[helm.git] / helm / papers / matita / matita2.tex

\documentclass[]{kluwer}
\usepackage{color}
\usepackage{graphicx}
\usepackage{hyperref}
\usepackage{color}
\usepackage{fancyvrb}
\usepackage[show]{ed}

\newcommand{\component}{component}
\newcommand{\components}{components}

\newcommand{\AUTO}{\textsc{Auto}}
\newcommand{\BOXML}{BoxML}
\newcommand{\COQ}{Coq}
\newcommand{\COQIDE}{CoqIde}
\newcommand{\ELIM}{\textsc{Elim}}
\newcommand{\GDOME}{Gdome}
\newcommand{\GTKMATHVIEW}{\textsc{GtkMathView}}
\newcommand{\HELM}{Helm}
\newcommand{\HINT}{\textsc{Hint}}
\newcommand{\IN}{\ensuremath{\dN}}
\newcommand{\INSTANCE}{\textsc{Instance}}
\newcommand{\IR}{\ensuremath{\dR}}
\newcommand{\IZ}{\ensuremath{\dZ}}
\newcommand{\LIBXSLT}{LibXSLT}
\newcommand{\LOCATE}{\textsc{Locate}}
\newcommand{\MATCH}{\textsc{Match}}
\newcommand{\MATHML}{MathML}
\newcommand{\MATITA}{Matita}
\newcommand{\MATITAC}{\texttt{matitac}}
\newcommand{\MATITADEP}{\texttt{matitadep}}
\newcommand{\METAHEADING}{Symbol & Position \\ \hline\hline}
\newcommand{\MOWGLI}{MoWGLI}
\newcommand{\NAT}{\ensuremath{\mathit{nat}}}
\newcommand{\NATIND}{\mathit{nat\_ind}}
\newcommand{\NUPRL}{NuPRL}
\newcommand{\OCAML}{OCaml}
\newcommand{\PROP}{\mathit{Prop}}
\newcommand{\REF}[3]{\ensuremath{\mathit{Ref}_{#1}(#2,#3)}}
\newcommand{\TEXMACRO}[1]{\texttt{\char92 #1}}
\newcommand{\UWOBO}{UWOBO}
\newcommand{\GETTER}{Getter}
\newcommand{\WHELP}{Whelp}
\newcommand{\DOT}{\ensuremath{\mbox{\textbf{.}}}}
\newcommand{\SEMICOLON}{\ensuremath{\mbox{\textbf{;}}}}
\newcommand{\BRANCH}{\ensuremath{\mbox{\textbf{[}}}}
\newcommand{\SHIFT}{\ensuremath{\mbox{\textbf{\textbar}}}}
\newcommand{\POS}[1]{\ensuremath{#1\mbox{\textbf{:}}}}
\newcommand{\MERGE}{\ensuremath{\mbox{\textbf{]}}}}
\newcommand{\FOCUS}[1]{\ensuremath{\mathtt{focus}~#1}}
\newcommand{\UNFOCUS}{\ensuremath{\mathtt{unfocus}}}
\newcommand{\SKIP}{\MATHTT{skip}}
\newcommand{\TACTIC}[1]{\ensuremath{\mathtt{tactic}~#1}}

\newcommand{\NT}[1]{\ensuremath{\langle\mathit{#1}\rangle}}
\newcommand{\URI}[1]{\texttt{#1}}
\newcommand{\OP}[1]{``\texttt{#1}''}
\newcommand{\FILE}[1]{\texttt{#1}}
\newcommand{\NOTE}[1]{\ednote{#1}{}}
\newcommand{\TODO}[1]{\textbf{TODO: #1}}

\definecolor{gray}{gray}{0.85} % 1 -> white; 0 -> black

\newenvironment{grafite}{\VerbatimEnvironment
 \begin{SaveVerbatim}{boxtmp}}%
 {\end{SaveVerbatim}\setlength{\fboxsep}{3mm}%
  \begin{center}
   \fcolorbox{black}{gray}{\BUseVerbatim[boxwidth=0.9\linewidth]{boxtmp}}
  \end{center}}

\newcounter{pass}
\newcommand{\PASS}{\stepcounter{pass}\arabic{pass}}

\newsavebox{\tmpxyz}
\newcommand{\sequent}[2]{
  \savebox{\tmpxyz}[0.9\linewidth]{
    \begin{minipage}{0.9\linewidth}
      \ensuremath{#1} \\
      \rule{3cm}{0.03cm}\\
      \ensuremath{#2}
    \end{minipage}}\setlength{\fboxsep}{3mm}%
  \begin{center}
   \fcolorbox{black}{gray}{\usebox{\tmpxyz}}
  \end{center}}

\bibliographystyle{alpha}

\begin{document}

\begin{opening}
 \title{The \MATITA{} Proof Assistant}

 \author{Andrea \surname{Asperti} \email{asperti@cs.unibo.it}}
 \author{Claudio \surname{Sacerdoti Coen} \email{sacerdot@cs.unibo.it}}
 \author{Enrico \surname{Tassi} \email{tassi@cs.unibo.it}}
 \author{Stefano \surname{Zacchiroli} \email{zacchiro@cs.unibo.it}}

 \institute{Department of Computer Science, University of Bologna\\
 Mura Anteo Zamboni, 7 --- 40127 Bologna, ITALY}

 \runningtitle{The \MATITA{} proof assistant}
 \runningauthor{Asperti, Sacerdoti Coen, Tassi, Zacchiroli}

 \begin{motto}
  ``We are nearly bug-free'' -- \emph{CSC, Oct 2005}
 \end{motto}

 \begin{abstract}
  \TODO{scrivere abstract}
 \end{abstract}

 \keywords{Proof Assistant, Mathematical Knowledge Management, XML, Authoring,
 Digital Libraries}
\end{opening}

% toc & co: to be removed in the final paper version
\tableofcontents
\listoffigures
\listoftables

\section{Introduction}
\label{sec:intro}

\MATITA{} is the Proof Assistant under development by the \HELM{}
team~\cite{mkm-helm} at the University of Bologna, under the direction of
Prof.~Asperti. The paper describes the overall architecture of
the system, focusing on its most distinctive and innovative 
features.

\subsection{Historical Perspective}
The origins of \MATITA{} go back to 1999. At the time we were mostly 
interested to develop tools and techniques to enhance the accessibility
via Web of formal libraries of mathematics. Due to its dimension, the
library of the \COQ~\cite{CoqManual} proof assistant (of the order of 35'000 theorems) 
was chosen as a privileged test bench for our work, although experiments
have been also conducted with other systems, and notably 
with \NUPRL~\cite{nuprl-book}.
The work, mostly performed in the framework of the recently concluded 
European project IST-33562 \MOWGLI~\cite{pechino}, mainly consisted in the 
following steps:
\begin{itemize}
\item exporting the information from the internal representation of
 \COQ{} to a system and platform independent format. Since XML was at the 
time an emerging standard, we naturally adopted this technology, fostering
a content-centric architecture~\cite{content-centric} where the documents
of the library were the the main components around which everything else 
has to be build;
\item developing indexing and searching techniques supporting semantic
 queries to the library; 
\item developing languages and tools for a high-quality notational 
rendering of mathematical information\footnote{We have been 
active in the \MATHML{} Working group since 1999.}; 
\end{itemize}

According to our content-centric commitment, the library exported from
\COQ{} was conceived as being distributed and most of the tools were developed
as Web services. The user could interact with the library and the tools by
means of a Web interface that orchestrates the Web services.

The Web services and the other tools have been implemented as front-ends
to a set of software components, collectively called the \HELM{} components.
At the end of the \MOWGLI{} project we already disposed of the following
tools and software components:
\begin{itemize}
\item XML specifications for the Calculus of Inductive Constructions,
with components for parsing and saving mathematical objects in such a
format~\cite{exportation-module};
\item metadata specifications with components for indexing and querying the
XML knowledge base;
\item a proof checker library (i.e. the {\em kernel} of a proof assistant), 
implemented to check that we exported from the \COQ{} library all the 
logically relevant content;
\item a sophisticated parser (used by the search engine), able to deal 
with potentially ambiguous and incomplete information, typical of the 
mathematical notation~\cite{disambiguation};
\item a {\em refiner} library, i.e. a type inference system, based on
partially specified terms, used by the disambiguating parser;
\item complex transformation algorithms for proof rendering in natural
language~\cite{remathematization};
\item an innovative, \MATHML-compliant rendering widget for the GTK 
graphical environment~\cite{padovani}, supporting 
high-quality bidimensional
rendering, and semantic selection, i.e. the possibility to select semantically
meaningful rendering expressions, and to paste the respective content into
a different text area.
\end{itemize}
Starting from all this, developing our own proof assistant was not
too far away: essentially, we ``just'' had to
add an authoring interface, and a set of functionalities for the
overall management of the library, integrating everything into a
single system. \MATITA{} is the result of this effort. 

\subsection{The system}

\MATITA{} is a proof assistant (also called interactive theorem prover).
It is based on the Calculus of (Co)Inductive Constructions
(CIC)~\cite{Werner} that is a dependently typed lambda-calculus \`a la
Church enriched with primitive inductive and co-inductive data types.
Via the Curry-Howard isomorphism, the calculus can be seen as a very
rich higher order logic and proofs can be simply represented and
stored as lambda-terms. \COQ{} and Lego are other systems that adopt
(variations of) CIC as their foundation.

The proof language of \MATITA{} is procedural, in the tradition of the LCF
theorem prover. Coq, NuPRL, PVS, Isabelle are all examples of others systems
whose proof language is procedural. Traditionally, in a procedural system
the user interacts only with the \emph{script}, while proof terms are internal
records kept by the system. On the contrary, in \MATITA{} proof terms are
praised as declarative versions of the proof. With this role, they are the
primary mean of communication of proofs (once rendered to natural language
for human audiences).

The user interfaces now adopted by all the proof assistants based on a
procedural proof language have been inspired by the CtCoq pioneering
system~\cite{ctcoq1}. One successful incarnation of the ideas introduced
by CtCoq is the Proof General generic interface~\cite{proofgeneral},
that has set a sort of
standard way to interact with the system. Several procedural proof assistants
have either adopted or cloned Proof General as their main user interface.
The authoring interface of \MATITA{} is a clone of the Proof General interface.

\TODO{item che seguono:}
\begin{itemize}
 \item sistema indipendente (da \COQ)
 \item compatibilit\`a con sistemi legacy
\end{itemize}

\subsection{Relationship with \COQ{}}

At first sight, \MATITA{} looks as (and partly is) a \COQ{} clone. This is
more the effect of the circumstances of its creation described 
above than the result of a deliberate design. In particular, we
(essentially) share the same foundational dialect of \COQ{} (the
Calculus of (Co)Inductive Constructions), the same implementation
language (\OCAML{}), and the same (script based) authoring philosophy.
However, the analogy essentially stops here and no code is shared by the
two systems.

In a sense; we like to think of \MATITA{} as the way \COQ{} would 
look like if entirely rewritten from scratch: just to give an
idea, although \MATITA{} currently supports almost all functionalities of
\COQ{}, it links 60'000 lines of \OCAML{} code, against the 166'000 lines linked
by \COQ{} (and we are convinced that, starting from scratch again,
we could reduce our code even further in sensible way).

Moreover, the complexity of the code of \MATITA{} is greatly reduced with
respect to \COQ. For instance, the API of the components of \MATITA{} comprise
989 functions, to be compared with the 4'286 functions of \COQ.

Finally, \MATITA{} has several innovative features over \COQ{} that derive
from the integration of Mathematical Knowledge Management tools with proof
assistants. Among them, the advanced indexing tools over the library and
the parser for ambiguous mathematical notation.

The size and complexity improvements over \COQ{} must be understood
historically. \COQ{} is a quite old
system whose development started 20 years ago.
\NOTE{Zack: verificato su Coq'Art, \`e iniziato nel 1985-86}
Since then
several developers have took over the code and several new research ideas
that were not considered in the original architecture have been experimented
and integrated in the system. Moreover, there exists a lot of developments
for \COQ{} that require backward compatibility between each pair of releases;
since many central functionalities of a proof assistant are based on heuristics
or arbitrary choices to overcome undecidability (e.g. for higher order
unification), changing these functionalities maintaining backward compatibility
is very difficult. Finally, the code of \COQ{} has been greatly optimized
over the years; optimization reduces maintainability and rises the complexity
of the code.

In writing \MATITA{} we have not been hindered by backward compatibility and
we have took advantage of the research results and experiences previously
developed by others, comprising the authors of \COQ. Moreover, starting from
scratch, we have designed in advance the architecture and we have split
the code in coherent minimally coupled components.

In the future we plan to exploit \MATITA{} as a test bench for new ideas and
extensions. Keeping the single components and the whole architecture as
simple as possible is thus crucial to foster future experiments and to
allow other developers to quickly understand our code and contribute.

%For direct experience of the authors, the learning curve to understand and
%be able to contribute to \COQ{}'s code is quite steep and requires direct
%and frequent interactions with \COQ{} developers.

\section{Architecture}
\label{architettura}

\begin{figure}[!ht]
 \begin{center}
  \includegraphics[width=0.9\textwidth,height=0.8\textheight]{pics/libraries-clusters}
  \caption[\MATITA{} components and related applications]{\MATITA{}
   components and related applications, with thousands of line of
   codes (klocs)}
  \label{fig:libraries}
 \end{center}
\end{figure}

Fig.~\ref{fig:libraries} shows the architecture of the \emph{\components}
(circle nodes) and \emph{applications} (squared nodes) developed in the HELM
project. Each node is annotated with the number of lines of source code
(comprising comments).

Applications and \components{} depend over other \components{} forming a
directed acyclic graph (DAG). Each \component{} can be decomposed in
a a set of \emph{modules} also forming a DAG.

Modules and \components{} provide coherent sets of functionalities
at different scales. Applications that require only a few functionalities
depend on a restricted set of \components{}.

Only the proof assistant \MATITA{} and the \WHELP{} search engine are
applications meant to be used directly by the user. All the other applications
are Web services developed in the HELM and MoWGLI projects and already described
elsewhere. In particular:
\begin{itemize}
 \item The \emph{\GETTER} is a Web service to retrieve an (XML) document
   from a physical location (URL) given its logical name (URI). The Getter is
   responsible of updating a table that maps URIs to URLs. Thanks to the Getter
   it is possible to work on a logically monolithic library that is physically
   distributed on the network. More information on the Getter can be found
   in~\cite{zack-master}.
 \item \emph{\WHELP} is a search engine to index and locate mathematical
   notions (axioms, theorems, definitions) in the logical library managed
   by the Getter. Typical examples of a query to Whelp are queries that search
   for a theorem that generalize or instantiate a given formula, or that
   can be immediately applied to prove a given goal. The output of Whelp is
   an XML document that lists the URIs of a complete set of candidates that
   are likely to satisfy the given query. The set is complete in the sense
   that no notion that actually satisfies the query is thrown away. However,
   the query is only approximated in the sense that false matches can be
   returned. Whelp has been described in~\cite{whelp}.
 \item \emph{\UWOBO} is a Web service that, given the URI of a mathematical
   notion in the distributed library, renders it according to the user provided
   two dimensional mathematical notation. \UWOBO{} may also embed the rendering
   of mathematical notions into arbitrary documents before returning them.
   The Getter is used by \UWOBO{} to retrieve the document to be rendered.
   \UWOBO{} has been described in~\cite{zack-master}.
 \item The \emph{Proof Checker} is a Web service that, given the URI of
   notion in the distributed library, checks its correctness. Since the notion
   is likely to depend in an acyclic way over other notions, the proof checker
   is also responsible of building in a top-down way the DAG of all
   dependencies, checking in turn every notion for correctness.
   The proof checker has been described in~\cite{zack-master}.
 \item The \emph{Dependency Analyzer} is a Web service that can produce
   a textual or graphical representation of the dependencies of an object.
   The dependency analyzer has been described in~\cite{zack-master}.
\end{itemize}

The dependency of a \component{} or application over another \component{} can
be satisfied by linking the \component{} in the same executable.
For those \components{} whose functionalities are also provided by the
aforementioned Web services, it is also possible to link stub code that
forwards the request to a remote Web service. For instance, the Getter
is just a wrapper to the \GETTER{} \component{} that allows the
\component{} to be used as a Web service. \MATITA{} can directly link the code
of the \GETTER{} \component, or it can use a stub library with the same
API that forwards every request to the Getter.

To better understand the architecture of \MATITA{} and the role of each
\component, we can focus on the representation of the mathematical information.
\MATITA{} is based on (a variant of) the Calculus of (Co)Inductive
Constructions (CIC). In CIC terms are used to represent mathematical
formulae, types and proofs. \MATITA{} is able to handle terms at
four different levels of specification. On each level it is possible to provide
a different set of functionalities. The four different levels are:
fully specified terms; partially specified terms; 
content level terms; presentation level terms.

\subsection{Fully specified terms}
\label{sec:fullyintro}

 \emph{Fully specified terms} are CIC terms where no information is
   missing or left implicit. A fully specified term should be well-typed.
   The mathematical notions (axioms, definitions, theorems) that are stored
   in our mathematical library are fully specified and well-typed terms.
   Fully specified terms are extremely verbose (to make type-checking
   decidable). Their syntax is fixed and does not resemble the usual
   extendible mathematical notation. They are not meant for direct user
   consumption.

   The \texttt{cic} \component{} defines the data type that represents CIC terms
   and provides a parser for terms stored in an XML format.

   The most important \component{} that deals with fully specified terms is
   \texttt{cic\_proof\_checking}. It implements the procedure that verifies
   if a fully specified term is well-typed. It also implements the
   \emph{conversion} judgement that verifies if two given terms are
   computationally equivalent (i.e. they share the same normal form).

   Terms may reference other mathematical notions in the library.
   One commitment of our project is that the library should be physically
   distributed. The \GETTER{} \component{} manages the distribution,
   providing a mapping from logical names (URIs) to the physical location
   of a notion (an URL). The \texttt{urimanager} \component{} provides the URI
   data type and several utility functions over URIs. The
   \texttt{cic\_proof\_checking} \component{} calls the \GETTER{}
   \component{} every time it needs to retrieve the definition of a mathematical
   notion referenced by a term that is being type-checked. 

   The Proof Checker is the Web service that provides an interface
   to the \texttt{cic\_proof\_checking} \component.

   We use metadata and a sort of crawler to index the mathematical notions
   in the distributed library. We are interested in retrieving a notion
   by matching, instantiation or generalization of a user or system provided
   mathematical formula. Thus we need to collect metadata over the fully
   specified terms and to store the metadata in some kind of (relational)
   database for later usage. The \texttt{hmysql} \component{} provides
   a simplified
   interface to a (possibly remote) MySql database system used to store the
   metadata. The \texttt{metadata} \component{} defines the data type of the
   metadata
   we are collecting and the functions that extracts the metadata from the
   mathematical notions (the main functionality of the crawler).
   The \texttt{whelp} \component{} implements a search engine that performs
   approximated queries by matching/instantiation/generalization. The queries
   operate only on the metadata and do not involve any actual matching
   (that will be described later on and that is implemented in the
    \texttt{cic\_unification} \component). Not performing any actual matching
   the query only returns a complete and hopefully small set of matching
   candidates. The process that has issued the query is responsible of
   actually retrieving from the distributed library the candidates to prune
   out false matches if interested in doing so.

   The Whelp search engine is the Web service that provides an interface to
   the \texttt{whelp} \component.

   According to our vision, the library is developed collaboratively so that
   changing or removing a notion can invalidate other notions in the library.
   Moreover, changing or removing a notion requires a corresponding change
   in the metadata database. The \texttt{library} \component{} is responsible
   of preserving the coherence of the library and the database. For instance,
   when a notion is removed, all the notions that depend on it and their
   metadata are removed from the library. This aspect will be better detailed
   in Sect.~\ref{sec:libmanagement}.
   
\subsection{Partially specified terms}
\label{sec:partiallyintro}

\emph{Partially specified terms} are CIC terms where subterms can be omitted.
Omitted subterms can bear no information at all or they may be associated to
a sequent. The formers are called \emph{implicit terms} and they occur only
linearly. The latters may occur multiple times and are called
\emph{metavariables}. An \emph{explicit substitution} is applied to each
occurrence of a metavariable. A metavariable stand for a term whose type is
given by the conclusion of the sequent. The term must be closed in the
context that is given by the ordered list of hypotheses of the sequent.
The explicit substitution instantiates every hypothesis with an actual
value for the variable bound by the hypothesis.

Partially specified terms are not required to be well-typed. However a
partially specified term should be \emph{refinable}. A \emph{refiner} is
a type-inference procedure that can instantiate implicit terms and
metavariables and that can introduce \emph{implicit coercions} to make a
partially specified term well-typed. The refiner of \MATITA{} is implemented
in the \texttt{cic\_unification} \component. As the type checker is based on
the conversion check, the refiner is based on \emph{unification} that is
a procedure that makes two partially specified term convertible by instantiating
as few as possible metavariables that occur in them.

Since terms are used in CIC to represent proofs, correct incomplete
proofs are represented by refinable partially specified terms. The metavariables
that occur in the proof correspond to the conjectures still to be proved.
The sequent associated to the metavariable is the conjecture the user needs to
prove.

\emph{Tactics} are the procedures that the user can apply to progress in the
proof. A tactic proves a conjecture possibly creating new (and hopefully
simpler) conjectures. The implementation of tactics is given in the
\texttt{tactics} \component. It is heavily based on the refinement and
unification procedures of the \texttt{cic\_unification} \component.

The \texttt{grafite} \component{} defines the abstract syntax tree (AST) for the
commands of the \MATITA{} proof assistant. Most of the commands are tactics.
Other commands are used to give definitions and axioms or to state theorems
and lemmas. The \texttt{grafite\_engine} \component{} is the core of \MATITA{}.
It implements the semantics of each command in the grafite AST as a function
from status to status.  It implements also an undo function to go back to
previous statuses.

As fully specified terms, partially specified terms are not well suited
for user consumption since their syntax is not extendible and it is not
possible to adopt the usual mathematical notation. However they are already
an improvement over fully specified terms since they allow to omit redundant
information that can be inferred by the refiner.

\subsection{Content level terms}
\label{sec:contentintro}

The language used to communicate proofs and especially formulae with the
user does not only needs to be extendible and accommodate the usual mathematical
notation. It must also reflect the comfortable degree of imprecision and
ambiguity that the mathematical language provides.

For instance, it is common practice in mathematics to speak of a generic
equality that can be used to compare any two terms. However, it is well known
that several equalities can be distinguished as soon as we care for decidability
or for their computational properties. For instance equality over real
numbers is well known to be undecidable, whereas it is decidable over
rational numbers.

Similarly, we usually speak of natural numbers and their operations and
properties without caring about their representation. However the computational
properties of addition over the binary representation are very different from
those of addition over the unary representation. And addition over two natural
numbers is definitely different from addition over two real numbers.

Formal mathematics cannot hide these differences and obliges the user to be
very precise on the types he is using and their representation. However,
to communicate formulae with the user and with external tools, it seems good
practice to stick to the usual imprecise mathematical ontology. In the
Mathematical Knowledge Management community this imprecise language is called
the \emph{content level} representation of formulae.

In \MATITA{} we provide two translations: from partially specified terms
to content level terms and the other way around. The first translation can also
be applied to fully specified terms since a fully specified term is a special
case of partially specified term where no metavariable or implicit term occurs.

The translation from partially specified terms to content level terms must
discriminate between terms used to represent proofs and terms used to represent
formulae. The firsts are translated to a content level representation of
proof steps that can easily be rendered in natural language. The representation
adopted has greatly influenced the OMDoc~\cite{omdoc} proof format that is now
isomorphic to it. Terms that represent formulae are translated to \MATHML{}
Content formulae. \MATHML{} Content~\cite{mathml} is a W3C standard
for the representation of content level formulae in an XML extensible format.

The translation to content level is implemented in the
\texttt{acic\_content} \component. Its input are \emph{annotated partially
specified terms}, that are maximally unshared
partially specified terms enriched with additional typing information for each
subterm. This information is used to discriminate between terms that represent
proofs and terms that represent formulae. Part of it is also stored at the
content level since it is required to generate the natural language rendering
of proofs. The terms need to be maximally unshared (i.e. they must be a tree
and not a DAG). The reason is that to the occurrences of a subterm in
two different positions we need to associate different typing informations.
This association is made easier when the term is represented as a tree since
it is possible to label each node with an unique identifier and associate
the typing information using a map on the identifiers.
The \texttt{cic\_acic} \component{} unshares and annotates terms. It is used
by the \texttt{library} \component{} since fully specified terms are stored
in the library in their annotated form.

We do not provide yet a reverse translation from content level proofs to
partially specified terms. But in \texttt{cic\_disambiguation} we do provide
the reverse translation for formulae. The mapping from
content level formulae to partially specified terms is not unique due to
the ambiguity of the content level. As a consequence the translation
is guided by an \emph{interpretation}, that is a function that chooses for
every ambiguous formula one partially specified term. The
\texttt{cic\_disambiguation} \component{} implements the
disambiguation algorithm we presented in~\cite{disambiguation} that is
responsible of building in an efficient way the set of all ``correct''
interpretations. An interpretation is correct if the partially specified term
obtained using the interpretation is refinable.

In Sect.~\ref{sec:partiallyintro} the last section we described the semantics of
a command as a
function from status to status. We also suggested that the formulae in a
command are encoded as partially specified terms. However, consider the
command ``\texttt{replace} $x$ \texttt{with} $y^2$''. Until the occurrence
of $x$ to be replaced is located, its context is unknown. Since $y^2$ must
replace $x$ in that context, its encoding as a term cannot be computed
until $x$ is located. In other words, $y^2$ must be disambiguated in the
context of the occurrence $x$ it must replace.

The elegant solution we have implemented consists in representing terms
in a command as functions from a context to a partially refined term. The
function is obtained by partially applying our disambiguation function to
the content term to be disambiguated. Our solution should be compared with
the one adopted in the Coq system, where ambiguity is only relative to De Brujin
indexes. In Coq variables can be bound either by name or by position. A term
occurring in a command has all its variables bound by name to avoid the need of
a context during disambiguation.  Moreover, this makes more complex every
operation over terms (i.e. according to our architecture every module that
depends on \texttt{cic}) since the code must deal consistently with both kinds
of binding. Also, this solution cannot cope with other forms of ambiguity (as
the context dependent meaning of the exponent in the previous example).

\subsection{Presentation level terms}
\label{sec:presentationintro}

Content level terms are a sort of abstract syntax trees for mathematical
formulae and proofs. The concrete syntax given to these abstract trees
is called \emph{presentation level}.

The main important difference between the content level language and the
presentation level language is that only the former is extendible. Indeed,
the presentation level language is a finite language that comprises all
the usual mathematical symbols. Mathematicians invent new notions every
single day, but they stick to a set of symbols that is more or less fixed.

The fact that the presentation language is finite allows the definition of
standard languages. In particular, for formulae we have adopt \MATHML{}
Presentation~\cite{mathml} that is an XML dialect standardized by the W3C. To
visually
represent proofs it is enough to embed formulae in plain text enriched with
formatting boxes. Since the language of formatting boxes is very simple,
many equivalent specifications exist and we have adopted our own, called
\BOXML.

The \texttt{content\_pres} \component{} contains the implementation of the
translation from content level terms to presentation level terms. The
rendering of presentation level terms is left to the application that uses
the \component. However, in the \texttt{hgdome} \component{} we provide a few
utility functions to build a \GDOME~\cite{gdome2} \MATHML+\BOXML{} tree from our
presentation
level terms. \GDOME{} \MATHML+\BOXML{} trees can be rendered by the
\GTKMATHVIEW{}
widget developed by Luca Padovani~\cite{padovani}. The widget is
particularly interesting since it allows to implement \emph{semantic
selection}.

Semantic selection is a technique that consists in enriching the presentation
level terms with pointers to the content level terms and to the partially
specified terms they correspond to. Highlight of formulae in the widget is
constrained to selection of meaningful expressions, i.e. expressions that
correspond to a lower level term, that is a content term or a partially or
fully specified term.
Once the rendering of a lower level term is
selected it is possible for the application to retrieve the pointer to the
lower level term. An example of applications of semantic selection is
\emph{semantic cut\&paste}: the user can select an expression and paste it
elsewhere preserving its semantics (i.e. the partially specified term),
possibly performing some semantic transformation over it (e.g. renaming
variables that would be captured or lambda-lifting free variables).

The reverse translation from presentation level terms to content level terms
is implemented by a parser that is also found in \texttt{content\_pres}.
Differently from the translation from content level terms to partially
refined terms, this translation is not ambiguous. The reason is that the
parsing tool we have adopted (CamlP4) is not able to parse ambiguous
grammars. Thus we require the mapping from presentation level terms
(concrete syntax) to content level terms (abstract syntax) to be unique.
This means that the user must fix once and for all the associativity and
precedence level of every operator he is using. In practice this limitation
does not seem too strong. The reason is that the target of the
translation is an ambiguous language and the user is free to associate
to every content level term several different interpretations (as a
partially specified term).

Both the direct and reverse translation from presentation to content level
terms are parameterized over the user provided mathematical notation. 
The \texttt{lexicon} \component{} is responsible of managing the lexicon,
that is the set of active notations. It defines an abstract syntax tree
of commands to declare and activate new notations and it implements the
semantics of these commands. It also implements undoing of the semantic
actions. Among the commands there are hints to the
disambiguation algorithm that are used to control and speed up disambiguation.
These mechanisms will be further discussed in Sect.~\ref{sec:disambiguation}.

Finally, the \texttt{grafite\_parser} \component{} implements a parser for
the concrete syntax of the commands of \MATITA. The parser process a stream
of characters and returns a stream of abstract syntax trees (the ones
defined by the \texttt{grafite} component and whose semantics is given
by \texttt{grafite\_engine}). When the parser meets a command that changes
the lexicon, it invokes the \texttt{lexicon} \component{} to immediately
process the command. When the parser needs to parse a term at the presentation
level, it invokes the already described parser for terms contained in
\texttt{content\_pres}.

The \MATITA{} proof assistant and the \WHELP{} search engine are both linked
against the \texttt{grafite\_parser} \components{}
since they provide an interface to the user. In both cases the formulae
written by the user are parsed using the \texttt{content\_pres} \component{} and
then disambiguated using the \texttt{cic\_disambiguation} \component.  However,
only \MATITA{} is linked against the \texttt{grafite\_engine} and
\texttt{tactics} components (summing up to a total of 11'200 lines of code)
since \WHELP{} can only execute those ASTs that correspond to queries
(implemented in the \texttt{whelp} component).

The \UWOBO{} Web service wraps the \texttt{content\_pres} \component,
providing a rendering service for the documents in the distributed library.
To render a document given its URI, \UWOBO{} retrieves it using the
\GETTER{} obtaining a document with fully specified terms. Then it translates
it to the presentation level passing through the content level. Finally
it returns the result document to be rendered by the user's
browser.\NOTE{\TODO{manca la passata verso HTML}}

The \components{} not yet described (\texttt{extlib}, \texttt{xml},
\texttt{logger}, \texttt{registry} and \texttt{utf8\_macros}) are 
minor \components{} that provide a core of useful functions and basic
services missing from the standard library of the programming language.
%In particular, the \texttt{xml} \component{} is used to easily represent,
%parse and pretty-print XML files.

\section{The interface to the library}
\label{sec:library}

A proof assistant provides both an interface to interact with its library and
an \emph{authoring} interface to develop new proofs and theories. According
to its historical origins, \MATITA{} strives to provide innovative
functionalities for the interaction with the library. It is more traditional
in its script based authoring interface.

In the remaining part of the paper we focus on the user view of \MATITA{}.
This section is devoted to the aspects of the tool that arise from the
document centric approach to the library. Sect.~\ref{sec:authoring} describes
the peculiarities of the authoring interface.

The library of \MATITA{} comprises mathematical concepts (theorems,
axioms, definitions) and notation. The concepts are authored sequentially
using scripts that are (ordered) sequences of procedural commands.
However, once they are produced we store them independently in the library.
The only relation implicitly kept between the notions are the logical,
acyclic dependencies among them. This way the library forms a global (and
distributed) hypertext. Several useful operations can be implemented on the
library only, regardless of the scripts. Examples of such operations
implemented in \MATITA{} are: searching and browsing (see Sect.~\ref{sec:indexing});
disambiguation of content level terms (see Sect.~\ref{sec:disambiguation});
automatic proof searching (see Sect.~\ref{sec:automation}).

The key requisite for the previous operations is that the library must
be fully accessible and in a logically consistent state. To preserve
consistency, a concept cannot be altered or removed unless the part of the
library that depends on it is modified accordingly. To allow incremental
changes and cooperative development, consistent revisions are necessary.
For instance, to modify a definition, the user could fork a new version
of the library where the definition is updated and all the concepts that
used to rely on it are absent. The user is then responsible to restore
the removed part in the new branch, merging the branch when the library is
fully restored.

To implement the proposed versioning system on top of a standard one
it is necessary to implement \emph{invalidation} first. Invalidation
is the operation that locates and removes from the library all the concepts
that depend on a given one. As described in Sect.~\ref{sec:libmanagement} removing
a concept from the library also involves deleting its metadata from the
database.

For non collaborative development, full versioning can be avoided, but
invalidation is still required. Since nobody else is relying on the
user development, the user is free to change and invalidate part of the library
without branching. Invalidation is still necessary to avoid using a
concept that is no longer valid.
So far, in \MATITA{} we address only this non collaborative scenario
(see Sect.~\ref{sec:libmanagement}). Collaborative development and versioning
is still under design.

Scripts are not seen as constituents of the library. They are not published
and indexed, so they cannot be searched or browsed using \HELM{} tools.
However, they play a central role for the maintenance of the library.
Indeed, once a notion is invalidated, the only way to restore it is to
fix the possibly broken script that used to generate it.
Moreover, during the authoring phase, scripts are a natural way to
group notions together. They also constitute a less fine grained clustering
of notions for invalidation.

In the rest of this section we present in more details the functionalities of
\MATITA{} related to library management and exploitation.
Sect.~\ref{sec:authoring} is devoted to the description of the peculiarities of
the \MATITA{} authoring interface.

\subsection{Indexing and searching}
\label{sec:indexing}

\subsection{Disambiguation}
\label{sec:disambiguation}

Software applications that involve input of mathematical content should strive
to require the user as less drift from informal mathematics as possible. We
believe this to be a fundamental aspect of such applications user interfaces.
Being that drift in general very large when inputing
proofs~\cite{debrujinfactor}, in \MATITA{} we achieved good results for
mathematical formulae which can be input using a \TeX-like encoding (the
concrete syntax corresponding to presentation level terms) and are then
translated (in multiple steps) to partially specified terms as sketched in
Sect.~\ref{sec:contentintro}.

The key component of the translation is the generic disambiguation algorithm
implemented in the \texttt{disambiguation} component of Fig.~\ref{fig:libraries}
and presented in~\cite{disambiguation}. In this section we present how to use
such an algorithm in the context of the development of a library of formalized
mathematics. We will see that using multiple passes of the algorithm, varying
some of its parameters, helps in keeping the input terse without sacrificing
expressiveness.

\subsubsection{Disambiguation aliases}
\label{sec:disambaliases}

Consider the following command to state a theorem over integer numbers.

\begin{grafite}
theorem Zlt_compat:
  \forall x, y, z. x < y \to y < z \to x < z.
\end{grafite}

The symbol \OP{<} is likely to be overloaded in the library
(at least over natural numbers). 
Thus, according to the disambiguation algorithm, two different
refinable partially specified terms could be associated to it.
\MATITA{} asks the user what interpretation he meant. However, to avoid
posing the same question in case of a future re-execution (e.g. undo/redo),
the choice must be recorded. Scripts need to be re-executed after invalidation.
Therefore the choice record must be permanently stored somewhere. The most
natural place is in the script itself.

In \MATITA{} disambiguation is governed by \emph{disambiguation aliases}.
They are mappings, stored in the library, from ambiguity sources
(identifiers, symbols and literal numbers at the content level) to partially
specified terms. In case of overloaded sources there exists multiple aliases
with the same source. It is possible to record \emph{disambiguation
preferences} to select one of the aliases of an overloaded source.

Preferences can be explicitely given in the script (using the
misleading \texttt{alias} commands), but
are also implicitly added when a new concept is introduced (\emph{implicit
preferences}) or after a sucessfull disambiguation that did not require
user interaction. Explicit preferences are added automatically by \MATITA{} to
record the disambiguation choices of the user. For instance, after the
disambiguation of the command above, the script is altered as follows:

\begin{grafite}
alias symbol "lt" = "integer 'less than'".
theorem Zlt_compat:
  \forall x, y, z. x < y \to y < z \to x < z.
\end{grafite}

The ``alias'' command in the example sets the preferred alias for the
\OP{lt} symbol.

Implicit preferences for new concepts are set since a concept just defined is
likely to be the preferred one in the rest of the script. Implicit preferences
learned from disambiguation of previous commands grant the coherence of
the disambiguation in the rest of the script and speed up disambiguation
reducing the search space.

Disambiguation preferences are included in the lexicon status
(see Sect.~\ref{sec:presentationintro} that is part of the authoring interface
status.  Unlike aliases, they are not part of the library.

When starting a new authoring session the set of disambiguation preferences
is empty. Until it contains a preference for each overloaded symbol to be
used in the script, the user can be faced with questions from the disambiguator.
To reduce the likelyhood of user interactions, we introduced
the \texttt{include} command. With \texttt{include} it is possible to import
at once in the current session the set of preferences that was in effect
at the end of the execution of a given script.

Preferences can be changed. For instance, at the start of the development
of integer numbers the preference for the symbol \OP{<} is likely
to be the one over natural numbers; sooner or later it will be set to the one
over integer numbers.

Nothing forbids the set of preferences to become incoherent. For this reason
the disambiguator cannot always respect the user preferences.
For example consider:

\begin{grafite}
theorem Zlt_mono: \forall x, y, k. x < y \to x < y + k.
\end{grafite}

No refinable partially specified term corresponds to the preferences:
\OP{+} over natural numbers, \OP{<} over integer numbers.
When the disambiguator fails, disambiguation is tried again with a less
strict set of preferences. Thus disambiguation is organized in
multiple \emph{passes}. In the first pass, called \emph{mono-preferences},
we consider only the aliases corresponding to the current preferences.
In the second pass, called \emph{multi-preferences},
we consider every alias corresponding to a current or past preference.
For instance, in the example above disambiguation succeeds in the
multi-preference pass 

\TODO{Disambiguazione: cambiato solo FINQUI}

and suppose that the \OP{+} operator is defined only on natural numbers. If
the alias for \OP{<} points to the integer version of the operator, no
refinable partially specified term matching the term could be found.

For this reason we chose to attempt \emph{multiple disambiguation passes}. A
first pass attempts to disambiguate using the last available disambiguation
aliases (\emph{mono aliases} pass); in case of failure the next pass tries
disambiguation again forgetting the aliases and using the whole library to
retrieve interpretation for ambiguous expressions (\emph{library aliases} pass).
Since the latter pass may lead to too many choices we intertwined an additional
pass among the two which use as interpretations all the aliases coming for
included parts of the library (\emph{multi aliases} phase). This is the reason
why aliases are \emph{one-to-many} mappings instead of one-to-one. This choice
turned out to be a well-balanced trade-off among performances (earlier passes
fail quickly) and degree of ambiguity supported for presentation level terms.

\subsubsection{Operator instances}
\label{sec:disambinstances}

Let us suppose now we want to define a theorem relating ordering relations on
natural and integer numbers. The way we would like to write such a theorem (as
we can read it in the \MATITA{} standard library) is:

\begin{grafite}
include "Z/z.ma".
include "nat/orders.ma".
..
theorem lt_to_Zlt_pos_pos:
  \forall n, m: nat. n < m \to pos n < pos m. 
\end{grafite}

Unfortunately, none of the passes described above is able to disambiguate its
type, no matter how aliases are defined. This is because the \OP{<} operator
occurs twice in the content level term (it has two \emph{instances}) and two
different interpretations for it have to be used in order to obtain a refinable
partially specified term. To address this issue, we have the ability to consider
each instance of a single symbol as a different ambiguous expression in the
content level term, and thus we can assign a different interpretation to each of
them. A disambiguation pass which exploit this feature is said to be using
\emph{fresh instances}.

Fresh instances lead to a non negligible performance loss (since the choice of
an interpretation for one instances does not constraint the choice for the
others). For this reason we always attempt a fresh instances pass only after
attempting a non-fresh one.

\paragraph{One-shot aliases} Disambiguation aliases as seen so far are
instance-independent. However, aliases obtained as a result of a disambiguation
pass which uses fresh instances ought to be instance-dependent, that is: to
ensure a term can be disambiguated in a batch fashion we may need to state that
an \emph{i}-th instance of a symbol should be mapped to a given partially
specified term. Instance-depend aliases are meaningful only for the term whose
disambiguation generated it. For this reason we call them \emph{one-shot
aliases} and \MATITA{} does not use it to disambiguate further terms down in the
script.

\subsubsection{Implicit coercions}
\label{sec:disambcoercions}

Let us now consider a theorem about derivation:

\begin{grafite}
theorem power_deriv:
  \forall n: nat, x: R. d x ^ n dx = n * x ^ (n - 1).
\end{grafite}

and suppose there exists a \texttt{R \TEXMACRO{to} nat \TEXMACRO{to} R}
interpretation for \OP{\^}, and a real number interpretation for \OP{*}.
Mathematicians would write the term that way since it is well known that the
natural number \texttt{n} could be ``injected'' in \IR{} and considered a real
number for the purpose of real multiplication. The refiner of \MATITA{} supports
\emph{implicit coercions} for this reason: given as input the above content
level term, it will return a partially specified term where in place of
\texttt{n} the application of a coercion from \texttt{nat} to \texttt{R} appears
(assuming it has been defined as such of course).

Nonetheless coercions are not always desirable. For example, in disambiguating
\texttt{\TEXMACRO{forall} x: nat. n < n + 1} we do not want the term which uses
two coercions from \texttt{nat} to \texttt{R} around \OP{<} arguments to show up
among the possible partially specified term choices. For this reason in
\MATITA{} we always try first a disambiguation pass which require the refiner
not to use the coercions and only in case of failure we attempt a
coercion-enabled pass.

It is interesting to observe also the relationship among operator instances and
implicit coercions. Consider again the theorem \texttt{lt\_to\_Zlt\_pos\_pos},
which \MATITA{} disambiguated using fresh instances. In case there exists a
coercion from natural numbers to (positive) integers (which indeed does, it is
the \texttt{pos} constructor itself), the theorem can be disambiguated using
twice that coercion on the left hand side of the implication. The obtained
partially specified term however would not probably be the expected one, being a
theorem which prove a trivial implication. For this reason we choose to always
prefer fresh instances over implicit coercions, i.e. we always attempt
disambiguation passes with fresh instances and no implicit coercions before
attempting passes with implicit coercions.

\subsubsection{Disambiguation passes}
\label{sec:disambpasses}

According to the criteria described above in \MATITA{} we choose to perform the
sequence of disambiguation passes depicted in Tab.~\ref{tab:disambpasses}. In
our experience that choice gives reasonable performance and minimize the need of
user interaction during the disambiguation.

\begin{table}[ht]
 \caption{Sequence of disambiguation passes used in \MATITA.\strut}
 \label{tab:disambpasses} 
 \begin{center}
  \begin{tabular}{c|c|c|c}
   \multicolumn{1}{p{1.5cm}|}{\centering\raisebox{-1.5ex}{\textbf{Pass}}}
   & \multicolumn{1}{p{3.1cm}|}{\centering\textbf{Disambiguation aliases}}
   & \multicolumn{1}{p{2.5cm}|}{\centering\textbf{Operator instances}}
   & \multicolumn{1}{p{2.5cm}}{\centering\textbf{Implicit coercions}} \\
   \hline
   \PASS & Mono aliases   & Shared          & Disabled \\
   \PASS & Multi aliases  & Shared          & Disabled \\
   \PASS & Mono aliases   & Fresh instances & Disabled \\
   \PASS & Multi aliases  & Fresh instances & Disabled \\
   \PASS & Mono aliases   & Fresh instances & Enabled  \\
   \PASS & Multi aliases  & Fresh instances & Enabled  \\
   \PASS & Library aliases& Fresh instances & Enabled
  \end{tabular}
 \end{center}
\end{table}

\subsection{Generation and Invalidation}
\label{sec:libmanagement}

The aim of this section is to describe the way \MATITA{} 
preserves the consistency and the availability of the library
using the \WHELP{} technology, in response to the user alteration or 
removal of mathematical objects.

As already sketched in Sect.~\ref{sec:fullyintro} what we generate 
from a script is split among two storage media, a
classical filesystem and a relational database. The former is used to
store the XML encoding of the objects defined in the script, the
disambiguation aliases and the interpretation and notational convention defined,
while the latter is used to store all the metadata needed by
\WHELP{}.

While the consistency of the data store in the two media has
nothing to do with the nature of
the content of the library and is thus uninteresting (but really
tedious to implement and keep bug-free), there is a deeper
notion of mathematical consistency we need to provide. Each object
must reference only defined object (i.e. each proof must use only
already proved theorems). 

We will focus on how \MATITA{} ensures the interesting kind
of consistency during the formalization of a mathematical theory, 
giving the user the freedom of adding, removing, modifying objects
without loosing the feeling of an always visible and browsable
library.

\subsubsection{Compilation}

The typechecker component guarantees that if an object is well typed
it depends only on well typed objects available in the library,
that is exactly what we need to be sure that the logic consistency of
the library is preserved. We have only to find the right order of
compilation of the scripts that compose the user development.

For this purpose we provide a tool called \MATITADEP{}
that takes in input the list of files that compose the development and
outputs their dependencies in a format suitable for the GNU \texttt{make} tool.
The user is not asked to run \MATITADEP{} by hand, but
simply to tell \MATITA{} the root directory of his development (where all
script files can be found) and \MATITA{} will handle all the compilation
related tasks, including dependencies calculation.
To compute dependencies it is enough to look at the script files for
inclusions of other parts of the development or for explicit
references to other objects (i.e. with explicit aliases, see
\ref{sec:disambaliases}). 

The output of the compilation is immediately available to the user
trough the \WHELP{} technology, since all metadata are stored in a
user-specific area of the database where the search engine has read
access, and all the automated tactics that operates on the whole
library, like \AUTO, have full visibility of the newly defined objects.

Compilation is rather simple, and the only tricky case is when we want
to compile again the same script, maybe after the removal of a
theorem. Here the policy is simple: clean the output before recompiling.
As we will see in the next section cleaning will ensure that
there will be no theorems in the development that depends on the
removed items.

\subsubsection{Cleaning}

With the term ``cleaning'' we mean the process of removing all the
results of an object compilation. In order to keep the consistency of
the library, cleaning an object requires the (recursive) cleaning
of all the objects that depend on it (\emph{reverse dependencies}).

The calculation of the reverse dependencies can be computed in two
ways, using the relational database or using a simpler set of metadata
that \MATITA{} saves in the filesystem as a result of compilation. The
former technique is the same used by the \emph{Dependency Analyzer}
described in~\cite{zack-master} and really depends on a relational
database.

The latter is a fall-back in case the database is not
available.\footnote{Due to the complex deployment of a large piece of
software like a database, it is a common practice for the \HELM{} team
to use a shared remote database, that may be unavailable if the user
workstation lacks network connectivity.} This facility has to be
intended only as a fall-back, since the queries of the \WHELP{}
technology depend require a working database.

Cleaning guarantees that if an object is removed there are no dandling
references to it, and that the part of the library still compiled is
consistent. Since cleaning involves the removal of all the results of
the compilation, metadata included, the library browsable trough the
\WHELP{} technology is always kept up to date.

\subsubsection{Batch vs Interactive}

\MATITA{} includes an interactive authoring interface and a batch
``compiler'' (\MATITAC). Only the former is intended to be used directly by the
user, the latter is automatically invoked when a
part of the user development is required (for example issuing an
\texttt{include} command) but not yet compiled.

While they share the same engine for compilation and cleaning, they
provide different granularity. The batch compiler is only able to
compile a whole script and similarly to clean only a whole script
(together with all the other scripts that rely on an object defined in
it). The interactive interface is able to execute single steps of
compilation, that may include the definition of an object, and
similarly to undo single steps. Note that in the latter case there is
no risk of introducing dangling references since the \MATITA{} user
interface inhibit undoing a step which is not the last executed.

\subsection{Automation}
\label{sec:automation}

\TODO{sezione sull'automazione}

\subsection{Naming convention}
\label{sec:naming}

A minor but not entirely negligible aspect of \MATITA{} is that of
adopting a (semi)-rigid naming convention for identifiers, derived by 
our studies about metadata for statements. 
The convention is only applied to identifiers for theorems 
(not definitions), and relates the name of a proof to its statement.
The basic rules are the following:
\begin{itemize}
\item each identifier is composed by an ordered list of (short)
names occurring in a left to right traversal of the statement; 
\item all identifiers should (but this is not strictly compulsory) 
separated by an underscore,
\item identifiers in two different hypothesis, or in an hypothesis
and in the conclusion must be separated by the string ``\verb+_to_+'';
\item the identifier may be followed by a numerical suffix, or a
single or double apostrophe.

\end{itemize}
Take for instance the theorem
\[\forall n:nat. n = plus \; n\; O\]
Possible legal names are: \verb+plus_n_O+, \verb+plus_O+, 
\verb+eq_n_plus_n_O+ and so on. 
Similarly, consider the theorem 
\[\forall n,m:nat. n<m \to n \leq m\]
In this case \verb+lt_to_le+ is a legal name, 
while \verb+lt_le+ is not.\\
But what about, say, the symmetric law of equality? Probably you would like 
to name such a theorem with something explicitly recalling symmetry.
The correct approach, 
in this case, is the following. You should start with defining the 
symmetric property for relations

\[definition\;symmetric\;= \lambda A:Type.\lambda R.\forall x,y:A.R x y \to R y x \]

Then, you may state the symmetry of equality as
\[ \forall A:Type. symmetric \;A\;(eq \; A)\]
and \verb+symmetric_eq+ is valid \MATITA{} name for such a theorem. 
So, somehow unexpectedly, the introduction of semi-rigid naming convention
has an important beneficial effect on the global organization of the library, 
forcing the user to define abstract notions and properties before 
using them (and formalizing such use).

Two cases have a special treatment. The first one concerns theorems whose
conclusion is a (universally quantified) predicate variable, i.e. 
theorems of the shape
$\forall P,\dots.P(t)$.
In this case you may replace the conclusion with the word
``elim'' or ``case''.
For instance the name \verb+nat_elim2+ is a legal name for the double
induction principle.

The other special case is that of statements whose conclusion is a
match expression. 
A typical example is the following
\begin{verbatim}
  \forall n,m:nat. 
      match (eqb n m) with
        [ true  \Rightarrow n = m 
        | false \Rightarrow n \neq m]
\end{verbatim}
where $eqb$ is boolean equality.
In this cases, the name can be build starting from the matched
expression and the suffix \verb+_to_Prop+. In the above example, 
\verb+eqb_to_Prop+ is accepted. 

\section{The authoring interface}
\label{sec:authoring}

The authoring interface of \MATITA{} is very similar to Proof General.  We
chose not to build the \MATITA{} UI over Proof General for two reasons. First
of all we wanted to integrate our XML-based rendering technologies, mainly
\GTKMATHVIEW{}.  At the time of writing Proof General supports only text based
rendering.\footnote{This may change with the future release of Proof General
based on Eclipse, but is not yet the case.} The second reason is that we wanted
to build the \MATITA{} UI on top of a state-of-the-art and widespread toolkit
as GTK is.

Fig.~\ref{fig:screenshot} is a screenshot of the \MATITA{} authoring interface,
featuring two windows. The background one is very like to the Proof General
interface. The main difference is that we use the \GTKMATHVIEW{} widget to
render sequents. Since \GTKMATHVIEW{} renders \MATHML{} markup we take
advantage of the whole bidimensional mathematical notation.

The foreground window, also implemented around \GTKMATHVIEW, is called
``cicBrowser''. It is used to browse the library, including the proof being
developed, and enable content based search over it. Proofs are rendered in
natural language, automatically generated from the low-level lambda-terms,
using techniques inspired by~\cite{natural,YANNTHESIS} and already described 
in~\cite{remathematization}.

Note that the syntax used in the script view is \TeX-like, however Unicode is 
fully supported so that mathematical glyphs can be input as such.

\begin{figure}[!ht]
 \begin{center}
  \includegraphics[width=0.95\textwidth]{pics/matita-screenshot}
  \caption{\MATITA{} look and feel}
  \label{fig:screenshot}
 \end{center}
\end{figure}

Since the concepts of script based proof authoring are well-known, the
remaining part of this section is dedicated to the distinguishing
features of the \MATITA{} authoring interface.

\subsection{Direct manipulation of terms}
\label{sec:directmanip}

While terms are input as \TeX-like formulae in \MATITA, they are converted to a
mixed \MATHML+\BOXML{} markup for output purposes and then rendered by
\GTKMATHVIEW. As described in~\cite{latexmathml} this mixed choice enables both
high-quality bidimensional rendering of terms (including the use of fancy
layout schemata like radicals and matrices) and the use of a
concise and widespread textual syntax.

Keeping pointers from the presentations level terms down to the
partially specified ones \MATITA{} enable direct manipulation of
rendered (sub)terms in the form of hyperlinks and semantic selection.

\emph{Hyperlinks} have anchors on the occurrences of constant and
inductive type constructors and point to the corresponding definitions
in the library. Anchors are available notwithstanding the use of
user-defined mathematical notation: as can be seen on the right of
Fig.~\ref{fig:directmanip}, where we clicked on $\not|$, symbols
encoding complex notations retain all the hyperlinks of constants or
constructors used in the notation.

\emph{Semantic selection} enables the selection of mixed
\MATHML+\BOXML{} markup, constraining the selection to markup
representing meaningful CIC (sub)terms. In the example on the left of
Fig.~\ref{fig:directmanip} is thus possible to select the subterm
$\mathrm{prime}~n$, whereas it would not be possible to select
$\to n$ since the former denotes an application while the
latter it not a subterm. Once a meaningful (sub)term has been
selected actions can be done on it like reductions or tactic
applications.

\begin{figure}[!ht]
 \begin{center}
  \includegraphics[width=0.40\textwidth]{pics/matita-screenshot-selection}
  \hspace{0.05\textwidth}
  \raisebox{0.4cm}{\includegraphics[width=0.50\textwidth]{pics/matita-screenshot-href}}
  \caption{Semantic selection and hyperlinks}
  \label{fig:directmanip}
 \end{center}
\end{figure}

\TODO{referenziarli e spostarli nella parte sulla libreria?}

\begin{figure}[!ht]
 \begin{center}
  \includegraphics[width=0.30\textwidth]{pics/cicbrowser-screenshot-browsing}
  \hspace{0.02\textwidth}
  \includegraphics[width=0.30\textwidth]{pics/cicbrowser-screenshot-query}
  \hspace{0.02\textwidth}
  \includegraphics[width=0.30\textwidth]{pics/cicbrowser-screenshot-con}
  \caption{Browsing and searching the library}
  \label{fig:cicbrowser}
 \end{center}
\end{figure}

\subsection{Patterns}
\label{sec:patterns}

In several situations working with direct manipulation of terms is 
simpler and faster than typing the corresponding textual 
commands~\cite{proof-by-pointing}.
Nonetheless we need to record actions and selections in scripts.

In \MATITA{} \emph{patterns} are textual representations of selections.
Users can select using the GUI and then ask the system to paste the
corresponding pattern in this script, but more often this process is
transparent: once an action is performed on a selection, the corresponding
textual command is computed and inserted in the script.

\subsubsection{Pattern syntax}

Patterns are composed of two parts: \NT{sequent\_path} and
\NT{wanted}; their concrete syntax is reported in Tab.~\ref{tab:pathsyn}.

\NT{sequent\_path} mocks-up a sequent, discharging unwanted subterms
with $?$ and selecting the interesting parts with the placeholder
$\%$.  \NT{wanted} is a term that lives in the context of the
placeholders.

Textual patterns produced from a graphical selection are made of the
\NT{sequent\_path} only. Such patterns can represent every selection,
but are quite verbose. The \NT{wanted} part of the syntax is meant to
help the users in writing concise and elegant patterns by hand.

\begin{table}
 \caption{Patterns concrete syntax.\strut}
 \label{tab:pathsyn}
\hrule
\[
\begin{array}{@{}rcll@{}}
  \NT{pattern} & 
    ::= & [~\verb+in+~\NT{sequent\_path}~]~[~\verb+match+~\NT{wanted}~] & \\
  \NT{sequent\_path} & 
    ::= & \{~\NT{ident}~[~\verb+:+~\NT{multipath}~]~\}~
      [~\verb+\vdash+~\NT{multipath}~] & \\
  \NT{multipath} & ::= & \NT{term\_with\_placeholders} & \\
  \NT{wanted} & ::= & \NT{term} & \\
\end{array}
\]
\hrule
\end{table}

\subsubsection{Pattern evaluation}

Patterns are evaluated in two phases. The first selects roots
(subterms) of the sequent, using the $\NT{sequent\_path}$,  while the
second searches the $\NT{wanted}$ term starting from these roots.
% Both are optional steps, and by convention the empty pattern selects
% the whole conclusion.

\begin{description}
\item[Phase 1]
  concerns only the $[~\verb+in+~\NT{sequent\_path}~]$
  part of the syntax. $\NT{ident}$ is an hypothesis name and
  selects the assumption where the following optional $\NT{multipath}$
  will operate. \verb+\vdash+ can be considered the name for the goal.
  If the whole pattern is omitted, the whole goal will be selected.
  If one or more hypotheses names are given the selection is restricted to 
  these assumptions. If a $\NT{multipath}$ is omitted the whole
  assumption is selected. Remember that the user can be mostly
  unaware of this syntax, since the system is able to write down a 
  $\NT{sequent\_path}$ starting from a visual selection.
  \NOTE{Questo ancora non va in matita}

  A $\NT{multipath}$ is a CIC term in which a special constant $\%$
  is allowed.
  The roots of discharged subterms are marked with $?$, while $\%$
  is used to select roots. The default $\NT{multipath}$, the one that
  selects the whole term, is simply $\%$.
  Valid $\NT{multipath}$ are, for example, $(?~\%~?)$ or $\%~\verb+\to+~(\%~?)$
  that respectively select the first argument of an application or
  the source of an arrow and the head of the application that is
  found in the arrow target.

  The first phase not only selects terms (roots of subterms) but
  determines also their context that will be eventually used in the
  second phase.

\item[Phase 2] 
  plays a role only if the $[~\verb+match+~\NT{wanted}~]$
  part is specified. From the first phase we have some terms, that we
  will see as subterm roots, and their context. For each of these
  contexts the $\NT{wanted}$ term is disambiguated in it and the
  corresponding root is searched for a subterm that can be unified to
  $\NT{wanted}$. The result of this search is the selection the
  pattern represents.

\end{description}

\subsubsection{Examples}
%To explain how the first phase works let us give an example. Consider
%you want to prove the uniqueness of the identity element $0$ for natural
%sum, and that you can rely on the previously demonstrated left
%injectivity of the sum, that is $inj\_plus\_l:\forall x,y,z.x+y=z+y \to x =z$.
%Typing
%\begin{grafite}
%theorem valid_name: \forall n,m. m + n = n \to m = O.
%  intros (n m H).
%\end{grafite}
%\noindent
Consider the following sequent 
\sequent{
n:nat\\
m:nat\\
H: m + n = n}{
m=O
}
\noindent
To change the right part of the equivalence of the $H$
hypothesis with $O + n$ the user selects and pastes it as the pattern
in the following statement.
\begin{grafite}
  change in H:(? ? ? %) with (O + n).
\end{grafite}
\noindent
To understand the pattern (or produce it by hand) the user should be
aware that the notation $m+n=n$ hides the term $(eq~nat~(m+n)~n)$, so
that the pattern selects only the third argument of $eq$.

The experienced user may also write by hand a concise pattern
to change at once all the occurrences of $n$ in the hypothesis $H$:
\begin{grafite}
  change in H match n with (O + n).
\end{grafite}
\noindent
In this case the $\NT{sequent\_path}$ selects the whole $H$, while
the second phase locates $n$.

The latter pattern is equivalent to the following one, that the system
can automatically generate from the selection.
\begin{grafite}
  change in H:(? ? (? ? %) %) with (O + n).
\end{grafite}
\noindent

\subsubsection{Tactics supporting patterns}

\TODO{mergiare con il successivo facendo notare che i patterns sono una
interfaccia comune per le tattiche}

In \MATITA{} all the tactics that can be restricted to subterm of the working
sequent accept the pattern syntax. In particular these tactics are: simplify,
change, fold, unfold, generalize, replace and rewrite.

\NOTE{attualmente rewrite e fold non supportano phase 2. per
supportarlo bisogna far loro trasformare il pattern phase1+phase2 
in un pattern phase1only come faccio nell'ultimo esempio. lo si fa
con una pattern\_of(select(pattern))}

\subsubsection{Comparison with \COQ{}}

\COQ{} has two different ways of restricting the application of tactics to
subterms of the sequent, both relaying on the same special syntax to identify
a term occurrence.

The first way is to use this special syntax to tell the
tactic what occurrences of a wanted term should be affected.
The second is to prepare the sequent with another tactic called
pattern and then apply the real tactic. Note that the choice is not
left to the user, since some tactics needs the sequent to be prepared
with pattern and do not accept directly this special syntax.

The base idea is that to identify a subterm of the sequent we can
write it and say that we want, for example, the third and the fifth
occurrences of it (counting from left to right). In our previous example,
to change only the left part of the equivalence, the correct command
is:

\begin{grafite}
  change n at 2 in H with (O + n)
\end{grafite} 

meaning that in the hypothesis $H$ the $n$ we want to change is the
second we encounter proceeding from left to right.

The tactic pattern computes a
$\beta$-expansion of a part of the sequent with respect to some
occurrences of the given term. In the previous example the following
command:
\begin{grafite}
  pattern n at 2 in H
\end{grafite}

would have resulted in this sequent:

\begin{grafite}
  n : nat
  m : nat
  H : (fun n0 : nat => m + n = n0) n
  ============================
   m = 0
\end{grafite}

where $H$ is $\beta$-expanded over the second $n$
occurrence. 

At this point, since Coq unification algorithm is essentially
first-order, the application of an elimination principle (of the
form $\forall P.\forall x.(H~x)\to (P~x)$) will unify 
$x$ with \texttt{n} and $P$ with \texttt{(fun n0 : nat => m + n = n0)}.

Since rewriting, replacing and several other tactics boils down to
the application of the equality elimination principle, the previous
trick deals the expected behaviour.

The idea behind this way of identifying subterms in not really far
from the idea behind patterns, but fails in extending to
complex notation, since it relays on a mono-dimensional sequent representation.
Real math notation places arguments upside-down (like in indexed sums or
integrations) or even puts them inside a bidimensional matrix.  
In these cases using the mouse to select the wanted term is probably the 
more effective way to tell the system what to do. 

One of the goals of \MATITA{} is to use modern publishing techniques, and
adopting a method for restricting tactics application domain that discourages 
using heavy math notation, would definitively be a bad choice.

\subsection{Tacticals}
\label{sec:tinycals}

There are mainly two kinds of languages used by proof assistants to recorder
proofs: tactic based and declarative. We will not investigate the philosophy
around the choice that many proof assistant made, \MATITA{} included, and we
will not compare the two different approaches. We will describe the common
issues of the tactic-based language approach and how \MATITA{} tries to solve
them.

\subsubsection{Tacticals overview}

Tacticals first appeared in LCF and can be seen as programming
constructs, like looping, branching, error recovery or sequential composition.
The following simple example shows three tacticals in action
\begin{grafite}
theorem trivial: 
  \forall A,B:Prop. 
    A = B \to ((A \to B) \land (B \to A)).
  intros (A B H).
  split; intro; 
    [ rewrite < H. assumption.
    | rewrite > H. assumption.
    ]
qed.
\end{grafite}

The first is ``\texttt{;}'' that combines the tactic \texttt{split}
with \texttt{intro}, applying the latter to each goal opened by the
former. Then we have ``\texttt{[}'' that branches on the goals (here
we have two goals, the two sides of the logic and).
The first goal $B$ (with $A$ in the context)
is proved by the first sequence of tactics
\texttt{rewrite} and \texttt{assumption}. Then we move to the second
goal with the separator ``\texttt{|}''. The last tactical we see here
is ``\texttt{.}'' that is a sequential composition that selects the
first goal opened for the following tactic (instead of applying it to
them all like ``\texttt{;}''). Note that usually ``\texttt{.}'' is
not considered a tactical, but a sentence terminator (i.e. the
delimiter of commands the proof assistant executes).

Giving serious examples here is rather difficult, since they are hard
to read without the interactive tool. To help the reader in
understanding the following considerations we just give few common
usage examples without a proof context.

\begin{grafite}
  elim z; try assumption; [ ... | ... ].
  elim z; first [ assumption | reflexivity | id ].
\end{grafite}

The first example goes by induction on a term \texttt{z} and applies
the tactic \texttt{assumption} to each opened goal eventually recovering if
\texttt{assumption} fails. Here we are asking the system to close all
trivial cases and then we branch on the remaining with ``\texttt{[}''.
The second example goes again by induction on \texttt{z} and tries to
close each opened goal first with \texttt{assumption}, if it fails it
tries \texttt{reflexivity} and finally \texttt{id}
that is the tactic that leaves the goal untouched without failing. 

Note that in the common implementation of tacticals both lines are
compositions of tacticals and in particular they are a single
statement (i.e. derived from the same non terminal entry of the
grammar) ended with ``\texttt{.}''. As we will see later in \MATITA{}
this is not true, since each atomic tactic or punctuation is considered 
a single statement.

\subsubsection{Common issues of tactic(als)-based proof languages}
We will examine the two main problems of tactic(als)-based proof script:
maintainability and readability. 

Huge libraries of formal mathematics have been developed, and backward
compatibility is a really time consuming task. \\
A real-life example in the history of \MATITA{} was the reordering of
goals opened by a tactic application. We noticed that some tactics
were not opening goals in the expected order. In particular the
\texttt{elim} tactic on a term of an inductive type with constructors
$c_1, \ldots, c_n$ used to open goals in order $g_1, g_n, g_{n-1}
\ldots, g_2$. The library of \MATITA{} was still in an embryonic state
but some theorems about integers were there. The inductive type of
$\mathcal{Z}$ has three constructors: $zero$, $pos$ and $neg$. All the
induction proofs on this type where written without tacticals and,
obviously, considering the three induction cases in the wrong order.
Fixing the behavior of the tactic broke the library and two days of
work were needed to make it compile again. The whole time was spent in
finding the list of tactics used to prove the third induction case and
swap it with the list of tactics used to prove the second case.  If
the proofs was structured with the branch tactical this task could
have been done automatically. 

From this experience we learned that the use of tacticals for
structuring proofs gives some help but may have some drawbacks in
proof script readability. We must highlight that proof scripts
readability is poor by itself, but in conjunction with tacticals it
can be nearly impossible. The main cause is the fact that in proof
scripts there is no trace of what you are working on. It is not rare
for two different theorems to have the same proof script (while the
proof is completely different).\\
Bad readability is not a big deal for the user while he is
constructing the proof, but is considerably a problem when he tries to
reread what he did or when he shows his work to someone else.  The
workaround commonly used to read a script is to execute it again
step-by-step, so that you can see the proof goal changing and you can
follow the proof steps. This works fine until you reach a tactical.  A
compound statement, made by some basic tactics glued with tacticals,
is executed in a single step, while it obviously performs lot of proof
steps.  In the fist example of the previous section the whole branch
over the two goals (respectively the left and right part of the logic
and) result in a single step of execution. The workaround does not work
anymore unless you de-structure on the fly the proof, putting some
``\texttt{.}'' where you want the system to stop.\\

Now we can understand the tradeoff between script readability and
proof structuring with tacticals. Using tacticals helps in maintaining
scripts, but makes it really hard to read them again, cause of the way
they are executed.

\MATITA{} uses a language of tactics and tacticals, but tries to avoid
this tradeoff, alluring the user to write structured proof without
making it impossible to read them again.

\subsubsection{The \MATITA{} approach: Tinycals}

\begin{table}
 \caption{Concrete syntax of \MATITA{} tacticals.\strut}
 \label{tab:tacsyn}
\hrule
\[
\begin{array}{@{}rcll@{}}
  \NT{punctuation} & 
    ::= & \SEMICOLON \quad|\quad \DOT \quad|\quad \SHIFT \quad|\quad \BRANCH \quad|\quad \MERGE \quad|\quad \POS{\mathrm{NUMBER}~} & \\
  \NT{block\_kind} & 
    ::= & \verb+focus+ ~|~ \verb+try+ ~|~ \verb+solve+ ~|~ \verb+first+ ~|~ \verb+repeat+ ~|~ \verb+do+~\mathrm{NUMBER} & \\
  \NT{block\_delimiter} & 
    ::= & \verb+begin+ ~|~ \verb+end+ & \\
  \NT{tactical} & 
    ::= & \verb+skip+ ~|~ \NT{tactic} ~|~ \NT{block\_delimiter} ~|~ \NT{block\_kind} ~|~ \NT{punctuation} ~|~& \\
\end{array}
\]
\hrule
\end{table}

\MATITA{} tacticals syntax is reported in Tab.~\ref{tab:tacsyn}.
While one would expect to find structured constructs like 
$\verb+do+~n~\NT{tactic}$ the syntax allows pieces of tacticals to be written.
This is essential for base idea behind \MATITA{} tacticals: step-by-step
execution.

The low-level tacticals implementation of \MATITA{} allows a step-by-step
execution of a tactical, that substantially means that a $\NT{block\_kind}$ is
not executed as an atomic operation. This has two major benefits for the user,
even being a so simple idea:
\begin{description}
\item[Proof structuring] 
  is much easier. Consider for example a proof by induction, and imagine you
  are using classical tacticals in one of the state of the
  art graphical interfaces for proof assistant like Proof General or \COQIDE.
  After applying the induction principle you have to choose: structure
  the proof or not. If you decide for the former you have to branch with
  ``\texttt{[}'' and write tactics for all the cases separated by 
  ``\texttt{|}'' and then close the tactical with ``\texttt{]}''. 
  You can replace most of the cases by the identity tactic just to
  concentrate only on the first goal, but you will have to go one step back and
  one further every time you add something inside the tactical. Again this is
  caused by the one step execution of tacticals and by the fact that to modify
  the already executed script you have to undo one step.
  And if you are board of doing so, you will finish in giving up structuring
  the proof and write a plain list of tactics.\\
  With step-by-step tacticals you can apply the induction principle, and just
  open the branching tactical ``\texttt{[}''. Then you can interact with the
  system reaching a proof of the first case, without having to specify any
  tactic for the other goals. When you have proved all the induction cases, you
  close the branching tactical with ``\texttt{]}'' and you are done with a 
  structured proof. \\
  While \MATITA{} tacticals help in structuring proofs they allow you to 
  choose the amount of structure you want. There are no constraints imposed by
  the system, and if the user wants he can even write completely plain proofs.
  
\item[Rereading]
  is possible. Going on step by step shows exactly what is going on.  Consider
  again a proof by induction, that starts applying the induction principle and
  suddenly branches with a ``\texttt{[}''. This clearly separates all the
  induction cases, but if the square brackets content is executed in one single
  step you completely loose the possibility of rereading it and you have to
  temporary remove the branching tactical to execute in a satisfying way the
  branches.  Again, executing step-by-step is the way you would like to review
  the demonstration. Remember that understanding the proof from the script is
  not easy, and only the execution of tactics (and the resulting transformed
  goal) gives you the feeling of what is going on.
\end{description}

\section{Standard library}
\label{sec:stdlib}

\MATITA{} is \COQ{} compatible, in the sense that every theorem of \COQ{}
can be read, checked and referenced in further developments. 
However, in order to test the actual usability of the system, a
new library of results has been started from scratch. In this case, 
of course, we wrote (and offer) the source script files, 
while, in the case of \COQ, \MATITA{} may only rely on XML files of
\COQ{} objects. 
The current library just comprises about one thousand theorems in 
elementary aspects of arithmetics up to the multiplicative property for 
Eulers' totient function $\phi$.
The library is organized in five main directories: \texttt{logic} (connectives,
quantifiers, equality, \ldots), \texttt{datatypes} (basic datatypes and type 
constructors), \texttt{nat} (natural numbers), \texttt{Z} (integers), \texttt{Q}
(rationals). The most complex development is \texttt{nat}, organized in 25
scripts, listed in Tab.~\ref{tab:scripts}.

\begin{table}[ht]
 \begin{tabular}{lll}
  \FILE{nat.ma}    & \FILE{plus.ma} & \FILE{times.ma}  \\
  \FILE{minus.ma}  & \FILE{exp.ma}  & \FILE{compare.ma} \\
  \FILE{orders.ma} & \FILE{le\_arith.ma} &  \FILE{lt\_arith.ma} \\   
  \FILE{factorial.ma} & \FILE{sigma\_and\_pi.ma} & \FILE{minimization.ma}  \\
  \FILE{div\_and\_mod.ma} & \FILE{gcd.ma} & \FILE{congruence.ma} \\
  \FILE{primes.ma} & \FILE{nth\_prime.ma} & \FILE{ord.ma} \\
  \FILE{count.ma}  & \FILE{relevant\_equations.ma} & \FILE{permutation.ma} \\ 
  \FILE{factorization.ma} & \FILE{chinese\_reminder.ma} &
  \FILE{fermat\_little\_th.ma} \\     
  \FILE{totient.ma} & & \\
 \end{tabular}
 \caption{Scripts on natural numbers in the standard library}
 \label{tab:scripts}
\end{table}

We do not plan to maintain the library in a centralized way, 
as most of the systems do. On the contrary we are currently
developing wiki-technologies to support a collaborative 
development of the library, encouraging people to expand, 
modify and elaborate previous contributions.

\section{Conclusions}
\label{sec:conclusion}

\TODO{conclusioni}

\acknowledgements
We would like to thank all the students that during the past
five years collaborated in the \HELM{} project and contributed to 
the development of \MATITA{}, and in particular
M.~Galat\`a, A.~Griggio, F.~Guidi, P.~Di~Lena, L.~Padovani, I.~Schena, M.~Selmi,
and V.~Tamburrelli.

\theendnotes

\bibliography{matita}

\end{document}