viewfinder.tex 13.7 KB
Newer Older
Michael Kohlhase's avatar
Michael Kohlhase committed
1
2
3
4
5
6
\begin{definition}\rm
Let $C$ be a corpus of theories with the same fixed meta-theory $M$.
We call the problem of finding theory morphisms between theories of $C$ the \defemph{view finding problem} and an algorithm that solves it a \defemph{view finder}. 
\end{definition}
Note that a view finder is sufficient to solve the theory classification use case from the introduction: 
Jane provides a $M$-theory $Q$ of beautiful sets, the view finder computes all (total) views from $Q$ into $C$, and returns presentations of target theories and the assignments made by the views.
Florian Rabe's avatar
Florian Rabe committed
7

Michael Kohlhase's avatar
Michael Kohlhase committed
8
The cost of this problem quickly explodes.
Florian Rabe's avatar
Florian Rabe committed
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
First of all, it is advisable to restrict attention to simple morphisms.
Eventually we want to search for arbitrary morphisms as well.
But that problem is massively harder because it subsumes theorem proving: a morphism from $\Sigma$ to $\Sigma'$ maps $\Sigma$-axioms to $\Sigma'$-proofs, i.e., searching for a theory morphism requires searching for proofs.

Secondly, if $C$ has $n$ theories, we have $n^2$ pairs of theories between which to search.
(It is exactly $n^2$ because the direction matters, and even morphisms from a theory to itself are interesting.)
Moreover, for two theories with $m$ and $n$ constants, there are $n^m$ possible simple morphisms.
(It is exactly $n^m$ because morphisms may map different constants to the same one.)
Thus, we can in principle enumerate and check all possible simple morphisms in $C$.
But for large $C$, it quickly becomes important to do so in an efficient way that eliminates ill-typed or uninteresting morphisms early on.

Thirdly, it is desirable to search for \emph{partial} theory morphisms as well.
In fact, identifying refactoring potential is only possible if we find partial morphisms: then we can refactor the involved theories in a way that yields a total morphism.
Moreover, many proof assistant libraries do not follow the little theories paradigm or do not employ any theory-like structuring mechanism at all.
These can only be represented as a single huge theory, in which case we have to search for partial morphisms from this theory to itself.
While partial morphism can be reduced to and then checked like total ones, searching for partial morphisms makes the number of possible morphisms that must be checked much larger.

Finally, even for a simple theory morphism, checking reduces to a set of equality constraints, namely the constraints $\vdash_{\Sigma'}\ov{\sigma}(E)=E'$ for the type-preservation condition.
Depending on $M$, this equality judgment may be undecidable and require theorem proving.

A central motivation for our algorithm is that equality in $M$ can be soundly approximated very efficiently by using a normalization function on $M$-expressions.
This has the additional benefit that relatively little meta-theory-specific knowledge is needed, and all such knowledge is encapsulated in a single well-understood function.
Thus, we can implement theory morphism--search generically for arbitrary $M$.

Our algorithm consists of two steps.
First, we preprocess all constant declarations in $C$ with the goal of moving as much intelligence as possible into a step whose cost is linear in the size of $C$.
Then, we perform the morphism search on the optimized data structures produced by the first step.

\subsection{Preprocessing}

The preprocessing phase computes for every constant declaration $c:E$ a normal form $E'$ and then the long abstract syntax tree for $E'$.
Both steps are described below.

\paragraph{Normalization}
Normalization involves two steps:
\begin{compactenum}
 \item MMT-level normalization performs generic transformations that do not depend on the meta-theory $M$.
Dennis Müller's avatar
Dennis Müller committed
46
47
   These include elaboration of structured theories and definition expansion, which we mentioned in Sect.~\ref{sec:prelim}.
 \item Meta-theory-level normalization applies an $M$-specific normalization function, which we assume as a black box for now and discuss further in Sect.~\ref{sec:preproc}.
Florian Rabe's avatar
Florian Rabe committed
48
49
50
51
\end{compactenum}

