-
Notifications
You must be signed in to change notification settings - Fork 23
/
metamath.tex
272 lines (232 loc) · 14.4 KB
/
metamath.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
\chapter{Metamathematical remarks}
\label{app:metamath}
Metamathematics is the study of mathematical theories as
mathematical objects in themselves.
This book is primarily a mathematical theory of symmetries.
Occasionally, however, we have made statements like
``the law of the excluded middle is not provable in our theory''.
This is a statement \emph{about}, and not \emph{in}, the type theory of this book.
As such it is a metamathematical statement.
Sometimes it is possible to encode statements
about a theory in the language of the theory itself.%
\marginnote{We leave aside that this sometimes can be done in different ways.
Historically, the first way was by ``Gödel-numbering'':
encoding all bits of syntax, including statements, as natural numbers,
so that the constructions and deductions of the theory correspond to definable operations on the encoding numbers.
In type theory, there are usually much more perspicacious ways of encoding mathematical theories
using types and type families.}
Even if true, the encoded metamathematical statement can
be unprovable in the theory itself.
The most famous example is G\"{o}del's second incompleteness theorem.\footnote{%
The original reference is~\citeauthor{Goedel2nd}\footnotemark{},
translated into English in~\citeauthor{Heijenoort-source}\footnotemark{}.
For an accessible introduction, see for instance~\citeauthor{Franzen-Goedel}\footnotemark{}
or~\citeauthor{Smullyan-Goedel}\footnotemark{}.}%
\addtocounter{footnote}{-4}%
\stepcounter{footnote}\footcitetext{Goedel2nd}%
\stepcounter{footnote}\footcitetext{Heijenoort-source}%
\stepcounter{footnote}\footcitetext{Franzen-Goedel}%
\stepcounter{footnote}\footcitetext{Smullyan-Goedel}.
G\"{o}del encoded, for any theory $T$ extending Peano Arithmetic and satisfying
some general assumptions, the statement that $T$ is consistent as
a statement $\Con(T)$ in Peano Arithmetic.
Then he showed that $\Con(T)$ is not provable in $T$.
We say that a metamathematical statement about a theory $T$
is \emph{internally} provable if its encoding is provable in $T$.
For example, the metamathematical statement ``if $P$ is unprovable in $T$,
then $T$ is consistent'' is internally provable in $T$, for any $T$ that
satisfies the assumptions of G\"{o}del's second incompleteness theorem.
The type theory in this book satisfies the assumptions of
G\"{o}del's second incompleteness theorem, which include, of course,
the assumption that $T$ is consistent. Thus there is no hope
that we can prove the consistency of our type theory internally.
Moreover, by the previous paragraph, we must be prepared that
no unprovability statement can be proved internally.
[TODO For consistency of UA, LEM, etc, refer to simplicial set model \footcite{KapLum}.
For unprovability of LEM, refer to cubical set model \footcite{BezCoqHub}.]
One property of type theory that we will use is \emph{canonicity}.
We call an expression \emph{closed} if it does not contain free variables.
One example of canonicity is that every closed expression of type $\NN$
is a \emph{numeral}, that is, either $0$ or $S(n)$ for some numeral $n$.
Another example of canonicity is that every closed expression of
type $L\coprod R$ is either of the form $\inl{l}$ for some $l:L$ or
of the form $\inr{r}$ for some $r:R$.
Both examples of canonicity above are clearly related to the
inductive definitions of the types involved: they are
expressed in terms of the constructors of the respective types.
One may ask what canonicity then means for the empty type $\false$,
defined in \cref{sec:finite-types} as the inductive type
with no constructors at all. The answer is that canonicity for $\false$
means that there cannot be a closed expression of type $\false$.
But this actually means that our type theory is consistent!
Therefore we cannot prove general canonicity internally.
[TODO no canonical forms: $x:\NN$, $\trp[P]{\ua(\id)}(0): \NN$,
with $P\defeq (p:\true \mapsto \NN)$
and (problematic) $\trp[Q]{\Sloop}(0): \NN$
with $Q\defeq (z:\Sc \mapsto \NN)$.]
[TODO A second important property of our theory is
that one can compute canonical forms.]
\section{Equality by definition}
\label{sec:defeq}
\subsection{Basics}
\label{sec:defeq-basics}
The concept of definition was introduced in \cref{univalent-mathematics},
together with what it means to be \emph{the same by definition}.
Being the same by definition (NB appears for the first time on p. 26!)
is a relationship between syntactic expressions.
In this section we provide more details about this relationship.
There are four basic forms of equality by definition:
\begin{enumerate}
\item\label{it:exp-defeq} Resulting from making an explicit definition,
e.g., $1 \defeq \Succ(0)$, after which we have $1 \jdeq \Succ(0)$;%
\footnote{The notation $\defeq$ tells the reader that we make a definition
(or reminds the reader that this definition has been made).}
\item\label{it:imp-defeq} Resulting from making an implicit definition,
like we do in inductive definitions,
e.g., $n+0 \defeq n$ and $n+\Succ(m) \defeq \Succ(n+m)$,
after which we have $n+0 \jdeq n$ and $n+\Succ(m) \jdeq \Succ(n+m)$;
\item\label{it:beta} Simplifying the application of an explicitly defined
function to an argument, e.g., $(x \mapsto e_x)(a) \jdeq e_a$;
\item\label{it:eta} Simplifying $(x \mapsto e_x)$ to $f$ when $e_x$
is the application of the function $f$ to the variable $x$,
e.g., $(x \mapsto S(x)) \jdeq S$.
