them here, but one diagram says “‚ is associative” and it! An element f pxq P Y is the same as
the other diagram says, “1 serves as an identity for ‚.” In summary, a function f x. And since categorically-
the data of a lax monoidal functor F : p1, b, 1q Ñ pSet, ˆ, t˚uq are minded folks (such as you and I) prefer
arrows over elements (Because of
Yoneda. Also, we might be rethinking
i) a set M set theory.), the notation f x—and more
generally, Fc as above—is preferred. By
ii) an associative binary operation ‚ : M ˆ M Ñ M the way, this is all related to Lawvere’s
philosophy of generalized elements,
iii) a special element 1 :“ 1pt˚uq P M that serves as a “multiplicative which is the idea that a morphism
identity” for ‚. A Ñ B is really an “A-shaped element
in B.” For some examples, check out
the articles “A Diagram is a Functor”
This triple pM, ‚, 1q is precisely a monoid! Or to borrow from Law- as well as “The Yoneda Embedding”
vere’s terminology, the functor (equivalently, the monoid) is one on Math3ma. Generalized elements are
interpretation of the category 1 in the Set. Interestingly, 1 may be closely related to functorial semantics,
so both links are worth a read!
interpreted in other categories as well. This leads to other famil-
iar monoidal structures. Indeed, if we replace pSet, ˆ, t˚uq by any
monoidal category pC, b, 1q, then a lax monoidal functor
p1, b, 1q Ñ pC, b, 1q
is a monoid in the category C. Sometimes this monoid goes by a

familiar name. Here are some examples.
1. Topological Monoid. Let pTop, ˆ, ˚q denote the category of

topological spaces and continuous functions, viewed as a monoidal
category with the Cartesian product ˆ, with the one-point space ˚
as monoidal unit. A lax monoidal functor
p1, b, 1q Ñ pTop, ˆ, ˚q
is a topological monoid. That is, a topological monoid is a monoid

The idea that “a monoid in C is a lax
in the category of topological spaces.
monoidal functor 1 Ñ C” is completely
analogous to claim that “an element in
2. Ring. Let pAbGroup, b, Zq denote the category of abelian groups X is a function t˚u Ñ X” made in the
(viewed as Z-modules) and abelian group homomorphisms, margin on the previous page. In both
cases, with have two objects A and B
viewed as a monoidal category with the tensor product b, with
of the same kind (monoidal categories
the integers Z as monoidal unit. A lax monoidal functor on this page; sets on the previous page)
together with a structure-preserving
p1, b, 1q Ñ pAbGroup, b, Zq map A Ñ B.
(Caveat: a lax monoidal functor is
only somewhat structure-preserving.
is a ring (with unit). That is, a ring is a monoid in the category of That’s why it’s called lax. And a func-
abelian groups. tion is vacuously structure-preserving
since sets don’t have any structure! But
3. Algebra. Let pFVect, b, kq denote the category of finite-dimensional I digress...)
In both cases the object A is trivial
vector spaces over a field k and linear maps, viewed as a monoidal (technically, terminal)—it’s just a point,
category with the tensor product b, with k as monoidal unit. A lax so to speak. And in both cases the
arrow provides an interpretation of that
monoidal functor
point within the context of B.
p1, b, 1q Ñ pFVect, b, kq To phrase it another way, we are
probing B with a point-shaped object. In
is an algebra (with unit). That is, an algebra is a monoid in the the case when B is a set, probing it with
category of vector spaces. a point will pick out an element. In the
case when B is a monoidal category,
probing it with a point will pick out a
4. Monad. Let C be a category and let EndC denote the category
monoid!
whose objects are functors C Ñ C and whose morphisms are
natural transformations. (So EndC is the category of endofunctors
on C.) Note that EndC can be given the structure of a monoidal
category: the monoidal product is composition of functors (i.e. if
F, G are objects in EndC , then the monoidal product of F and G is
F ˝ G), and the monoidal unit is the identity functor 1C on C (i.e.
1C assigns each object and morphism in C to itself). Then a lax
monoidal functor
p1, b, 1q Ñ pEndC , ˝, 1C q
is a monad. That is, a monad is a monoid in the category of end-
ofunctors on C.
You’ll notice that in each of these examples, a change in the seman-

tics category C gives rise to a different interpretation of 1, which
served as our syntax category. Pretty neat, right? For more details on
the examples, see Emily Riehl’s Category Theory in Context Definitions
1.6.3 and 5.1.1.
That’s the idea behind functorial semantics. Now, how is it used

in applied category theory? We’ll see the answer when we look at
the two examples—two research projects from the field—one from
chemistry and one from natural language processing. In both exam-
ples, the key is the existence of a (structure-preserving) functor from a
syntax category to a semantics category. Here’s a sneak preview:
In Section 3.1 we’ll see how the behavior of a chemical reaction net-
work is modeled by a functor
Petri nets Ñ dynamical systems
as shown in “A Compositional Framework for Reaction Networks”

by John Baez and Blake Pollard. In Section 3.2, we’ll see how a model
for natural language can be described by a functor
grammar Ñ meanings of words
via the work of Bob Coecke, Mehrnoosh Sadrzadeh, and Stephen

Clark in “Mathematical Foundations for a Compositional Distribu-
tional Model of Meaning”. And perhaps you’re wondering, “How
do Petri nets, dynamical systems, grammar, and meanings of words form
categories? And what’s a Petri net, anyway?” We’ll answer these
questions in the pages to come, but first I’d like to introduce another
important theme in applied category theory: compositionality.
1.2 Compositionality
Compositionality, also known as the principal of compositionality,
also known as Frege’s principle, is the idea that the meaning of a Frege as in Gottlob Frege.
complex expression is determined by
1. the meanings of its constituent parts, and
2. the rules for how those parts are combined.

Or, as succinctly stated on the homepage of the brand new journal of
applied-category-theory-and-related-fields,
compositionality describes and quantifies how complex things can be
assembled out of simpler parts.
As it turns out, the name of that journal is itself Compositionality,4 4

A contending title was Applied Cate-
which hints at the importance of this concept within the field. gory Theory, but in the end Composition-
ality had the most votes.
In Section 3.1, which contains our example from chemistry, the
complex expression will be a network—a big complicated directed
multigraph, if you like. Its constituent parts are simply smaller
chunks of the network.
In Section 3.2, our example from natural language, the complex Matrix factorization provides another
illustration of compositionality in math-
expression will be a sentence; its constituent parts are the words that
ematics. As an example, every n ˆ m
comprise the sentence. matrix M has a singular value decomposi-
In both examples, functorial semantics and the principle of com- tion, which means it can be written as a
product of three matrices M “ UDV :
positionality will go hand-in-hand. The former prompts us to model where U and V are unitary square
behavior using a functor between syntax and semantics categories. matrices (here V : denotes the conjugate
The latter encourages us to take things one at a time: To model a huge transpose of V) and D is a rectangular
diagonal matrix. Intuitively then, the
system, compositionality tells us, it’s enough to model smaller pieces of linear transformation M can be broken
it and then stick those pieces together. Simple enough. But what does it down into a rotation followed by a
shear followed by another rotation. So
mean to “stick pieces together” mathematically? The answer is pro- you can analyze your transformation
vided by the structure of a monoidal category. And that is the first of (or your data set, if that’s what M is
encoding) by understanding its con-
our two main constructions in ACT.
stituent pieces—the factors—and how
they compose together. More generally,
I like to think that tensor networks are
a good example of compositionality, but
such a discussion might take us too far
off course. Perhaps another day!
1.3 Further Reading

For more on functorial semantics and compositionality:
• Take a look at (this small notice on) William Lawvere’s 1963 PhD
thesis “Functorial Semantics of Algebraic Theories” for the formal
foundations for functorial semantics.
• You might also enjoy this discussion on doctrines over at the n-

Category Café: https://golem.ph.utexas.edu/category/2006/09/doctrines.html
• The preface to Brendan Fong’s PhD thesis, “The Algebra of Open

and Interconnected Systems”, has a nice discussion on the princi-
pal of compositionality and includes various references. And while
you’re at it, take a look at the entire thesis, which is wonderfully
written and provides the backbone of much of John Baez’s current
research in network theory, which we’ll talk a little bit about in
Section 3.1.
• Applied category theorist Jules Hedges has also written a nice

exposition on compositionality, appropriately entitled “On Com-
positionality.” In the article, you’ll find a link to the Stanford Ency-
clopedia on Philosophy’s entry on compositionality, which gives a
thorough overview of the topic.
2 Two Constructions
Two constructions that appear over and over (and over and over and
over) in (some projects in) applied category theory are monoidal
categories and decorated cospans. Let’s talk about the first one first.
2.1 Monoidal Categories

Monoids and monoidal categories were
the main focus in the digression on
Actually, before we talk about monoidal categories, let’s talk about
page 7, but now I’ll proceed as if they
monoids. Here are three examples of monoids: the integers Z, the are new to the reader.
rational numbers Q, and the set of all n ˆ n matrices with real-number
entries Mn pRq. Well, technically these are monoids:
pZ, `, 0q pQ, ¨, 1q pMn pRq, ¨, 1q
Each example consists of a set X equipped with an associative

binary operation, which I’ll denote by ‚. Moreover, there is a special A “binary operation on X” is just the
element in the set, let’s call it 1, that serves as an identity for the fancy name for a function X ˆ X Ñ X.
So ‚ is a function ‚ : X ˆ X Ñ X. It
operation. Those three things—a set, an associative binary operation, assigns to a pair px, yq P X ˆ X an
an identity—comprise a monoid. Usually, we write this triple as element x ‚ y P X. To say that an
element 1 P X serves as an identity for ‚
means it satisfies 1 ‚ x “ x ‚ 1 “ x for all
pX, ‚, 1q x P X.
Not too bad, right?
Great. Now imagine replacing the set X by a category C, and re-

placing the binary operation ‚ : X ˆ X Ñ X by a functor ‚ : C ˆ C Ñ C,
and replacing the identity element 1 P X by an object 1 in C. The
resulting triad
pC, ‚, 1q
is called a monoidal category. The object 1 is often called the monoidal
unit, and people usually prefer to write b (and call it the monoidal
product) instead of ‚ so let’s do that too:
There’s a little more to the story here

since we want b to be associative: for
any objects A, B, C in the category C,
we want A b pB b Cq “ pA b Bq b C.
Alas, things are rarely equal on the
nose. To compensate for this, we ask
instead that there exist an isomorphism
A b pB b Cq – pA b Bq b C, which
should behave nicely. I won’t go into
the details here, but of course you can
find more on the Wikipedia page on
monoidal categories. For a delightful
exposition on the richness of monoidal
categories, I strongly recommend “A
Rosetta Stone” by John Baez and Mike
Stay. It is a gem.
In short, a monoidal category is a category in which it makes sense
to “combine” objects and morphisms.5 As we’ll see in Sections 3.1
and 3.2, each of the four categories mentioned on page 10—Petri nets,
dynamical systems, grammar, and meanings of words,—are monoidal
categories! Here are some more examples.
pSet, ˆ, t˚uq pTop, \, ∅q pFVect, b, kq
By the way, if there is an isomorphism A b B – B b A for all ob- 5

Allow me to explain the “and mor-
phisms” part. First remember, b is a
jects A and B that behaves nicely in a sense that can be made precise, functor! That means it’s an assignment
then we say that pC, b, 1q is a symmetric monoidal category. Each of on objects and on morphisms. Consider
the three examples above are symmetric monoidal. Monoidal cate- pSet, ˆ, t˚uq, for example, where ˆ as-
signs to a pair of sets pA, Bq their Carte-
gories come in other flavors too (braided, Cartesian, closed, Cartesian sian product A ˆ B. And given two
closed, closed braided,...), depending on which properties are satis- functions f : A Ñ B and g : A1 Ñ B1 ,
it assigns to the pair p f , gq the function
fied. f ˆ g : A ˆ B Ñ A1 ˆ B1 , which is
defined by: p f ˆ gqpa, bq :“ p f a, gbq.
The main takeaway here is that monoidal categories are the bread This is one example of the action of the
monoidal product on morphisms. More
generally, what f b g is depends on the
explicit definition of b.
and butter of many applied category theorists. One reason for this
is that monoidal categories provide a good setting in which to view
morphisms Ñ as physical processes and objects A, B, . . . as states. As a
non-technical example, let’s suppose A is a bunch of lemon meringue
pie ingredients while B is a fully-assembled-yet-unbaked lemon
meringue pie. We might view a morphism A Ñ B as the process of
mixing the raw ingredients together and then pouring the resulting
concoction into a pre-baked crust.
As it turns out, this pie example isn’t so silly after all. It’s one of the
motivating examples that Brendan Fong and David Spivak use in
their excellent book Seven Sketches in Compositionality: An Invitation
to Applied Category Theory to illustrate both the ubiquity and the sim-
plicity of monoidal categories. (If you haven’t read Seven Sketches yet,
you really must.) Below is a copy of their lemon meringue pie dia-
gram, where I’ve drawn our A and B on the left as input and right as
output.
Now that we’ve zoomed in, we can see that our process
is actually made up of a bunch of other processes! This isn’t too
surprising as there are several steps that go into preparing a lemon
pie: separating the eggs, making the lemon filling, filling the crust, and
so on. Fong and Spivak’s diagram illustrates just how those those
individual steps combine to form the single process prepare lemon
meringue pie. What’s neat is that we can describe these steps using the
language of monoidal categories! We’ll go into more detail later in
this section, but here’s a quick preview:
• The category’s composition ˝ corresponds to using one box’s out-

put wire as another box’s input wire. For example,
• The monoidal product b corresponds to stacking boxes on top of

each other. For example,
In other words, ˝ means “do the processes in series” while the

monoidal product b means “do the processes in parallel.” The re-
sulting picture is called a string diagram—a graphical representation
of a process (or equation of processes) in a monoidal category. I’ll
give more detail on how string diagrams work in a second. But first,
I’m reminded of something else about monoidal categories that I
want to tell you! So let me tell you this new bit of information, then
we’ll come back to string diagrams. This small digression will, in
fact, tie things together quite nicely. Bear with me.
Earlier, I mentioned that the word “symmetric” can be used as

an adjective for “monoidal categories”:
Similarly, there is another flavor of monoidal categories that we

should know about. This one will provide the main setting for our
example in Section 3.2:
I’ll explain.
Interlude: Compact Closed Categories and String Diagrams

Do you know what a finite dimensional vector space over R is? Then
you know what a compact closed category is! Or rather, you know an
example of one:
FVect is a compact closed category.
What makes that sentence true? Answer: every finite dimensional

vector space V has a dual space V ˚ “ hompV, Rq.
That’s it.
A compact closed category just the name for a monoidal category in

which every object has a dual!6 But what does “has a dual” mean? 6
Technically, every object must have a
left dual and a right dual. We need the
In other words, what makes a dual dual? Before I tell you the answer
distinction because not all monoidal
cite the definition, let’s think back to the category FVect, by way of categories are symmetric monoidal.
motivation, and let’s assume each vector space comes with an inner
product. In this case, there are two very important linear maps be-
tween the ground field—let’s say it’s R for now—and a vector space
V tensored with its dual:
ηV : R Ñ V b V ˚ eV : V ˚ b V Ñ R
In fact, there’s a nice fact from linear algebra, namely that once we fix
a basis te1 , . . . , en u for V then there is an isomorphism V – V ˚ . So We’ll need η and e for a computation
in Section 3.2, so it’s good to see what
let’s fix that basis (the standard one) and write V instead of V ˚ . Also,
they look like explicitly.
the subscript is a little cumbersome, so let’s drop it for now. So we
have two maps
η: R Ñ V bV e: V b V Ñ R (1)
The map η is called the unit 7 , and it assigns to every real number a 7
Note: this unit is not to be confused
vector in V b V, namely: with “monoidal unit”!
n
ÿ
ηp1q “ ei b ei (and extend linearly)
i “1
The map e is called the counit8 , and it assigns to every vector in 8

Note: counit is pronounced “coh-yew-
V b V a real number, namely: nit” not “cow-nit.” This is important.
˜ ¸
n
ÿ n
ÿ
e cij vi b w j “ cij pvi ¨ w j q where ¨ is the inner product
i“1 i “1
Intuitively, we can think of e as an evaluation map. That’s because

there is always a map V ˚ b V Ñ R given by evaluation. Indeed, if
v P V and f P V ˚ , then we can pair the two together to obtain f v P R.
And if we view f as a 1 ˆ n matrix and v as an n ˆ 1 matrix, then f v

is their inner product:
» . fi
.
” ı — . ffi
¨¨¨ f ¨¨¨ —–vfl “ a number
ffi
..
.
The e map just extends this linearly. That is, if we now have any
vector ij cij vi b w j in V b V, then e : V b V Ñ R is given by
ř
¨ ˛
ÿ ÿ
e ˝ cij vi b wi ‚ “ cij pvi ¨ w j q
ij ij
as above.
Finally, the unit η and counit e interact nicely with each other
because they satisfy some equations called the yanking equations,
which I’ll explain shortly. The bottom line is that all the above—
the maps η and e and the equations they satisfy—makes V ˚ into a
bona fide dual for V. The upshot is that compact closed categories
generalize these notions.
Definition 2.1. A compact closed category is a monoidal category

pC, b, 1q where for every object c in C there exists objects cl and cr
and morphisms The η maps are called the left and right
units, and the e maps are called the left
ηcl : 1 Ñ c b cl ecl : cl b c Ñ 1 and right counits.
ηcr : 1 Ñ cr b c ecr : c b cr Ñ 1
that satisfy the “yanking (or snake) equations” Here idc denotes the identity morphism
idc : c Ñ c.
pidc bel q ˝ pη l b idc q “ idc per b idc q ˝ pidc bη r q “ idc
(2)
pel b idcl q ˝ pidcl bη l q “ idcl pidcr ber q ˝ pη r b idcr q “ idcr
Yikes. What do these equations MEAN?
I’m glad you asked.
To answer, it’s time to revisit our previous discussion on string

diagrams! String diagrams are loved by applied category theorists far Ð In that sentence, I meant to convey
that string diagrams make life so much
and wide because they make life SO much easier. As we saw earlier, a
easier, but one may argue that applied
string diagram is a picture that represents morphisms in a monoidal category theorists also (are working to)
category C. For now let’s take that category to be FVect so that our make life so much easier.
objects are vector spaces V, W, . . .. In this case, the left and right dual
of space V is its vector space dual
V˚ “ Vr “ Vl
We’ll get to the yanking equations shortly, but first: If this document
has been your first introduction into string diagrams, then here is
THE KEY thing to know:
In category theory, we often draw an object as a dot ‚ and a morphism
as an arrow ‚ Ñ ˝. To draw a string diagram, just do the opposite!
The lemon pie diagram that we saw
(This goes back to Poincaré duality in topology.) To draw a string
on page 15 is an example of a string
diagram, draw an object as an arrow and a morphism as a dot or, even diagram!
better, a box.
With this small artistic adjustment, we can represent the monoidal

product b pictorially as well. The product of two spaces V b W
is drawn as two lines, side-by-side. A similar picture holds for the
product of two morphisms. Composition ˝ is represented by gluing
strings together.
And as we saw above, every object in a compact closed category such

as FVect has a dual. Its picture is given by an arrow pointing in the
opposite direction.
Another special object in a compact closed category is the monoidal

unit, for instance R in FVect. Because the unit is an object, it’s de-
picted as an arrow, too. People like to draw this arrow in a special
way, namely as the “empty” arrow. In other words, people don’t like
to draw an arrow. That’s because the monoidal unit R satisfies9 9
More generally, the monoidal unit 1 in
a monoidal category pC, b, 1q satisfies
VbR – V – RbV for all V, 1 b c – c – c b 1 for all objects c in C.
which suggests that the unit is “invisible.” But I like to draw it any-
way, shaded:
Now we are ready to get back to the yanking equations. Remem-

ber, part of the data of a compact closed category is that each object
V has left and right duals V r , V l together with morphisms
l
ηV : R Ñ V b Vl l
eV : Vl b V Ñ R
r
ηV : R Ñ Vr b V r
eV : V b Vr Ñ R
Again, to simplify the notation we’ll use the fact that that for vector
spaces, V ˚ “ V r “ V l . I’ll also drop the subscripts to keep things
clean.
Graphically, the ηs and es are drawn as below. The reason we
have two versions of each map is because the “information flow” can
either flow up or it can flow down.
Alternatively, some folks will rotate the

e and η diagrams by 90˝ clockwise and
counterclockwise, respectively, which is
the reason for their common nickname
of “cups and caps.”
Note: the direction of the (invisible)

arrow for the unit R can go either way.
The monoidal unit is always self dual!
10
Remember, the sentence “FVect is
symmetric monoidal” means there is
an isomorphism V b W – W b V for
And since FVect is a symmetric monoidal category, and since the left every pair of vector spaces V and W. In
and right duals are both V ˚ , there is really only one unit and one string diagram calculus, this means that
the order in which we draw our arrows
counit for vector spaces.10 doesn’t matter:
That is, in FVect
η “ ηr “ ηl and e “ er “ e l ,
and these are precisely the η and e defined on page 17! So in this For fun, verify that the unit and counit
maps on page 17 do indeed satisfy
example, the four yanking equations of (2) reduce down to just two: these two equations.
pe b idV q ˝ pidV bηq “ idV

(3)
pidV ˚ b eq ˝ pη b idV ˚ q “ idV ˚
Graphically, these equations can be represented as follows:
After yanking the strings taut, you’ll notice that information flows
rightwards in the first equation, while it flows leftwards in the second
equation.
Since we’re in a symmetric monoidal category, nothing changes if

we reverse the arrows in the pre-yanked strings. If, however, the
monoidal product b is not symmetric, then we obtain two more
diagrams.
This gives a grand total of four equations—the four displayed in

(2). And that’s what gives us a compact closed category. The category
FVect together with the unit and counit defined in (1) will make
another appearance in Section 3.2. In that same section, we’ll also
seen an example of a compact closed category that is not symmetric.
By the way, a key feature of FVect (and more generally, all sym-
metric compact closed categories) is that processes, i.e. morphisms,
V Ñ W are in bijection with states R Ñ W b V ˚ – V ˚ b W, which is
the special name given to morphisms whose domain is the monoidal
unit. This bijection is sometimes called process-state duality, and in
the context of FVect, it means we can view linear maps as vectors in a
tensor product11 and vice versa! 11
While a linear map R Ñ V ˚ b W is
not itself a vector in V ˚ b W, it can be
identified with one, namely with the
image of 1 in R! More generally, for any
finite-dimensional vector space A over
R, you can always think of hompR, Aq
as A itself, at least at the set level.
That’s because the forgetful functor
U : FVect Ñ Set is representable with
representing object R. In other words,
linear maps R Ñ A are in one-to-one
correspondence with the vectors in A,
viewed as elements of its underlying
set,
hompR, Aq – U A
This is completely analogous to how
functions t˚u Ñ X from the one-
point set to a set X are in one-to-one
correspondence with the elements in X,
hompt˚u, Xq – X
and is another manifestation of the
“probing” idea we saw in the margin
on page 9.
I like to think of it this way: when V and W are Rn and Rm with

the standard bases, process-state duality—taken together with the
margin note on the previous page—is akin to the observation that
matrices can be viewed as vectors and vice versa. That is, a linear
map Rn Ñ Rm has an m ˆ n matrix representation which can be
reshaped into an nm ˆ 1 column vector and then identified with a
vector in Rn b Rm . Conversely, there’s a way to identify a vector in
Rn b Rm with an nm ˆ 1 column vector that can be reshaped into an
n ˆ m matrix, which gives rise to a linear map Rn Ñ Rm . Here we’re making the mental identifi-
cation
Aside: There is, I think, a nice categorical way to piece this together.
linear map ù process
First note that there is a category MatpRq whose objects are natural
vector ù state
numbers n, m, . . . and whose morphisms f : n Ñ m are m ˆ n matrices
with real entries. The identity n Ñ n is the n ˆ n identity matrix which is closely related to the process-
and composition is given by matrix multiplication. This category is state duality seen in quantum physics.
It’s not quite the same, though—we’d
actually a compact closed category! The monoidal product on objects
need to replace MatpRq by the category
is given by multiplication n b m :“ nm and on morphisms is given of completely positive maps, another
by the Kronecker product of matrices. The monoidal unit is 1 P N. compact closed category! For more,
For the compact closed structure, each object is self-dual, n˚ :“ n, see Example 2.4 of “A categorical
and for each n P N the unit map ηn : 1 Ñ n2 is the n2 ˆ 1 column semantics for causal structure” by Aleks
Kissinger and Sander Uijlen, as well as
vector obtained by stacking the standard bases vectors e1 , . . . , en on
section 4.1.2 and chapter 6 of Picturing
top of each other. In other words, η is given by the Kronecker delta Quantum Processes by Bob Coecke and
function ηij :“ δij . The counit map en : n2 Ñ 1 is the 1 ˆ n2 row vector Aleks Kissinger.
obtained by taking the transpose of ηn . For example, if n “ 3 then
” ıJ
η3 “ 1 0 0 0 1 0 0 0 1 and e3 “ η3J . Since MatpRq is So
“ the entries of η3 are ‰
compact closed, it exhibits process-state duality, too: δ11 δ12 δ13 δ21 δ22 δ23 δ31 δ32 δ33 .
hompn, mq – homp1, nmq
This correspondence is precisely the reshaping of n ˆ m matrices

into nm ˆ 1 column vectors and vice versa, which can be verified
by using the unit and counit maps in a way analogous to the string
diagrams shown at the bottom of the previous page. To tie this in to
the remark about Rn and Rm in the previous paragraph, note that
–
there is a functor FVect Ñ MatpRq sending a vector space V to its
dimension dimpVq and a linear map f : V Ñ W to its corresponding
dimpWq ˆ dimpVq matrix representation, and it defines an equivalence
of categories! For details, see the discussion on page 30 as well as
Corollary 1.5.11 of Emily Riehl’s Category Theory in Context.
As we’ll see in Section 3.2, process-state duality pairs very nicely

with our intuition about language. There we’ll discover that a verb
can either be represented as a vector in a tensor product of vector
spaces or as a linear map, i.e. a process. In other words, a verb is an
action in the eyes of both grammar and mathematics!
Another digression: A conjunction with adjunctions?

If you’re familiar with adjunctions in category theory, then you might
wonder about this choice of naming and notation:
η ù “unit” e ù “counit”
Is it a coincidence that these two words are also used in the defini-
tion of an adjunction? NOPE. They are closely related. Specifically,
the data V, V ˚ , η, and e together with the yanking equations are an
instance of a categorical adjunction! I think this is a neat fact,12 so 12
which appears on the first page
let’s take yet another digression. Happily, it will tie in quite nicely of “Coherence for Compact Closed
Categories” by Kelley and LaPlaza.
with our discussion on string diagrams. We’ll begin by recalling the
definition of an adjunction.
Equivalently, L and R form an adjunc-
tion if for all objects c P C, d P D there is
Definition 2.2. An adjunction between categories C and D is a pair of
an isomorphism
functors –
homD pLc, dq ÐÑ homC pc, Rdq
L: C D: R
that’s natural in both c and d.
and a pair of natural transformations
There, idC denotes the identity func-
η : idC ùñ RL e : LR ùñ idD tor on C. It assigns each object and
morphism in C to itself.
called the unit and counit respectively, such that these two triangles
Here, L ˝ η denotes the natural trans-
commute:
formation whose components are of
L˝η η ˝R the form Lηc : c Ñ LRLc, while e ˝ L is
L LRL R RLR the natural transformation with compo-
nents e Lc : LRLc Ñ Lc. A similar story
e˝ L R˝e
id L idR holds for η ˝ R and R ˝ e. (As per the
L R margin comment on page 7, I’d prefer
to omit the composition symbol ˝, but
The adjunction is denoted L % R, and L is said to be left adjoint to R I’m writing it now for good reason, as
we’ll soon see!)
while R is said to be right adjoint to L.
Believe it or not, those commuting triangles—often called the

triangle identities—are closely related to the yanking equations in (3)!
Indeed, “these triangles commute” means that these two equations
hold:
pe ˝ Lq ˝ pL ˝ ηq “ id L
(4)
pR ˝ eq ˝ pη ˝ Rq “ idR
Now lets compare them to (3):
pe b idV q ˝ pidV bηq “ idV
pidV ˚ b eq ˝ pη b idV ˚ q “ idV ˚
Why are (3) and (4) so similar? What’s going on here? Is there a sense
in which a vector space and its dual form an adjunction?
Is V % V ˚ a thing?
Yes!
But to make sense of V % V ˚ we’ll need to venture into the world

of 2-categories. A 2-category is an appropriate setting in which to talk
about adjunctions, among other things. Here’s why. As we know
from the definition above, an adjunction consists of
i. some objects (categories) Recall: The data of an adjunction are

These form a category! functors between categories
ii. some arrows (functors)
L: C D: R
iii. some arrows between the arrows (natural transformations) and natural transformations
η : idC ùñ RL e : LR ùñ idD
You’ll notice that the objects and the arrows themselves form a cate-
gory, namely Cat, the category of all categories. The objects of Cat are
categories and the morphisms are functors.
Nice.
It’d be even nicer, though, if the natural transformations were also

part of the data. That is, it’d be super nice if the threesome itself con-
stituted a known categorical construction. But as it stands, it doesn’t.
There is no room for a notion of “arrows between arrows” in the
definition of a category.
So what do we do?
We expand the definition. Literally. We add an extra dimension,

which results in a 2-category. That is, a 2-category consists of What’s more, in any 2-category there
is a composition rule for 2-morphisms
i. objects, now called 0-cells ‚, ˝, . . . just like there is for 1-morphisms
in an ordinary category! In fact, in
ii. morphisms that go between objects, which are the usual arrows, a 2-category we require that the set
homp‚, ˝q be more than a set. We ask
but now we’ll call them 1-cells ‚ ÝÑ ˝ that it be a category itself! Its objects
are 1-cells ‚ Ñ ˝, and its morphisms
iii. morphisms that go between 1-cells, which are not surprisingly
are 2-cells ‚ ˝ There is also
an identity 1-cell id‚ for each 0-cell ‚
called 2-cells ‚ ˝
and there is an identity 2-cell idid‚ for
each id‚ . Confusingly, both of these
As you might guess, the quintessential example of a 2-category is Cat, identity morphisms are sometimes
denoted as ‚. And of course, there are
where the the usual identity and associativity
axioms, though I won’t write them here.
i. 0-cells are categories C, D, . . .
F
ii. 1-cells are functors C ÝÑ D
iii. 2-cells are natural transformations C D
G
So a 2-category is a good generalization of the relationship we see

exhibited among categories, functors, and natural transformations.
Having generalized this trio, it becomes very easy to talk about “ad-
junctions” in any 2-category. Parallel to Definition 2.2, we might lay
down the following proposed definition:
Definition (proposed). An adjunction between 0-cells ‚ and ˝ is a pair
of 1-cells l and r,
l: ‚ ˝: r
and a pair of 2-cells η and e, called the unit and counit respectively,
13
If you do a Google search for “defini-
id‚ id˝ tion of 2-category” you’ll soon find that
2-cells can be composed in two ways:
‚ η ‚ and ˝ e ˝ “vertically” and “horizontally.” I didn’t
mention this earlier, but now’s a good
r˝l l˝r time to do so. Suppose we have three
1-cells from ‚ to ˝ and 2-cells η and e as
such that these two triangles commute shown below on the left,
l˝η η˝r η
l l˝r˝l r r˝l˝r ‚ ˝ ù ‚ e˛η ˝
e
e˝l r˝e
idl idr
l r then vertical composition ˛ gives a 2-cell
e ˛ η as shown above on the right. This
i.e. such that the following equations hold13 is composition along a common 1-cell Ñ.
On the other hand, given four 1-cells as
pe ˝ lq ˝ pl ˝ ηq “ idl and pr ˝ eq ˝ pη ˝ rq “ idr shown below left,
where l ˝ η :“ idl ˝η, and similarly for e ˝ l and so on. We’ll say l is ‚ η ‹ e ˝ ù ‚ e˝η ˝
a left adjoint of r, and r is a right adjoint of l, and we’ll denote the
adjunction by l % r. horizontal composition ˝ gives a 2-cell
e ˝ η as shown above right. This is
composition along a common 0-cell ‹.
Alright, fine. But what does this have to do with vector spaces? Moreover, the triangle identities involve
both compositions. That is, the actual
equations are
The answer lies in the following neat fact.
pe ˛ lq ˝ pl ˛ ηq “ idl and pr ˛ eq ˝ pη ˛ rq “ idr
Neat Fact: Every monoidal category pC, b, 1q can be viewed
Take note of the diamonds vs. the
as a 2-category! circles!
Er, actually, I shouldn’t spread rumors.
Neat Fact: Every monoidal category pC, b, 1q can be viewed

as a 2-category!
Here’s the correct statement: 14

In a 2-category, the composition of
1-cells is associative, i.e. f pghq “ p f gqh
Neat Fact: Every monoidal category pC, b, 1q can be viewed for any composable triple of 1-cells
f , g, h. In a bicategory, however, we
as a bicategory! weaken this. Instead of asking for
equality, we ask for the existence of
A bicategory is basically a 2-category—the data is completely the an invertible 2-cell f pghq ðñ p f gqh.
same. There are 0-cells, 1-cells, and 2-cells. The only difference is As we’ll see below, the category FVect
gives rise to a bicategory rather than
what’s in the margin.14 So any monoidal category pC, b, 1q gives rise a 2-category because the two vector
to a bicategory C where the spaces V b pW b Uq and pV b Wq b U
are not equal, but there certainly is a
linear isomorphism
–
V b pW b Uq ÐÑ pV b Wq b U!
i. only 0-cell is the category C
ii. 1-cells are the objects of C; composition is b
iii. 2-cells are the morphisms of C; composition is composition ˝ in C
Therefore it makes sense to talk about 1-cells in C (i.e. objects in C)

having adjoints! And it makes sense to talk about 2-cells in C (i.e.
morphisms in C) being units and counits of the adjunction, vis-a-vis
our Proposed Definition! In particular, this is true of the symmetric
monoidal category pFVect, b, Rq. It gives rise to a bicategory F Vect
where the
i. only 0-cell is the category FVect
ii. 1-cells are vector spaces; composition is the tensor product b
iii. 2-cells are linear maps; composition is the usual composition ˝
So there is an adjunction of vector spaces V % W whenever the

conditions of our Proposed Definition hold. Of course, those condi- You’ll notice that the monoidal unit R is
tions hold precisely when W “ V ˚ and η and e are defined as in (1). taking the place of id‚ in the Proposed
Definition. Indeed, id‚ and R are
Explicitly:
comparable since both are 1-cells that
act as an identity on other 1-cells: For
There is an adjunction V % V ˚ in the bicategory F Vect since there are f
linear maps all 1-cells ‚ ÝÑ ˝ in C
η
RÝÑV ˚ b V
e
and V b V ˚ ÝÑR f ˝ id‚ “ f “ id˝ ˝ f
so that the following triangles commute and for all vector spaces V in F Vect,
V b R – V – R b V.
Vbη ηbV ˚
V – VbR V b V˚ b V V˚ – R b V˚ V˚ b V b V˚
ebV V˚ b e
idV idV ˚
RbV – V V˚ b R – V˚
i.e. so that the following equations hold On the leftmost triangle, the notation
V b η denotes the linear map
pe b idV q ˝ pidV bηq “ idV and pidV ˚ b eq ˝ pη b idV ˚ q “ idV ˚
idV bη : V b R Ñ V b V ˚ b V
and these are precisely the string diagram equations shown in the that appears in the first equation. A
chart on page 21. similar statement holds for e b V, etc.
Also, take note of the different symbols
b and ˝ and compare them with the
Voila! diamond ˛ and circle ˝ in the margin
on the previous page.
Finally, notice that the above holds for every vector space V in
FVect. On the other hand, there are certainly 2-categories in which
not every 1-cell is dualizable, i.e. has an adjoint. Take Cat for instance!
Not every functor is part of an adjunction. There is, however, a spe-
cial name given to those bicategories C that do arise from a monoidal The punchline for this section is that
monoidal categories are an appropriate
category C and in which every 1-cell has an adjoint.
framework for stacking things together,
and the calculus of string diagrams
That name is compact closed. allows us to replace complicated, messy
equations by simple, neat pictures. In
Section 3, we’ll see two examples of
how this can be put into practice.
2.2 Decorated Cospans

A second construction that appears in some work within applied
category is the decorated cospan. In any category, a diagram that looks
like
A Ñ C Ð B.
is called a cospan. In the next section, we’ll only consider the case
when A, B, and C are finite sets and the arrows are functions between
them. A decorated cospan is a cospan where the middle set C has
been endowed with some extra structure. That’s the intuitive defini-
tion, though I’d like to postpone a more precise definition until the
next section.
Now you might think it strange to give a name to a simple dia-
gram like A Ñ C Ð B, but cospans come in handy quite often! For
instance, if for some reason you can’t possibly hope to find a mor-
phism between objects A and B, a common technique15 is to instead 15
I learned this from Brendan during
look for a “larger” object C that “contains” both A and B. Then al- the 2018 ACT workshop. Thanks,
Brendan!
though you don’t have maps between A and B, you do have maps
A Ñ C Ð B. In that case, your cospan is the next best thing.
Admittedly, this section is bite-sized compared to the behemoth

on monoidal categories that we just finished, but that’s not because
decorated cospans are any less important! In fact, Brendan Fong
developed the theory of decorated cospans as part of his PhD the-
sis “The Algebra of Open and Interconnected Systems”, which has
served as the foundation for much progress in applied category the-
ory, as I mentioned earlier. But in these notes, we’ll only use the
cospan construction in our brief discussion on chemical reaction net-
works in Section 3.1. On the other hand, we will need the language
of monoidal categories in both Sections 3.1 and 3.2. In fact, as we’ll
soon see, decorated cospans themselves form a monoidal category!
2.3 Further Reading

For more on monoidal categories and string diagrams:
• Read Chapters 3 and 4 of Picturing Quantum Processes by Bob Co-

ecke and Aleks Kissinger. There you’ll also find more information
on the interpretation of morphisms in a monoidal category as
processes and objects as systems.
• Take a look at TheCatsters videos on string diagrams, by Euge-

nia Cheng and Simon Willerton. On second thought, their entire
collection of videos is great. Go watch them all!
• If you like 8-categories, you’ll be delighted to know that a ver-

sion of string diagrams (affectionately called “strictly undulating
squiggles”) and the yanking equations (!) make an appearance in
chapter 8 of Elements of 8-Category Theory, a new book on model-
independent 8-category theory by Emily Riehl and Dominic Ver-
ity.
For more on decorated cospans:
• Read “Decorated Cospans” a blog post by John Baez on the n-

Category Café.
• Read Chapter 6 of Seven Sketches in Compositionality by Brendan

Fong and David Spivak. In Section 6.1, the authors give the follow-
ing bit of motivation:
...we produce a certain monoidal category—namely that of cospans

in [a category] C, denoted CospanC —that can conveniently package
C’s colimits in terms of its own basic operations: composition and
monoidal structure. In summary, the first part of this chapter is
devoted to the slogan ‘colimits model connection.’ (emphasis theirs)
As we’ll see in Section 3.1, objects in CospanC are cospans in C

and a morphism between two cospans is given by a construction
called a colimit. Like the composition ˝ and product b in a general
monoidal category, a colimit is a categorical construction that al-
lows you to connect things together. But for the sake of “time” (i.e.
so that this document doesn’t accidentally turn into a book...), I’ll
assume familiarity with colimits. But if you’d like to an intuitive
introduction of colimits, as well as their dual construction, limits, I
recommend that you
• Take a look at “Limits and Colimits (Part 1)” a blog post on

Math3ma. Also see chapters 3 and 6 of Seven Sketches and chap-
ter 3 of Category Theory in Context by Emily Riehl.
3 Two Examples
Having taken a leisurely stroll through two themes (functorial se-
mantics and compositionality) and two constructions (monoidal
categories and decorated cospans) within applied category theory, it’s
time to see them come to life in two examples. As mentioned in the
introduction, we’ll walk through the first example—chemical reaction
networks—relatively quickly. There are several excellent resources
available online, including John Baez’s expositions on the n-Category
Café as well as on his personal webpage. (I’ve included a few links to
these in Section 3.3.) Afterwards we’ll take a longer stroll through the
second example—natural language processing—in Section 3.2.
3.1 Chemical Reaction Networks

The first example comes from a paper by John Baez and Blake Pol-
lard called “A Compositional Framework for Reaction Networks.”
Specifically, they provide a compositional framework for modeling
chemical reaction networks. A chemical reaction network is, well, a
networks of chemical reactions. And a chemical reaction is exactly
what you think it is. It’s what you learned back in high school: You
start with some reactants and some products, and there’s a chemical
process that takes one to the other.
What’s nice is that these reactions can be depicted graphically:
Of course, you can imagine that there might be lots of various reac-
tants, products, and chemical processes. The corresponding network
would then be a (possibly huge) collection of these graphs stacked
side-by-side, perhaps with connecting edges and loops and so on.

For instance, this chemical reaction network made a cameo appear-
ance in Baez’s 2016 talk “The Mathematics of Networks”:
Graphs such as these are examples of Petri nets. A Petri net is es-
sentially a bipartite directed (multi)graph that allows us to visually
represent reactions, though they are used outside of chemistry as
well.
But if we do wish to model chemical reactions, then an important
thing we’d like to account for is the rate at which one or more chemi-
cals change over to another. A Petri net with rates included is called,
appropriately, a Petri net with rates. More specifically, it’s a bipartite
directed graph whose two types of vertices are called places, which
represent chemical species, and transitions, which represent chemical
reactions. Moreover, each transition τi is assigned a rate ri , a positive
real number that describes how fast or how likely it is for τi to occur.
These rates then allow us to write down differential equations that
describe the system. A Petri net with rates is thus a pictorial repre-
sentation of a set of differential equations that describe a system. So,
for instance, if you did watch Baez’s “The Mathematics of Networks”
talk then this example will look familiar:
It tells us that, for example, substances with concentrations A and

B combine and produce a substance with concentration C at a rate
proportional to r1 . The differential equations you see are due to the
law of mass action, which says that the rate with which a chemical
reaction will occur is equal to its rate constant ri multiplied by the
product of the concentration of the reactants, i.e. the concentration of
the “inputs” of the reaction.
By the way, the rates themselves could change with time, which
might suggest the presence of a dynamical system. What’s more, a
system such as the above could potentially interact with its environ-
ment, which is to say there might be some quantities that flow in and
some quantities that flow out, resulting in an open Petri net with
rates:
These quantities can be incorporated into the equations, too, result-

ing in an open dynamical system. As Baez and Pollard summarize,
the goal is to use these observations “to build up a reaction net-
work from smaller pieces, in such a way that its rate equation can
be determined from those of the pieces.” This is what is meant by a
compositional framework, and is a prime is example of the principal of
compositionality mentioned in Section 1.2. What’s more, a key step
towards achieving this goal is given by functorial semantics! That
is, we start by thinking of a Petri net as syntax and a set of differen-
tial equations as semantics. And if we have a collection of Petri nets
that model a very large network, then—guided by the principal of
compositionality—we would like to compose them by gluing graphs
together, and we would like to aggregate them by stacking graphs
on top of each other. In other words, we hope that Petri nets form a
monoidal category! Similarly, one would hope that there is a sense
in which dynamical systems form a monoidal category so that differ-
ential equations can be “composed” and “aggregated” as well. One
would also wish for a monoidal functor Petri nets Ñ dynamical systems.
That’s a lot of wishes, but amazingly they all come true, for this is
precisely what Baez and Pollard proved in their paper! But how ex- 16
This is the catch-all phrase for a
network that interacts with its environ-
actly? How was it all formalized? The key is the decorated cospan
ment so that stuff can either flow in or
construction of Brendan Fong that we mentioned in Section 2.2. flow out
(What’s amazing is that Fong’s construction is general enough to
model other open reaction networks16 as well! But more on that
later.)
The reason the arrows from X and

Y point in is that X and Y might be
thought of as “leftputs and rightputs”
rather than as inputs and outputs. In
other words, you’d like the freedom
to think of things as flowing either
in or out of either end. For example,
a physical pipe doesn’t know the
different between left and right. Water
can flow in or out at either direction. As
Baez notes, “The main reason for these
designations is to remember that when
we screw together two pipes, we attach
A cospan in the category of finite sets, for example, is just a diagram the output of the first to the input of the
of the form X Ñ V Ð Y, where we’re meant to think of V as the second.”
set of places (i.e. chemical species) in the Petri net. To account for the
edges in the graph, we ask that V is “decorated” with extra structure,
namely source and target maps from the set of edges. This results in
a decorated cospan, and Fong proved that these constructions form
17
A pushout is a type of colimit, a
a category! That is, there is a category where objects are finite sets major construction in category theory
X, Y, . . ., and a morphism X Ñ Y is a decorated cospan whose feet that‘s a bit like like mathematical glue.
are X and Y. Composition is given by the pushout17 , which amounts Anytime you mush two mathematical
objects together—like the graphs in the
to gluing graphs together. (This composition is only associative up picture—you’ve probably got a colimit
to isomorphism, so the morphisms are really isomorphism classes of construction. More intuition behind
colimits and their dual construction,
cospans. Also, the identity X Ñ X is the empty graph.) limits, can be found on Math3ma.
What’s more, Fong showed that this category has a symmetric

monoidal structure by stacking graphs on top of each other, i.e. by A hypergraph category is a sym-
metric monoidal category in which
taking their disjoint union. In fact, it’s compact closed and also a
every object has a special commutative
hypergraph category! Frobenius structure. This allows more
In summary, Fong’s constructions quickly give rise to key results, freedom (i.e. messiness) when compos-
ing morphisms, reflecting the messiness
which I’ll summarize here. The first is that the syntax category of of most network diagrams!
Baez and Pollard is indeed a category:
Theorem 3.1 (Baez, Pollard). There is a symmetric monoidal category

Petri where
• objects are finite sets X, Y, . . .
• a morphism X Ñ Y is a open Petri net with rates, i.e. a cospan Really, it’s an isomorphism class of
cospans. Also, you’ll notice that in
Theorem 12 of Baez and Pollard’s “A
V
Compositional Framework,” their
i o syntax category is something called
RxNet. That stands for the category
X Y of open reaction networks with rates. An
open reaction network is very nearly
together with a Petri net with rates whose places are comprised of V. the same as an open Petri net, though
I’m glossing over this a bit.
The next corollary provides the same statement for the semantics
category:
Theorem 3.2 (Baez, Pollard). There is a symmetric monoidal category

Dynam where
• objects are finite sets X, Y, . . .
• a morphism X Ñ Y is an open dynamical system, i.e. a cospan Again, it’s really an isomorphism class
of cospans. And again we can think of
V as the set of all places in a Petri net
V
where, as before, there may be a real
i o number ri attached to each vertex than
can vary with time. The description
X Y of how these things vary in time is
precisely a vector field on RV .
together with a smooth vector field on RV .
Finally, another result of Baez and Pollard shows the existence of a

symmetric monoidal functor ˝ : Petri Ñ Dynam from the syntax to the
semantics.
Theorem 3.3 (Baez, Pollard). There is a symmetric monoidal functor

˝ : Petri Ñ Dynam sending any open Petri net with rates to the correspond- But see Theorem 18 of Baez and Pollard
ing open dynamical system. where, since RxNet is used in lieu of
Petri, the symmetric monoidal functor
RxNet Ñ Dynam is a slightly different
The upshot is that functoriality and monoidality gray boxing functor.
˝p f ˝ gq “ ˝ f ˝ ˝g
˝p f b gq “ ˝ f b ˝g
tell us that if you want to understand the open dynamical systems

of the composite (or tensor product) of two open Petri nets, then you
just have to find the open dynamical systems of each one and then
compose (tensor). This is exactly the principle of compositionality:
to determine the behavior of a big complicated thing, you need only
understand the behaviors of its components, and then assemble them
together. And by the way, this works for many other kinds of net-
work graphs, not just Petri nets. It’s all part of Baez’s larger body of
work on a general categorical framework for a theory of networks
which encompasses electrical circuits, Markov processes, signal-flow
graphs in control theory, and more!
This rapid tour through chemical reaction networks is only one

way that compositionality, functorial semantics, and monoidal cat-
egories are being used in applications. The next example gives a
second way: natural language processing.
3.2 Natural Language Processing

At long last, we’ve made it to our second application of category
theory—natural language processing! It is, simply put, a branch
of artificial intelligence that aims to train computers to understand
human language. What’s nice is that computers can understand
meanings of words (through models like Word2vec, for instance18 ) 18
In the literature, these models are
and computers can understand grammar (through parts of speech often called distributional models of
meanings.
tagging, for instance19 ). But what’s not-so-nice is that computers 19
These are often called symbolic or
aren’t too good at understanding meanings of sentences and longer compositional models of meaning.
bodies of text.
In 2010, Bob Coecke, Mehrnoosh Sadrzadeh, and Stephen Clark
sought to address this problem in “Mathematical Foundations for a
Compositional Distributional Model of Meaning.” In this paper, the
authors rely heavily on the principle of compositionality—the idea
that the meaning of a sentence can be determined by the meanings of
its individual words together with the grammatical rules for combin-
ing them. So if a computer can understand meanings of individual
words and if it can understand grammatical rules, then then only
thing it needs help with is knowing how to combine them to form
a meaningful whole. And that’s where the category theory comes
in! Guided by functorial semantics, Coecke et. al. model natural lan-
guage as a (monoidal) functor between compact closed categories
grammar Ñ meanings of words
This functor assigns a grammar type to a word, and the monoidal

structures provide a way to combine the meanings of those words
(and their grammar types) to form a sentence, whose meaning can be
determined via the principal of compositionality.
in the remaining pages, we’ll dive into the details by answering

the following questions:
i. (the syntax category) How can we make sense of grammar,

mathematically? Specifically, how does grammar form a compact
closed category?
ii. (the semantics category) How can we make sense of meanings

of words, mathematically? That is, how do meanings of words
form a compact closed category?
iii. (the functor) How is the functor grammar Ñ meanings of words

defined and how does it allow one to determine the meaning of
a full sentence?
Let’s start by answering the first question.

The Syntax Category: Pregroup

If you were to develop your own
categorical-compositional-distributional
Following the work of Coecke et. al., we can model grammar al-
model (often called DisCoCat mod-
gebraically via a pregroup, a construction due to mathematician els) of meaning, you might wish to
Joachim Lambek in the early 1990s. Informally, a pregroup is cooked work with a construction other than
pregroups. And that’s fine. But the
up from the following recipe: nice thing about a pregroup is that (as
we’ll soon see) it has the exact same
poset + monoid + “duals” = pregroup categorical structure as our semantics
category, namely compact closure.
That means we consider a functor from
a pregroup into the semantics cate-
gory that preserves the compact-closed
structure. More generally, then, you
might define a DisCoCat-type language
model to be any monoidal functor
C Ñ meanings of words where C is any
compact closed category accounting for
grammar.
In other words, a pregroup is
• a poset pP, ďq
• that has a multiplication (we’ll denote it by juxtaposition) that’s

compatible with the partial order, i.e. if p ď q then ap ď aq and
pa ď qa for all a P P,
• together with a unit 1 satisfying 1p “ p1 “ p for all p P P,
• and moreover each element p has both a left dual pl and a right
dual pr with maps
el ηl er ηr
pl p ď 1 ď ppl and ppr ď 1 ď pr p.
that are required to satisfy the yanking equations in (2).
I’ve referred to the inequalities a maps (and have labeled them as

such) because they are actually morphisms in a category! Indeed,
every poset pP, ďq can be viewed as a category: an object is an el-
ement in P and there is an arrow p Ñ q if and only if p ď q. (In
particular, there is at most one arrow between any two elements in
a poset.) Composition is given by transitivity: if p ď q and q ď r
then p ď r, and associativity is immediate. Also, every element has
an identity arrow since the partial order is reflexive: p ď p for all
p P P. So because a pregroup is a poset, we may also view it as a
category. Therefore I’ll now draw an arrow Ñ in lieu of the partial
order ď. Moreover, a pregroup is really a poset-with-extra-structure
and therefore we may view it as a category-with-extra-structure. Not
surprisingly, given the reappearance of the yanking equations, that
extra structure is compact closure! In summary, a pregroup is an ex-

ample of a compact closed category. It is, in particular, a compact
closed category that is not symmetric. Indeed, those four inequalities
above are really the unit and counit maps
el ηl er ηr
pl p ÝÑ 1 ÝÑ ppl and ppr ÝÑ 1 ÝÑ pr p.
discussed in Section 2.1, and the yanking equations (2) amount to the
following:
1¨ η r e r ¨1
p “ p ¨ 1 ÝÑ ppr p ÝÑ 1 ¨ p “ p
η l ¨1 1¨ η l
p “ 1 ¨ p ÝÑ ppl p ÝÑ p ¨ 1 “ p
η r ¨1 1¨ e r
p “ 1 ¨ pr ÝÑ pr ppr ÝÑ pr ¨ 1 “ pr In first equality of the third line we’re
rewriting pr as 1 ¨ pr rather than pr ¨ 1
because neither of the ηs nor es provide
1¨ η l e l ¨1 a map 1 Ñ ppr . Similarly for the
p “ pl ¨ 1 ÝÑ pl ppl ÝÑ 1 ¨ pl “ pl
last line, write pl “ pl ¨ 1 rather than
pl “ 1 ¨ pl since there’s no map 1 Ñ pl p.
Let’s look at two examples of pregroups. The first is an arith-

metic example, which will help to get our feet wet. The second is a
grammatical example, which is used in the DisCoCat model of Co-
ecke, Sadrzadeh, and Clark.
Example 3.4. The set
t f : Z Ñ Z | f is monotone and unboundedu
is a pregroup. The partial order is given pointwise: f ď g if and only Fun fact: the pair p f l , f q forms a special
if f n ď gn for all n. The monoid multiplication is given by function kind of categorical adjunction called a
Galois connection since it satisfies
composition f ¨ g :“ f ˝ g. The monoidal unit is idZ . Given such a
function f , its left and right duals are given by f ln ď m ô n ď f m.
Indeed if n is even, then n{2 ď m ô
f l n :“ mintm P Z | n ď f mu f r n “ maxtm P Z | f m ď nu. n ď 2m. And if n is odd, then pn `
1q{2 ď m which means n ` 1 ď 2m
For example, if f m “ 2m, then which is true iff n ď 2m. Similarly, the
pair p f , f r q forms a Galois connection
$ $ since
&n if n is even, &n if n is even,
f ln “ 2 fr “ 2 fn ď m ô n ď f r m.
n ` 1 % n ´1
%
2 if n is odd 2 if n is odd.
Indeed, if n is even then 2n ď m ô
n ď m{2. And if n is odd, then 2n ď m
In short, f l n “ t n` 1 r n
2 u and f n “ t 2 u.
means n ď m{2 which is true iff
n ď pm ´ 1q{2.
You can find this example in “Iterated Galois Connections in
For a couple of great introductions to
Arithmetic and Linguistics” by Lambek, which appears in the Springer Galois connections (They are super cool
book Galois Connections and Applications. You’ll also find mention of it and appear in lots of places in math!)
take a look at Lecture 4 of John Baez’s
in the “Mathematical Foundations” paper of Coecke et. al. online course on applied category
theory as well as Section 1.5 of Seven
Sketches by Fong and Spivak.
While arithmetic is fun, this next example is the one we’re most
interested in.
Example 3.5. Given any finite poset X, we can construct the free pre-
group generated by X, denoted PregX. For a simple example, suppose
X “ tn, su whose elements we’ll think of as basic grammar types:
n is the type of a noun and s is the type of a (declarative) sentence.
Elements of Pregtn, su are concatenations of the letters n and s and
their left and right duals and iterations of those duals and so on. For
example, some grammatical types in Pregtn, su are:
The strings of letters are called compound types, and there is a mor-
phism a Ñ b between compound types if and only if a can reduce to b
by application of one or more of the counit maps er and el .
Consider a banana, for example. It has type n, of course, while
the adjective yellow has type nnl . The reason that adjectives have
grammar type nnl is that an adjective can always be paired on the left
with a noun, resulting in a new noun—e.g. yellow banana.
yellow banana
nnl n
Indeed, to verify that the grammar type of yellow banana is n, we start

by concatenating the types of the individual words to obtain nnl n. The dot ¨ in el ¨ 1n is meant to suggest
“apply el to nnl while simultaneously
Then we apply the counit map el : nnl Ñ 1 together with the identity
applying 1n to n.”
map 1n : n Ñ n (this is given to us by the reflexivity axiom of posets:
n ď n) to see that nnl n reduces down to n:
This tells us that the phrase yellow banana has grammar type n.
That’s good. A yellow banana is a noun!
In light of this discussion on yellow bananas, you might enjoy tak-
ing a few seconds to think about why nr snl represents the grammar
type of transitive verb.
<ponder> ... </ponder>
A transitive verb is a word that accepts a noun on the right and an-
other noun on the left such that the resulting phrase is a full sen-
tence. Since we like bananas, here’s another fruit-based example:
bananas are fruit

n nr snl n
To determine the grammar type of this phrase, we concatenate the

grammatical types of the individual words and then apply the counit
maps to reduce, as before:
Here’s that same reduction written out step-by-step. For clarity, I’ll
indicate the concatenation with a dot:
e r ¨1 ¨1 1 ¨e l
nnr snl n “ nnr ¨ s ¨ nl n ÝÑ
s n
1 ¨ s ¨ nl n “ s ¨ nl n ÝÑ
s
s¨1 “ s
In words, we’ve used the counit maps to reduce the concatenation of

the grammar types for bananas are fruit to the letter s, which confirms
that “bananas are fruit” is indeed a grammatically correct sentence.
More generally, for any a, b P Pregtn, su, we draw a morphism a Ñ b
if and only if a can be reduced to b in a similar fashion.
Alright, that’s (a very condensed version of) the pregroup story!

To summarize, the language model of Coecke et. al. amounts to a
structure-preserving functor
syntax Ñ semantics
In this section, we’ve just shown that the syntax category is taken to
be a pregroup freely generated on a finite set of basic grammar types,
i.e. syntax “ PregX. Let’s move on to semantics now.
The Semantics Category: Vector Spaces

As stated in the introduction to Section 3.2, computers are able to
understand meanings of individual words pretty well. That’s because
computers understand numbers! For example, a great way to inform
a computer of the meaning of the word banana is to represent banana
by a number and then give that number to the computer.
What number?
Well, it’s not exactly a number. It’s an array of numbers—a vector.
Okay, what vector?
The answer is simple, though I’d like to motivate it by sharing the

following theorem.
Theorem (The Yoneda Lemma for Linguistics). You shall know a word 20
Firth, J. R. A synopsis of linguistic
by the company it keeps. theory, 1930–1955. In Selected Papers
of JR Firth, 1952–59 (ed. J. Firth and F.
Proof. John Firth20 Palmer). Indiana University Press.
Okay, so it’s not a theorem. But it is a great quote! Firth’s idea

is that words that appear in similar contexts will have similar meaning. Are you wondering why I’ve referred to
In the linguistics community, this is referred to as the distributional Firth’s idea as the Yoneda Lemma? To
find out why, I recommend reading up
hypothesis. So you might imagine that apple is more similar to banana on the Yoneda Perspective.
than it is to puppy since apples and bananas often occur near words
such as sweet, snack, green, eat, etc., whereas puppy occurs more
often near words such as pet, cute, furry, bark, and so on. As another
example, you might not know what the word yegg means (or perhaps
you do), but you can probably infer it from this sentence:
The cops grabbed him and another yegg for a Philadelphia store burglary.21 James Lardner and Thomas Reppetto,
21
NYPD: A City and Its Police, 2000
So we can represent the meaning of a word by a vector. This is

often called a distributional model of meaning. But what, exactly, is the
assignment word ÞÑ vector? Suppose we have a fixed corpus—your
favorite book, say. Start by choosing a set of so-called context words
tw1 , . . . , wn u. This can be every word in the corpus or some subset of
it. By representing each wi as the ith standard basis vector
ith spot
hnlj
wi “ p0, . . . , 1 , . . . , 0q
we obtain a basis tw1 , . . . , wn u for a vector space V. Then any word

w in the corpus has a vector representation given by a linear combi-
nation of the context words
n
ÿ
w“ ci w i
i “1
The coefficients ci are real numbers that indicate the number of times
that w occurs near22 wi in the corpus. 22
You can decide what “near” means.
That is, the context of w is the set of
Here’s an example. Suppose we’re reading a book that contains
words within k words of w, where k “ 1
the words or 2 or 3 or whatever you like.
tsweet, green, furryu
Let’s choose them to be our context words and make the assignment
so that
» fi » fi » fi
1 0 0
sweet “ –0fl green “ –1fl furry “ –0fl
— ffi — ffi — ffi
0 0 1
Then if banana, puppy and fruit are also words in our book, we might
have something like
» fi » fi » fi
21 8 43
banana “ – 9 fl puppy “ – 1 fl fruit “ –19fl
0 32 0
In other words, we’ve used data from the corpus to embed these
words as vectors inside of a three-dimensional vector space. This
prompts us to say that the meaning of the word banana is the vector
p21, 9, 0q, the meaning of puppy is p8, 1, 32q, and the meaning of fruit
is p43, 19, 0q.
And this works! That is, you can feed distributional models into
your computer, and they’ll ace the word-similarity portion of your
SAT exam. Or your can compute the dot product between words, and
you’ll find that vectors are closer together precisely when the words
they represent have the same meaning. It’s all familiar territory for
NLP practitioners. The semantics category for Coecke et. al. is thus
the category of finite dimensional vector spaces over R. That is,
meanings of words “ FVect
Unfortunately, the distributional model does not work for

sentences. The same sentence rarely occurs twice in a given docu-
ment, therefore we can’t follow the same procedure above. This is
where category theory can help lend a hand. In light of the princi-
ple of compositionality, the meaning of a sentence should be able
to be computed given the meanings of its individual words and the
rules of grammar for combining them. And we can pair meanings-

of-words with grammatical types via a map from syntax (grammar) to
semantics (meanings of words), i.e. via a functor
F : PregX Ñ FVect
where X is a finite set of basic grammar types. In fact, as we know

from Section 2.1, both FVect and PregX are compact closed categories,
so we’ll ask that F be a strong monoidal functor—one that preserves
the compact closed structure. That’s the gist behind the categorical
compositional distributional model of Coecke et. al.
But how is F actually defined? Let’s talk about that next.
The Functor: Syntax Ñ Semantics

In this section, we’ll give an explicit description of the functor
F : syntax Ñ semantics
or more specifically,
F : PregX Ñ FVect
For simplicity, let’s take X “ tn, su as we did before. Now to define
a functor, we need simply to say what it does on objects and mor-
phisms. So let’s do that. On objects,
• F assigns to the noun type n a vector space N :“ Fn, which we’ll

call a noun space
• F assigns to the sentence type s a vector space S :“ Fs, which we’ll

call a sentence space
and on morphisms
r Fr
• F assigns to a type reduction a ÝÑ b a linear map Fa ÝÑ Fb that
sends the vector corresponding to a word or phrase of type a in Fa
to the vector corresponding to a word or phrase of type b in Fb.
Moreover, asking that F preserve the compact closed structure means

that
• units and counits in Pregtn, su map to units and counits in FVect
e.g. given n P Pregtn, su,
ηr ηl er el
n
Fp1 ÝÑ nr nq “ Fp1 ÝÑ
n
nnl q “ η N and Fpnnr ÝÑ
n
1q “ Fpnl n ÝÑ
n
1q “ e N
where η N : R Ñ N b N and e N : N b N Ñ R are the linear maps

that we defined on p. 17. A similar idea holds if we replace n by
any element of Pregtn, su.
• duals map to duals
e.g. Fnr “ Fnl “ N ˚ . But our vector spaces are finite dimensional
and so N ˚ – N and therefore Fnr “ Fnl “ N.
• a compound type is assigned to a tensor product of vector spaces.
e.g. Fpnr snl q – Fnr b Fs b Fnl “ N b S b N
And that’s it!
Except... this might not be very enlightening yet. It’ll surely be

helpful to look at a toy example. So in the next couple of pages, let’s
use the DisCoCat model to compute the meaning of the sentence
bananas are fruit
By “compute the meaning,” I mean the following: we want to be

able to view the sentence bananas are fruit as a vector, then feed that
vector into the functor F and get an output vector that encodes for
the meaning of the sentence.
Fpbananas are f ruitq “ ??
That output vector will be the “meaning” of the sentence. Our goal is
to find that meaning.
Goal: Compute the meaning of bananas are fruit.
Let’s proceed systematically. I’ll list the computations step-by-

step, starting from the beginning.
Step 1: Assign each word a grammar type in Pregtn, su.
That’s easy enough:
bananas ù n
are ù nr snl
fruit ù n
Step 2: Fix a noun space Fn “ N and a sentence space Fs “ S.
Let’s suppose N is the three-dimensional space spanned by the basis

vectors
sweet, green, furry
which we can represent as column vectors

» fi » fi » fi
1 0 0
w1 “ sweet “ –0fl , w2 “ green “ –1fl , w3 “ furry “ –0fl
0 0 1
as before. These basis vectors generate the noun space. But what
about the sentence space S? For simplicity, let’s define S to be a “true
or false” space so that it’s a one-dimensional vector space spanned
by a single vector ~1. The origin 0 P S corresponds to “false” while ~1
corresponds to “true.” What about scalar multiplies of ~1? If you like,
you’re more than welcome to think of a positive scalar multiple of ~1
as the meaning vector for sentence that is super true. The larger the
scalar, the more true the sentence!
Finally, note that once we’ve established N and S, the verb space
comes for free:
Fpnr snl q “ N b S b N
This is a nine-dimensional space spanned by vectors of the form
wi b~1 b wj where i and j range between 1 and 3.
Step 3: Determine the vector representations of each word in

the sentence.
We’ll simply recycle the vectors we used earlier:

» fi » fi
21 43
bananas “ – 9 fl , fruit “ –19fl
— ffi — ffi
0 0
Note that both of these vectors live in the noun space N since each
word has grammar type n. But what about the transitive verb are? By
Step 1, we know that are has grammar type nr snl and is therefore a
vector in the tensor product N b S b N. That is, there are coefficients
cij P R so that
are “ c11 sweet b~1 b sweet ` c12 sweet b~1 b green ` ¨ ¨ ¨ ` c33 furry b~1 b furry.
Eek. That looks uncomely.
Fortunately, we learned in Section 2.1 that FVect is a compact closed

category and therefore it exhibits process-state duality, which is the
sophisticated way of saying
every vector in a tensor product can be identified with a linear map,
which is the long way of saying
every vector is really a matrix!

And that is excellent news, for if we know “what is what,” i.e. if we

know that puppies are f urry but not green and so on, then we can
re-express the vector for are as a 3 ˆ 3 matrix. The ijth entry of this
matrix is the coefficient cij which is
$
&1, if w is w ,
i j
cij “
%0, otherwise
It’s no surprise that we get the identity
The upshot is that the transitive verb are has matrix representation matrix. Being is all about identity.
That is, the verb are tells you when
» fi something IS something else.
1 0 0
are “ –0 1 0fl
— ffi
0 0 1
Step 4: Choose a type reduction in Pregtn, su
In this step, which takes place in the grammar category, we simply

perform the type reduction already done on page 39. To recap, we
know the grammar types of bananas and are and f ruit, and so we
concatenate those types to obtain nnr snl n. Using the left and right
counit maps, this string of letters reduces down to s, which confirms
that the phrase bananas are f ruit is a tried-and-true sentence. In Step
4, we simply take that reduction morphism
e r ¨1 s ¨ e l
nnr snl n nÝÑ n n
and hold on to it. We’ll need to use it in Step 5.
Aisde: You might wonder about the word “Choose” in “Step 4: Choose
a type reduction.” What’s up with that? Incidentally, no choice was
needed in this toy example of ours, so the purpose of this aside might
be unclear. Indeed, there’s only one way to parse the sentence bananas
are fruit. But there exist sentences that can be parsed in more than one
way. Consequently, the grammar type of such sentences may reduce
down to type s via more than one reduction morphism. In Step 4, we are
required to choose one. As an illustration, here is a nice sentence:
I saw a man with a telescope.
How did you parse it? Perhaps
I saw (a man with a telescope).
or perhaps
I saw (a man) with a telescope.

Those are two parsings of the same sentence, each of which corre-
sponds to a different type reduction in the pregroup. In turn, this gives
rise to different meaning vectors! And rightly so. Those two sentences
have different meanings! Step 4 is simply reminding us of this fact.
Step 5: Apply F!
This is the fun part! We have a morphism in the pregroup
e r ¨1 s ¨ e l
nnr snl n nÝÑ n n
and we can apply F to get a linear map of vector spaces

e N b1 S b e N
NbNbSbNbN ÝÑ S
where e : N b N Ñ R is the linear map given on page 17 and 1S : S Ñ

S denotes the identity map on S. Finally, apply this linear map to the
vector corresponding to the sentence
e N b 1S b e N pbananas b are b fruitq
which amounts to a simple matrix multiplication

» fi » fi
” ı 1 0 0 43
e N b 1S b e N pbananas b are b fruitq “ 21 9 0 –0 1 0fl –19fl “ 1074
— ffi — ffi
0 0 1 0
Conclusion? The meaning of the sentence bananas are f ruit is
1074~1
which is super true. Voila!

Some Closing Remarks

This functor described above is somewhat reminiscent of a topological
quantum field theory, which is a functor from the category of cobor-
disms (another compact closed category) to the category of complex
Hilbert spaces. But in 2014 Anne Preller showed that the only func-
tors from a pregroup P freely generated on a finite set of basic types
to FVect are those mapping to one-dimensional spaces. The key to her
proof is the fact that P is a poset and hence there is at most one mor-
phism between any two objects. In particular, any morphism from
an object to itself must be the identity. As a consequence, if a is in P
then the morphism p1a ¨ ηa q ˝ pea ¨ 1a q from aar a Ñ aar a must equal the
identity on aar a. Graphically:
a ar a a ar a
a ar a a ar a
Now consider a functor P Ñ FVect. It assigns a in P to a vector space
A in FVect, and it assigns p1a ¨ ηa q ˝ pea ¨ 1a q to the corresponding
linear map, p1 A ¨ η A q ˝ pe A ¨ 1 A q : A b A˚ b A Ñ A b A˚ b A, which
we’ll just denote by f ,
A A˚ A
f “
A A˚ A
and which must be an isomorphism. Now if the dimension of A is at
least 2, then we can choose orthogonal basis vectors e1 and e2 so that
e1 b e2˚ b e1 P A b A˚ b A. And since e A computes the inner product
between e1 and e2˚ , we have f pe1 b e2˚ b e1 q “ 0. Therefore f is not
injective, and so it cannot be an isomorphism.
The intuition is, perhaps, that pregroups have too few morphisms
to capture the semantics. In particular, pregroups do not allow us
to distinguish different parsings of strings of types. One string may
reduce in several ways—e.g. (Men and women) whom I like vs. Men
and (women whom I like)—and the morphisms in a pregroup do
not account for this. So in some sense, there isn’t enough “wiggle
room” for meaning in pregroup syntax, so the output can only be a
one-dimensional vector space. But all is not lost! As Preller showed,
the problem can be fixed by replacing a free pregroup with a free
compact closed category. For more details, see her paper “From
Logical to Distributional Models.”
3.3 Further Reading

For more on the work of Baez and Pollard:
• Certainly take a look at their paper, “A Compositional Framework

for Reaction Networks”. The n-Category Café also contains exposi-
tions, including a post by Baez with the same title and “Dynamical
Systems and Their Steady States” by Maru Sarazola.
• The examples in Section 3.1 can be found in “The Mathematics of

Networks”, a wonderfully accessible talk by Baez on YouTube.
• All of the decorated cospan formalism can be found in Brendan

Fong’s thesis, “The Algebra of Open and Interconnected Systems.”
As part of the ACT workshop, Jonathan Lorand and Fabrizio Gen-
ovese wrote about Fong’s thesis in an article titled “Hypergraph
Categories of Cospans” on the n-Category Café.
• The framework of chemical reaction networks is also being used

to model ATP coupling! This project was birthed during the ACT
workshop. For more, take a look at “Coupling Through Emergent
Conservation Laws ” on Baez’s Azimuth blog.
For more on the work of Coecke, Sadrzadeh, and Clark:
• There is a delightful 5-minute YouTube video (made in the style

of Minute Physics!) called “How Quantum Theory Can Help
Understanding Natural Language” that does an excellent job of
explaining the ideas behind DisCoCat in a non-technical way.
• The yellow banana example of the previous section was just a toy
example meant to showcase the functor of the DisCoCat model of
meaning. But this is a document on applied category theory, and
so you’d surely like to see some applications! For empirical data
arising from actual implementations of the DisCoCat model, take a
look at:
– “Experimental support for a categorical compositional dis-

tributional model of meaning” by Edward Grefenstette and
Mehrnoosh Sadrzadeh
– “Prior disambiguation of word tensors for constructing sentence
vectors” by Dimitri Kartsaklis and Mehrnoosh Sadrzadeh
– “Quantization, Frobenius and bi-algebras from the categorical
framework of quantum mechanics to natural language seman-
tics” by Mehrnoosh Sadrzadeh
• The DisCoCat model was also featured on the n-Category Café

as part of the Applied Category Theory Workshop. In fact, it was
featured twice! The two blog posts are:
– “Linguistics Using Category Theory” by Corey Griffith and Jade

Master. This wonderfully written article gives another recap of
the DisCoCat model, as well as another toy example. Check out
the comment section, too. There are some nice discussion going
on there!
– “Cognition, Convexity, and Category Theory” by Brad Theilman
and me. This blog post is a summary of “Interacting Conceptual
Spaces I,” a paper in which the authors change the semantics
category of the DisCoCat model from FVect to something that
attempts to model human cognition more closely, namely con-
vex spaces! This is a great example of tweaking the semantics
part of functorial semantics.
• The last suggested resource is unrelated to DisCoCat, but it’s still

in the vein of machine learning + applied category theory and
so I thought I’d share: Did you know that backpropagation is a
functor?
4 But Wait! There’s More...

There’s much more to applied category theory—I’ve only presented
a very tiny subset of hand-selected ideas. But there’s so much more
to see, learn, and do! So to close out these notes, I’ll leave you with
a few more links where you can discover other themes, constructions,
and examples of applied category theory.
• To start, there’s the main Applied Category Theory webpage,

which has
– a description of the 2018 workshop that took place at the

Lorentz Center in Leiden, Netherlands
– a call for the 2019 workshop, to be hosted by Bob Coecke in
Oxford.
• Jelle Harold and the folks at Statebox filmed most of the 2018
workshop talks, and you can watch them here: https://statebox.org/events/act-
leiden.html. Speakers include Samson Abramsky, John Baez, Bob
Coecke, Kathryn Hess, Aleks Kissinger, Tom Leinster, David Spi-
vak, and many more!
• Back in March 2018, there was an applied category theory work-

shop hosted at the National Institute of Standards and Technology.
Slides and videos of the talk can be found here:
http://www.appliedcategorytheory.org/nist-workshop-slides/
• There is also Seven Sketches in Compositionality (subtitle: “An Invi-

tation to Applied Category Theory”) by Brendan Fong and David
Spivak. I’ve referenced this book several times already, but that’s
because it’s such a gem! (I was sold just after reading the preface.)
It’s a delightful and insightful introduction to more themes, more
constructions, and more examples within applied category theory.
Even better, no prior knowledge of category theory is assumed.
The book is based on a MIT course the authors taught. You can
find videos of their lectures here: http://math.mit.edu/ dspi-
vak/teaching/sp18/
• And as if all of these great resources weren’t enough, John Baez is

running a free online course on applied category theory. Participants
have been working through the Seven Sketches book. The lectures
and ensuing discussions are a treasure trove of exciting mathe-
matics: https://forum.azimuthproject.org/categories/applied-
category-theory-course. Also be sure to take a look at the “applied
category” tag on his blog, Azimuth.

Applied Cat-Theory

Uploaded by

Copyright:

Available Formats

Applied Cat-Theory

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Applied Cat-Theory

Uploaded by

Copyright:

Available Formats

TA I - D A N A E B R A D L E Y

arXiv:1809.05923v1 [math.CT] 16 Sep 2018

What is applied category theory?

#1 Applied category theory? Isn’t that an oxymoron?

A quick note to the reader

• I’ll assume the reader is comfortable with the basics of cate-

While diagrams of networks have been independently introduced in

I like to imagine that category theory is like a cup of black coffee,

I hope you’ll enjoy it as much as I do!

1.1 Functorial Semantics

refers to the meaning of those things. A functor

provides a way to bring the syntax to life.

grammar ù syntax meaning ù semantics. become quite literal in applied category

It is grammatically correct, yet it has no meaning. The point here is

it might not seem to have anything to do with triangles. Fortunately,

Here R is a rotation by 60˝ while S is reflection across the x-axis.

G ù syntax AutpVq ù semantics

So a group representation is like a (structure-preserving) morphism

Example: a monoid is the image of a functor from a certain syntax cate-

In footnote 3 you’ll notice I dropped

1 : t˚u Ñ F1 i.e. 1 : t˚u Ñ M f pxq “ f ˝ x “ f x

where on the right hand side, I’ve

is a monoid in the category C. Sometimes this monoid goes by a

1. Topological Monoid. Let pTop, ˆ, ˚q denote the category of

is a topological monoid. That is, a topological monoid is a monoid

You’ll notice that in each of these examples, a change in the seman-

That’s the idea behind functorial semantics. Now, how is it used

Petri nets Ñ dynamical systems

as shown in “A Compositional Framework for Reaction Networks”

grammar Ñ meanings of words

via the work of Bob Coecke, Mehrnoosh Sadrzadeh, and Stephen

2. the rules for how those parts are combined.

As it turns out, the name of that journal is itself Compositionality,4 4

1.3 Further Reading

• You might also enjoy this discussion on doctrines over at the n-

• The preface to Brendan Fong’s PhD thesis, “The Algebra of Open

• Applied category theorist Jules Hedges has also written a nice

2.1 Monoidal Categories

pZ, `, 0q pQ, ¨, 1q pMn pRq, ¨, 1q

Each example consists of a set X equipped with an associative

Not too bad, right?

Great. Now imagine replacing the set X by a category C, and re-

There’s a little more to the story here

pSet, ˆ, t˚uq pTop, \, ∅q pFVect, b, kq

By the way, if there is an isomorphism A b B – B b A for all ob- 5

• The category’s composition ˝ corresponds to using one box’s out-

• The monoidal product b corresponds to stacking boxes on top of

In other words, ˝ means “do the processes in series” while the

Earlier, I mentioned that the word “symmetric” can be used as

Similarly, there is another flavor of monoidal categories that we

Interlude: Compact Closed Categories and String Diagrams

FVect is a compact closed category.

What makes that sentence true? Answer: every finite dimensional

A compact closed category just the name for a monoidal category in

The map e is called the counit8 , and it assigns to every vector in 8

Intuitively, we can think of e as an evaluation map. That’s because

And if we view f as a 1 ˆ n matrix and v as an n ˆ 1 matrix, then f v

Definition 2.1. A compact closed category is a monoidal category