\paragraph{Abstract Syntax Trees}
We define \textbf{abstract syntax trees} as pairs $(t,s)$ where $t$ is subject to the grammar
Dennis Müller's avatar
Dennis Müller committed
52
 \[t ::= C_{Nat} \mid V_{Nat} \mid \ombind{t}{t^+}{t^+}\]
Florian Rabe's avatar
Florian Rabe committed
53
(where $Nat$ is a non-terminal for natural numbers) and $s$ is a list of constant names.
Dennis Müller's avatar
Dennis Müller committed
54
We obtain an abstract syntax tree from an MMT expression $E$ by (i) switching to de-Bruijn representation of bound variables and (ii) replacing all occurrences of constants with $C_i$ in such a way that every $C_i$ refers to the $i$-th element of $s$.
Florian Rabe's avatar
Florian Rabe committed
55
56
57
58
59
60

Abstract syntax trees have the nice property that they commute with the application of simple morphisms $\sigma$:
If $(t,s)$ represents $E$, then $\sigma(E)$ is represented by $(t,s')$ where $s'$ arises from $s$ by replacing every constant with its $\sigma$-assignment.

The above does not completely specify $i$ and $s$ yet, and there are several possible canonical choices among the abstract syntax trees representing the same expression.
The trade-off is subtle because we want to make it easy to both identify and check theory morphisms later on.
Dennis Müller's avatar
Dennis Müller committed
61
We call $(t,s)$ the \textbf{long} abstract syntax tree for $E$ if $C_i$ replaces the $i$-th occurrence of a constant in $E$ when $E$ is read in left-to-right order.
Florian Rabe's avatar
Florian Rabe committed
62
In particular, the long tree does not merge duplicate occurrences of the same constant into the same number.
Dennis Müller's avatar
Dennis Müller committed
63
The \textbf{short} abstract syntax tree for $E$ arises from the long one by removing all duplicates from $s$ and replacing the $C_i$ accordingly.
Florian Rabe's avatar
Florian Rabe committed
64

Dennis Müller's avatar
Dennis Müller committed
65
66
\begin{newpart}{DM}
\paragraph{} As an example, consider again the term $\lambda x,y:\mathbb N.\; (x + \cn{one})\cdot(y+\cn{one})$ with internal representation $\ombind{\lambda}{x:\mathbb N,y : \mathbb N}{\oma{\cdot}{\oma{+}{x,\cn{one}},\oma{+}{y,\cn{one}}}}$.
Florian Rabe's avatar
Florian Rabe committed
67

Dennis Müller's avatar
Dennis Müller committed
68
69
70
71
72
73
The \emph{short} syntax tree and list of constants associated with this term would be:
\[t = \ombind{C_1}{C_2,C_2}{\oma{C_3}{\oma{C_4}{V_2,C_5},\oma{C_4}{V_1,C_5}}} \]
\[ s = (\lambda,\mathbb N,\cdot,+,\cn{one}) \]
\end{newpart}

\paragraph{} In our algorithm, we pick the \emph{long} abstract syntax tree, which may appear surprising.
Florian Rabe's avatar
Florian Rabe committed
74
75
76
77
The reason is that shortness is not preserved when applying a simple morphism: whenever a morphism maps two different constants to the same constant, the resulting tree is not short anymore.
Longness, on the other hand, is preserved.
The disadvantage that long trees take more time to traverse is outweighed by the advantage that we never have to renormalize the trees.

Dennis Müller's avatar
Dennis Müller committed
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
%\begin{oldpart}{FR: replaced with the above}
%
%\subsection{The General Algorithm} The aim is to find typing-preserving morphisms between theories, i.e. given a constant $C:t$ in a theory $T_1$, we want to find a view $V:T_1\to T_2$ such that if $V(C)=C'$ and $C':t'\in T_2$, we have $V(t)=t'$. Correspondingly, we need to consider the types of the constants in our two theories, which we assume to be flat.
%
%To run a proper unification algorithm is in our situation infeasible, since the flat version of a theory can become prohibitively large (and obviously finding two unifiable types in two theories is a search problem quadratic in the number of constants). To solve that problem, we first preprocess our theories such that pre-selecting plausibly ``unifiable'' constants becomes as fast as possible.
%
%\paragraph{} We do so by first transforming each constant $C$ in a theory to an \emph{encoding} $(\cn{h}(C),\cn{syms}(C))$ the following way:
%
%Consider the syntax tree of the type $t$ of a constant $C$. We first systematically replace the leaves by an abstract representation, yielding a data structure $\cn{tree}(C)$. We can eliminate variables by replacing them by their De Bruijn index, and symbol references by enumerating them and storing the symbol's names in a list $\cn{syms}(C)$. 
%
%As a result, we get a pair $(\cn{tree}(C),\cn{syms}(C))$. Now an assignment $V(C)=D$ is valid, iff $\cn{tree}(C)=\cn{tree}(D)$, the lists $\cn{syms}(C)$ and $\cn{syms}(D)$ have the same length and their pairwise assignments $V(\cn{syms}(C)_i)=\cn{syms}(D)_i$ are all valid.
%
%Furthermore, since we only need $\cn{tree}(C)$ for an equality check, we can immediately replace $\cn{tree}(C)$ by an integer hashcode $\cn{h}(C)$.
%
%Now given a constant $C\in T_1$, we can find valid matches in $T_2$ by computing the encodings for each constant in $T_1$ and $T_2$, taking the list of constants $C'\in T_2$ with $\cn{h}(C')=\cn{h}(C)$ and recursing into their respective lists of symbols.
%\end{oldpart}
Florian Rabe's avatar
Florian Rabe committed
94
95
96
97

\subsection{Search}

Consider two constants $c:E$ and $c':E'$ preprocessed into long abstract syntax trees $(t,s)$ and $(t',s')$.
Dennis Müller's avatar
Dennis Müller committed
98
99
100
101
102
It is now straightforward to show the following Lemma:
\begin{lemma}
The assignment $c\mapsto c'$ is well-typed in a morphism $\sigma$ if $t=t'$ (in which case $s$ and $s'$ must have the same length $l$) and $\sigma$ also contains $s_i\mapsto s'_i$ for $i=1,\ldots,l$.
\end{lemma}
Of course, the condition about $s_i\mapsto s'_i$ may be redundant if $s$ contain duplicates; but because $s$ has to be traversed anyway, it is cheap to skip all duplicates. We call the set of assignments $s_i\mapsto s'_i$ the \textbf{prerequisites} of $c\mapsto c'$.
Florian Rabe's avatar
Florian Rabe committed
103
104

This lemma is the center of our search algorithm:
Dennis Müller's avatar
Dennis Müller committed
105
\paragraph{\textbf{Core Algorithm}} For two given constant declarations $c$ and $c'$, we build a morphism by starting with $\sigma=c\mapsto c'$ and recursively adding all prerequisites to $\sigma$ until
Florian Rabe's avatar
Florian Rabe committed
106
107
\begin{compactitem}
 \item either the process terminates, in which case we have found a morphism,
Dennis Müller's avatar
Dennis Müller committed
108
 \item or $\sigma$ contains two different assignments for the same constant (a contradiction), in which case we fail.
Florian Rabe's avatar
Florian Rabe committed
109
110
\end{compactitem}

Dennis Müller's avatar
Dennis Müller committed
111
112
113
114
115
116
117
118
\begin{newpart}{DM}
	\paragraph{} The general procedure for finding views between two theories $T_1,T_2$ now is as follows:
	\begin{enumerate}
		\item Compute the long abstract syntax trees for all constants in $T_1$ and $T_2$
		\item Find the set $P$ of all pairs $(c,c')$ with $c\in T_1$, $c'\in T_2$ and long abstract syntax trees $(t,s)$ and $(t',s')$ such that $t=t'$.
		\item For each pair in $P$, execute the core algorithm, yielding a set of partial morphisms $M$.
	\end{enumerate}
\end{newpart}
Dennis Müller's avatar
update    
Dennis Müller committed
119

Dennis Müller's avatar
Dennis Müller committed
120
% \ednote{continue description of algorithm: how do we pick $c\mapsto c'$, and so on}
Dennis Müller's avatar
update    
Dennis Müller committed
121

Dennis Müller's avatar
Dennis Müller committed
122
\subsection{Improvements and Parameters of the Algorithm}\label{sec:algparams}
Dennis Müller's avatar
update    
Dennis Müller committed
123