\end{enumerate}
Equality by definition is the \emph{congruence closure} of these
four basic forms, that is, the smallest reflexive, symmetric, transitive
and congruent relation that contains all instances of the four basic forms.
Here a congruent relation is a relation that is closed under all syntactic
operations of type theory. One such operation is substitution, so that we
get from the examples above that, e.g., $1+0 \jdeq 1$
and $n+\Succ(\Succ(m)) \jdeq \Succ(n+\Succ(m))$. Another important
operation is application. For example, we can apply $\Succ$ to each of
the sides of $n+\Succ(m) \jdeq \Succ(n+m)$ and get
$\Succ(n+\Succ(m)) \jdeq \Succ(\Succ(n+m))$, and also
$n+\Succ(\Succ(m)) \jdeq \Succ(\Succ(n+m))$ by transitivity.
Let's elaborate $\id \circ f \jdeq f$ claimed on page~\pageref{page:idofetaf}.
The definitions used on the left hand side are
$\id \defeq (y \mapsto y)$ and $g \circ f \defeq (x \mapsto g(f(x)))$.
In the latter definition we substitute $\id$ for $g$ and get
$\id \circ f \jdeq (x \mapsto \id(f(x)))$. Unfolding $\id$ we get
$(x \mapsto \id(f(x))) \jdeq (x \mapsto (y\mapsto y)(f(x)))$.
Applying \ref{it:beta} we can substitute $f(x)$ for $(y\mapsto y)(f(x)))$
and get $(x \mapsto (y\mapsto y)(f(x))) \jdeq (x \mapsto f(x))$.
By \ref{it:eta} the right hand side is equal to $f$ by definition.
Indeed $\id \circ f \jdeq f$ by transitivity.
Equality by definition is also relevant for typing.
For example, let $A: \UU$ and $P: A\to\UU$. If $B\jdeq A$,
then $ (B\to\UU)\jdeq(A\to\UU)$ by congruence, and also $P: B\to\UU$,
and even $\prod_{x:B} P(x) \jdeq \prod_{x:A} P(x)$.
%Discuss later: If syntactic expressions $e$ and $e'$ are equal by definition, denoted by $e\jdeq e'$, then $e$ and $e'$ have the same type $T$, then $\refl e : e=_T e'$; the converse is not true.
\subsection{Deciding equality by definition (not updated yet)}
\label{sec:defeq-computation}
By a \emph{decision procedure} we mean a terminating algorithmic procedure that
answers a yes/no question.
Although it is possible to enumerate all true equalities by definition,
this does not give a test that answers whether or not a given instance $e\jdeq e'$ holds.
In particular when $e\jdeq e'$ does not hold, such an enumeration will not terminate.
A test of equality by definition is important for type checking,
as the examples in the last paragraph of the previous section show.
A better approach to a test of equality by definition is the following.
First direct the four basic forms of equality by definition from left to right
as they are given.\footnote{%
TODO: think about the last, $\eta$.}
For the first two forms this can be viewed as unfolding definitions,
and for the last two forms as simplifying function application and (unnecessary)
abstraction, respectively.
This defines a basic reduction relation, and we write $e\to e'$ if $e'$ can
be obtained by a basic reduction of a subexpression in $e$.
The reflexive transitive closure of $\to$ is denoted by $\to^*$.
The symmetric closure of $\to^*$ coincides with $\jdeq$.
We mention a few important properties of the relations $\to,\to^*$ and $\jdeq$.
The first is called the Church--Rosser property, and states that,
if $e\jdeq e'$, then there is an expression $c$ such that $e\to^* c$
and $e'\to^* c$. The second is called type safety and states that,
if $e:T$ and $e\to e'$, then also $e':T$.
The third is called termination and states that for well-typed expressions $e$
there is no infinite reduction sequence starting with $e$.
The proofs of Church--Rosser and type safety are long and tedious, but pose no essential
difficulties. For a non-trivial type theory such as in this book the last property,
termination, is extremely difficult and has not been carried out in full detail.
The closest come results on the Coq \footcite{Coq} (TODO: find good reference).
Testing whether or not two given well-typed terms $e$ and $e'$
are equal by definition can now be done
by reducing them with $\to$ until one reaches irreducible expressions $n$ and $n'$
such that $e\to^* n$ and $e'\to^* n'$, and then comparing $n$ and $n'$.
Now we have: $e\jdeq e'$ iff $n\jdeq n'$ iff (by Church--Rosser)
there exists a $c$ such that $n\to^* c$ and $n'\to^* c$.
Since $n$ and $n'$ are irreducible the latter is equivalent to
$n$ and $n'$ being identical syntactic expressions.
\section{The Limited Principle of Omniscience}
\label{sec:LPO}
\begin{remark}\label{rem:LPO-solves-halting problem}
Recall the Limited Principle of Omniscience (LPO), \cref{LPO}:
for any function $P:\NN\to\bn 2$,
either there is a smallest number $n_0:\NN$ such that $P(n_0)=1$,
or $P$ is a constant function with value $0$.
We will show that LPO is not provable in our theory.
The argument is based on the halting problem: given a Turing machine
$M$ and an input $n$, determine whether $M$ halts on $n$.
It is known that the halting problem cannot be solved by an algorithm
that can be implemented on a Turing machine.\footnote{It's commonly accepted that
every algorithm \emph{can} be thus implemented.}
We use a few more facts from computability theory.
First, Turing machines can be enumerated. We denote the $n$\th Turing machine $M_n$,
so we can list the Turing machines in order: $M_0,M_1,\ldots$.
Secondly, there exists a function $T(e,n,k)$ such that $T(e,n,k) = 1$
if $M_e$ halts on input $n$ in at most $k$ steps, and $T(e,n,k) = 0$
otherwise. This function $T$ can be implemented in our theory.
Towards a contradiction, assume we have a closed proof $t$ of LPO in our theory.
We assume as well that $t$ does not depend on any axiom.\footnote{It is possible to weaken the notion
of canonicity so that the argument still works even if the proof $t$ uses the Univalence Axiom.
Of course, the argument must fail if we allow $t$ to use LEM!}
It is clear that $k\mapsto T(e,n,k)$ is a constant function with value $0$
if and only if $M_e$ does not halt on input $n$. Now consider $t(k\mapsto T(e,n,k))$,
which is an element of a type of the form $L\coprod R$.
We now explain how to solve the halting problem.
Let $e$ and $n$ be arbitrary numerals.
Then $t(k\mapsto T(e,n,k))$ is a closed element of $L\coprod R$.
Hence we can compute its canonical form. If $t(k\mapsto T(e,n,k))\jdeq\inr{r}$ for some
$r:R$, then $k\mapsto T(e,n,k)$ is a constant function with value $0$,
and $M_e$ does not halt on input $n$. If $t(k\mapsto T(e,n,k))\jdeq\inl{l}$ for some
$l:L$, then $M_e$ does halt on input $n$.
Thus we have an algorithm to solve the halting problem
for all $e$ and $n$. Since this is impossible, we have refuted the assumption
that there is a closed proof $t$ of LPO in our theory.
\end{remark}
\section{Topology}
\label{sec:topology}
In this section we will explain how our intuition about types relates to our intuition about topological spaces.
INSERT AN INTRODUCTORY PARAGRAPH HERE.
[Intuitively, the types of type theory can be modeled by topological spaces,
and elements as points thereof. However, this is not so easy to achieve, and the first
model of homotopy theory theory was in simplicial sets. Topological spaces and simplicial sets are both \emph{models} of homotopy types. And by a lucky coincidence, it makes sense
to say that homotopy type theory is a theory of homotopy types.]
Some references include: \textcite{Hatcher-AlgTop,May-Concise,MayPonto-MoreConcise}
\begin{remark}
\label{rem:injectionsurjectionisnotwhatyouthink}
Our definitions of injections and surjections are lifted directly from the intuition about sets. However, types need not be sets, and
thinking of types as spaces may at this point lead to a slight confusion.
The real line is contractible and the inclusion of the discrete subspace $\{0,1\}$ is, well, an inclusion (of sets, which is the same thing as
an inclusion of spaces). However, $\{0,1\}$ is not connected, seemingly contradicting the next result.
This apparent contradiction is resolved once one recalls the myopic nature of our setup: the contractibility of the real line means that ``all
real numbers are identical'', and \emph{our} ``preimage of $3{.}25$'' is not a proposition: it contains \emph{both} $0$ and $1$. Hence
``$\{0,1\}\subseteq\mathbb R$'' would not count as an injection in our sense.
We should actually have been more precise above: we were referring to the \emph{homotopy type} of the real line, rather than the real line itself.\footnote{\label{ft:cohesive}%
We don't define this formally here,
see \citeauthor{Shulman-Real-Cohesive}\footnotemark{} for a synthetic account.
The idea is that the homotopy type $\constant{h}(X)$ of a type $X$
has a map from $X$, $\iota : X \to \constant{h}(X)$,
and any continuous function $f : [0,1] \to X$
gives rise to a path
$\iota(f(0)) = \iota(f(1))$ in
$\constant{h}(X)$.}\footcitetext{Shulman-Real-Cohesive}
We shall later (in the chapters on geometry) make plenty of use of the latter,
which is as usual a set with uncountably many elements.
\end{remark}
\input{choicefin}
%%% Local Variables:
%%% mode: latex
%%% fill-column: 144
%%% latex-block-names: ("lemma" "theorem" "remark" "definition" "corollary" "fact" "properties" "conjecture" "proof" "question" "proposition")
%%% TeX-master: "book"
%%% End: