cat/doc/introduction.tex

\chapter{Introduction}
This thesis is a case-study in the application of cubical Agda in the
context of category theory. At the center of this is the notion of
\nomenindex{equality}. In type-theory there are two pervasive notions
of equality: \nomenindex{judgmental equality} and
\nomenindex{propositional equality}. Judgmental equality is a property
of the type system. Judgmental equality on the other hand is usually
defined \emph{within} the system. When introducing definitions this
report will use the symbol $\defeq$. Judgmental equalities will be
denoted with $=$ and for propositional equalities the notation
$\equiv$ is used.

The rules of judgmental equality are related with $β$- and
$η$-reduction which gives a notion of computation in a given type
theory.
%
There are some properties that one usually want judgmental equality to
satisfy. It must be \nomenindex{sound}, enjoy \nomenindex{canonicity}
and be a \nomenindex{congruence relation}. Soundness means that things
judged to be equal are equal with respects to the \nomenindex{model}
of the theory or the \emph{meta theory}. It must be a congruence
relation because otherwise the relation certainly does not adhere to
our notion of equality. One would be able to conclude things like: $x
\equiv y \rightarrow f\ x \nequiv f\ y$. Canonicity means that any
well typed term evaluates to a \emph{canonical} form. For example for
a closed term $e \tp \bN$ it will be the case that $e$ reduces to $n$
applications of $\mathit{suc}$ to $0$ for some $n$; $e =
\mathit{suc}^n\ 0$. Without canonicity terms in the language can get
``stuck'' meaning that they do not reduce to a canonical form.

To work as a programming languages it is necessary for judgmental
equality to be \nomenindex{decidable}. Being decidable simply means
that that an algorithm exists to decide whether two terms are equal.
For any practical implementation the decidability must also be
effectively computable.

For propositional equality the decidability requirement is relaxed. It
is not in general possible to decide the correctness of logical
propositions (cf.\ Hilbert's \emph{entscheidigungsproblem}).

There are two flavors of type-theory. \emph{Intensional-} and
\emph{extensional-} type theory (ITT and ETT respectively). Identity
types in extensional type theory are required to be
\nomen{propositions}{proposition}. That is, a type with at most one
inhabitant. In extensional type theory the principle of reflection
%
$$a ≡ b → a = b$$
%
is enough to make type checking undecidable. This report focuses on
Agda which at a glance can be thought of as a version of intensional
type theory. Pattern-matching in regular Agda lets one prove
\nomenindex{Uniqueness of Identity Proofs} (UIP). UIP states that any
two identity proofs are propositionally identical.

The usual notion of propositional equality in ITT is quite
restrictive. In the next section a few motivating examples will
highlight this. There exist techniques to circumvent these problems,
as we shall see. This thesis will explore an extension to Agda that
redefines the notion of propositional equality and as such is an
alternative to these other techniques. The extension is called cubical
Agda. Cubical Agda drops UIP as this does not permit
\nomenindex{functional extensionality} and
\nomenindex{univalence}. What makes this extension particularly
interesting is that it gives a \emph{constructive} interpretation of
univalence. What all this means will be elaborated in the following
sections.
%
\section{Motivating examples}
%
In the following two sections I present two examples that illustrate
some limitations inherent in ITT and -- by extension -- Agda.
%
\subsection{Functional extensionality}
\label{sec:functional-extensionality}%
Consider the functions:
%
\begin{align*}%
\var{zeroLeft}  & \defeq \lambda\; (n \tp \bN) \to (0 + n \tp \bN) \\
\var{zeroRight} & \defeq \lambda\; (n \tp \bN) \to (n + 0 \tp \bN)
\end{align*}%
%
The term $n + 0$ is \nomenindex{definitionally} equal to $n$, which we
write as $n + 0 = n$. This is also called \nomenindex{judgmental
  equality}.  We call it definitional equality because the
\emph{equality} arises from the \emph{definition} of $+$ which is:
%
\begin{align*}
  +             & \tp \bN \to \bN \to \bN      \\
  n + 0         & \defeq n                     \\
  n + (\suc{m}) & \defeq \suc{(n + m)}
\end{align*}
%
Note that $0 + n$ is \emph{not} definitionally equal to $n$. This is
because $0 + n$ is in normal form. I.e.\ there is no rule for $+$
whose left hand side matches this expression. We do however have that
they are \nomen{propositionally}{propositional equality} equal, which
we write as $n \equiv n + 0$. Propositional equality means that there
is a proof that exhibits this relation. We can do induction over $n$
to prove this:
%
\begin{align}
\label{eq:zrn}
\begin{split}
\var{zrn}\                & \tp    ∀ n → n ≡ \var{zeroRight}\ n \\
\var{zrn}\ \var{zero}     & \defeq \var{refl} \\
\var{zrn}\ (\var{suc}\ n) & \defeq \var{cong}\ \var{suc}\ (\var{zrn}\ n)
\end{split}
\end{align}
%
This show that zero is a right neutral element hence the name $\var{zrn}$.
Since equality is a transitive relation we have that $\forall n \to
\var{zeroLeft}\ n \equiv \var{zeroRight}\ n$. Unfortunately we don't
have $\var{zeroLeft} \equiv \var{zeroRight}$. There is no way to
construct a proof asserting the obvious equivalence of
$\var{zeroLeft}$ and $\var{zeroRight}$. Actually showing this is
outside the scope of this text. Essentially it would involve giving a
model for our type theory that validates all our axioms but where
$\var{zeroLeft} \equiv \var{zeroRight}$ is not true. We cannot show
that they are equal even though we can prove them equal for all
points. For functions this is exactly the notion of equality that we
are interested in: Functions are considered equal when they are equal
for all inputs. This is called \nomenindex{pointwise equality}, where
the \emph{points} of a function refer to its arguments.
%
\subsection{Equality of isomorphic types}
%
Let $\top$ denote the unit type -- a type with a single constructor.
In the propositions as types interpretation of type theory $\top$ is
the proposition that is always true. The type $A \x \top$ and $A$ has
an element for each $a \tp A$. So in a sense they have the same shape
(Greek;
\nomenindex{isomorphic}). The second element of the pair does not
add any ``interesting information''. It can be useful to identify such
types. In fact, it is quite commonplace in mathematics. Say we look at
a set $\{x \mid \phi\ x \land \psi\ x\}$ and somehow conclude that
$\psi\ x \equiv \top$ for all $x$. A mathematician would immediately
conclude $\{x \mid \phi\ x \land \psi\ x\} \equiv \{x \mid \phi\ x\}$
without thinking twice. Unfortunately such an identification can not
be performed in ITT.

More specifically what we are interested in is a way of identifying
\nomenindex{equivalent} types. I will return to the definition of
equivalence later in section \S\ref{sec:equiv}, but for now it is
sufficient to think of an equivalence as a one-to-one correspondence.
We write $A \simeq B$ to assert that $A$ and $B$ are equivalent types.
The principle of univalence says that:
%
$$\mathit{univalence} \tp (A \simeq B) \simeq (A \equiv B)$$
%
In particular this allows us to construct an equality from an equivalence
%
$$\mathit{ua} \tp (A \simeq B) \to (A \equiv B)$$
%
and vice versa.

\section{Formalizing Category Theory}
%
The above examples serve to illustrate a limitation of ITT. One case
where these limitations are particularly prohibitive is in the study
of Category Theory. At a glance category theory can be described as
``the mathematical study of (abstract) algebras of functions''
(\cite{awodey-2006}). By that token functional extensionality is
particularly useful for formulating Category Theory. In Category
theory it is also commonplace to identify isomorphic structures and
univalence gives us a way to make this notion precise. In fact we can
formulate this requirement within our formulation of categories by
requiring the \emph{categories} themselves to be univalent as we shall
see in \S\ref{sec:univalence}.

\section{Context}
\label{sec:context}
%
The idea of formalizing Category Theory in proof assistants is not new. There
are a multitude of these available online. Notably:
%
\begin{itemize}
\item
  A formalization in Agda using the setoid approach:
  \url{https://github.com/copumpkin/categories}
\item
  A formalization in Agda with univalence and functional
  extensionality as postulates:
  \url{https://github.com/pcapriotti/agda-categories}
\item
  A formalization in Coq in the homotopic setting:
  \url{https://github.com/HoTT/HoTT/tree/master/theories/Categories}
\item
  A formalization in \emph{CubicalTT} -- a language designed for
  cubical type theory. Formalizes many different things, but only a
  few concepts from category theory:
  \url{https://github.com/mortberg/cubicaltt}
\end{itemize}
%
The contribution of this thesis is to explore how working in a cubical setting
will make it possible to prove more things and to reuse proofs and to try and
compare some aspects of this formalization with the existing ones.

There are alternative approaches to working in a cubical setting where
one can still have univalence and functional extensionality. One
option is to postulate these as axioms. This approach, however, has
other shortcomings, e.g. you lose \nomenindex{canonicity}
(\cite[p. 3]{huber-2016}).

Another approach is to use the \emph{setoid interpretation} of type
theory (\cite{hofmann-1995,huber-2016}). With this approach one works
with \nomenindex{extensional sets} $(X, \sim)$. That is a type $X \tp
\MCU$ and an equivalence relation $\sim\ \tp X \to X \to \MCU$ on that
type. Under the setoid interpretation the equivalence relation serve
as a sort of ``local'' propositional equality. Since the developer
gets to pick this relation it is not a\~priori a congruence
relation. So this must be verified manually by the developer.
Furthermore, functions between different setoids must be shown to be
setoid homomorphism, that is; they preserve the relation.

This approach has other drawbacks; it does not satisfy all
propositional equalities of type theory a priori. That is, the
developer must manually show that e.g.\ the relation is a congruence.
Equational proofs $a \sim_{X} b$ are in some sense `local' to the
extensional set $(X , \sim)$. To e.g.\ prove that $x ∼ y → f\ x ∼
f\ y$ for some function $f \tp A → B$ between two extensional sets $A$
and $B$ it must be shown that $f$ is a groupoid homomorphism. This
makes it very cumbersome to work with in practice (\cite[p.
  4]{huber-2016}).

\section{Conventions}
In the remainder of this paper I will use the term \nomenindex{Type}
to describe -- well -- types. Thereby departing from the notation in
Agda where the keyword \texttt{Set} refers to types.  \nomenindex{Set}
on the other hand shall refer to the homotopical notion of a set. I
will also leave all universe levels implicit. This of course does not
mean that a statement such as $\MCU \tp \MCU$ means that we have
type-in-type but rather that the arguments to the universes are
implicit.

And I use the term
\nomenindex{arrow} to refer to morphisms in a category,
whereas the terms
\nomenindex{morphism},
\nomenindex{map} or
\nomenindex{function}
shall be reserved for talking about type theoretic functions; i.e.
functions in Agda.

As already noted $\defeq$ will be used for introducing definitions $=$
will be used to for judgmental equality and $\equiv$ will be used for
propositional equality.

All this is summarized in the following table:
%
\begin{samepage}
\begin{center}
\begin{tabular}{ c c c }
Name & Agda & Notation \\
\hline

\varindex{Type}            & \texttt{Set}         & $\Type$            \\

\varindex{Set}             & \texttt{Σ Set IsSet} & $\Set$             \\
Function, morphism, map & \texttt{A → B}       & $A → B$            \\
Dependent- ditto        & \texttt{(a : A) → B} & $∏_{a \tp A} B$  \\

\varindex{Arrow}           & \texttt{Arrow A B}   & $\Arrow\ A\ B$     \\

\varindex{Object}          & \texttt{C.Object}    & $̱ℂ.Object$     \\
Definition              & \texttt{=}           & $̱\defeq$       \\
Judgmental equality     & \null                & $̱=$            \\
Propositional equality  & \null                & $̱\equiv$
\end{tabular}
\end{center}
\end{samepage}