Dennis Müller's avatar
Dennis Müller committed
124
\paragraph{Extending the abstract syntax tree} By the very nature of the approach described in Section \ref{sec:oaf}, many symbols will be common to domain and codomain of a given viewfinding problem: Most importantly, assuming a common meta-theory, we most certainly do not want to reassign the symbols therein.
Dennis Müller's avatar
update    
Dennis Müller committed
125

Dennis Müller's avatar
Dennis Müller committed
126
Therefore, we consider the \emph{common} symbols as a fixed part of the abstract syntax tree of a constant. The symbols will hence be encoded in the component $t$ instead of the list $s$. This will not only exclude spurious matches, but also reduce the number of plausible matches and consequently speed up the algorithm.
Dennis Müller's avatar
update    
Dennis Müller committed
127

Dennis Müller's avatar
Dennis Müller committed
128
\paragraph{Picking starting pairs:} Note, that we will still find many spurious matches if executed in its general form. The reason being that (for example) all atomic types match each other, as will all binary operators etc. Most of these results will not be interesting. Furthermore, since the algorithm needs to recurse into the lists $s$, many potential matches will need to be checked repeatedly. Both problems can be massively reduced by selecting specific pairs of encodings as \emph{starting pairs} for the algorithm, so that the majority of matching constants will only be considered if the algorithm runs into them during recursing. Potential useful starting points are:
Dennis Müller's avatar
update    
Dennis Müller committed
129
\begin{itemize}
Dennis Müller's avatar
Dennis Müller committed
130
131
	\item \emph{Axioms:} Since we are mostly interested in matching constants that share the same mathematical properties, by using axioms as starting point we can ensure that the algorithm only matches constants that have at least one (axiomatic) property in common (e.g. only commutative, or associative operators).
	\item \emph{The length of $s$ in the short abstract syntax tree:} By choosing a minimal length for $s$, we can guarantee that only morphisms will be produced that relate a minimal number of distinct constants. 
Dennis Müller's avatar
update    
Dennis Müller committed
132
\end{itemize}
Dennis Müller's avatar
update    
Dennis Müller committed
133

Dennis Müller's avatar
Dennis Müller committed
134
135
\paragraph{Picking starting theories:} If we try to find views between whole collections of theories, we can obviously disregard all theories that are already included in some other theory in our collections, since we work with a normalized (and dependency-closed) version of a theory. Consequently, by only using maximal theories we do not find any fewer views but speed up the search significantly.
\begin{newpart}{DM}
Dennis Müller's avatar
Dennis Müller committed
136
\paragraph{Morphism aggregation:} The morphisms found by the algorithm are always induced by a single assignment $c\mapsto c'$. Depending on what we intend to do with the results, we might prefer to consider them individually (e.g. to yield \emph{alignments} in the sense of \cite{KKMR:alignments:16}), aggregate them into ideally total views (by merging compatible morphisms, where two morphisms $v_1,v_2$ are compatible if there are no assignments $(c\mapsto c_1)\in v_1$ and $(c\mapsto c_2)\in v_2$ with $c_1\neq c_2$) with varying degrees of modularity.
Dennis Müller's avatar
Dennis Müller committed
137
\end{newpart}
Dennis Müller's avatar
Dennis Müller committed
138
\paragraph{Storing Encodings} Finally, for computing the encodings of a theory we only need to know the symbols to be fixed in the component $t$ of an abstract syntax tree, for which only the meta-theories of the theories are relevant. They also determine the specific preprocessings and translations we want to likely use. Since there is only a small number of meta-theories incolved that are relevant in practice, we can store and retrieve the encodings for the most important situations. Since computing the encodings (as well as sorting the theories in a library by their dependencies) is the most expensive part of the algorithm, this -- once computed and stored -- makes the viewfinding process itself rather efficent.
Dennis Müller's avatar
update    
Dennis Müller committed
139

Michael Kohlhase's avatar
Michael Kohlhase committed
140
141
142
143
144
%%% Local Variables:
%%% mode: latex
%%% eval: (visual-line-mode) (set-fill-coloumn 5000)
%%% TeX-master: "paper"
%%% End: