doc-src/Logics/LK.tex
author wenzelm
Mon, 28 Aug 2000 13:52:38 +0200
changeset 9695 ec7d7f877712
parent 7116 8c1caac3e54e
child 19152 d81fae81f385
permissions -rw-r--r--
proper setup of iman.sty/extra.sty/ttbox.sty;

%% $Id$
\chapter{First-Order Sequent Calculus}
\index{sequent calculus|(}

The theory~\thydx{LK} implements classical first-order logic through Gentzen's
sequent calculus (see Gallier~\cite{gallier86} or Takeuti~\cite{takeuti87}).
Resembling the method of semantic tableaux, the calculus is well suited for
backwards proof.  Assertions have the form \(\Gamma\turn \Delta\), where
\(\Gamma\) and \(\Delta\) are lists of formulae.  Associative unification,
simulated by higher-order unification, handles lists
(\S\ref{sec:assoc-unification} presents details, if you are interested).

The logic is many-sorted, using Isabelle's type classes.  The class of
first-order terms is called \cldx{term}.  No types of individuals are
provided, but extensions can define types such as {\tt nat::term} and type
constructors such as {\tt list::(term)term}.  Below, the type variable
$\alpha$ ranges over class {\tt term}; the equality symbol and quantifiers
are polymorphic (many-sorted).  The type of formulae is~\tydx{o}, which
belongs to class {\tt logic}.

LK implements a classical logic theorem prover that is nearly as powerful as
the generic classical reasoner.  The simplifier is now available too.

To work in LK, start up Isabelle specifying  \texttt{Sequents} as the
object-logic.  Once in Isabelle, change the context to theory \texttt{LK.thy}:
\begin{ttbox}
isabelle Sequents
context LK.thy;
\end{ttbox}
Model logic and linear logic are also available, but unfortunately they are
not documented.


\begin{figure} 
\begin{center}
\begin{tabular}{rrr} 
  \it name      &\it meta-type          & \it description       \\ 
  \cdx{Trueprop}& $[sobj\To sobj, sobj\To sobj]\To prop$ & coercion to $prop$\\
  \cdx{Seqof}   & $[o,sobj]\To sobj$    & singleton sequence    \\
  \cdx{Not}     & $o\To o$              & negation ($\neg$)     \\
  \cdx{True}    & $o$                   & tautology ($\top$)    \\
  \cdx{False}   & $o$                   & absurdity ($\bot$)
\end{tabular}
\end{center}
\subcaption{Constants}

\begin{center}
\begin{tabular}{llrrr} 
  \it symbol &\it name     &\it meta-type & \it priority & \it description \\
  \sdx{ALL}  & \cdx{All}  & $(\alpha\To o)\To o$ & 10 & 
        universal quantifier ($\forall$) \\
  \sdx{EX}   & \cdx{Ex}   & $(\alpha\To o)\To o$ & 10 & 
        existential quantifier ($\exists$) \\
  \sdx{THE} & \cdx{The}  & $(\alpha\To o)\To \alpha$ & 10 & 
        definite description ($\iota$)
\end{tabular}
\end{center}
\subcaption{Binders} 

\begin{center}
\index{*"= symbol}
\index{&@{\tt\&} symbol}
\index{*"| symbol}
\index{*"-"-"> symbol}
\index{*"<"-"> symbol}
\begin{tabular}{rrrr} 
    \it symbol  & \it meta-type         & \it priority & \it description \\ 
    \tt = &     $[\alpha,\alpha]\To o$  & Left 50 & equality ($=$) \\
    \tt \& &    $[o,o]\To o$ & Right 35 & conjunction ($\conj$) \\
    \tt | &     $[o,o]\To o$ & Right 30 & disjunction ($\disj$) \\
    \tt --> &   $[o,o]\To o$ & Right 25 & implication ($\imp$) \\
    \tt <-> &   $[o,o]\To o$ & Right 25 & biconditional ($\bimp$) 
\end{tabular}
\end{center}
\subcaption{Infixes}

\begin{center}
\begin{tabular}{rrr} 
  \it external          & \it internal  & \it description \\ 
  \tt $\Gamma$ |- $\Delta$  &  \tt Trueprop($\Gamma$, $\Delta$) &
        sequent $\Gamma\turn \Delta$ 
\end{tabular}
\end{center}
\subcaption{Translations} 
\caption{Syntax of {\tt LK}} \label{lk-syntax}
\end{figure}


\begin{figure} 
\dquotes
\[\begin{array}{rcl}
    prop & = & sequence " |- " sequence 
\\[2ex]
sequence & = & elem \quad (", " elem)^* \\
         & | & empty 
\\[2ex]
    elem & = & "\$ " term \\
         & | & formula  \\
         & | & "<<" sequence ">>" 
\\[2ex]
 formula & = & \hbox{expression of type~$o$} \\
         & | & term " = " term \\
         & | & "\ttilde\ " formula \\
         & | & formula " \& " formula \\
         & | & formula " | " formula \\
         & | & formula " --> " formula \\
         & | & formula " <-> " formula \\
         & | & "ALL~" id~id^* " . " formula \\
         & | & "EX~~" id~id^* " . " formula \\
         & | & "THE~" id~     " . " formula
  \end{array}
\]
\caption{Grammar of {\tt LK}} \label{lk-grammar}
\end{figure}




\begin{figure} 
\begin{ttbox}
\tdx{basic}       $H, P, $G |- $E, P, $F

\tdx{contRS}      $H |- $E, $S, $S, $F ==> $H |- $E, $S, $F
\tdx{contLS}      $H, $S, $S, $G |- $E ==> $H, $S, $G |- $E

\tdx{thinRS}      $H |- $E, $F ==> $H |- $E, $S, $F
\tdx{thinLS}      $H, $G |- $E ==> $H, $S, $G |- $E

\tdx{cut}         [| $H |- $E, P;  $H, P |- $E |] ==> $H |- $E
\subcaption{Structural rules}

\tdx{refl}        $H |- $E, a=a, $F
\tdx{subst}       $H(a), $G(a) |- $E(a) ==> $H(b), a=b, $G(b) |- $E(b)
\subcaption{Equality rules}
\end{ttbox}

\caption{Basic Rules of {\tt LK}}  \label{lk-basic-rules}
\end{figure}

\begin{figure} 
\begin{ttbox}
\tdx{True_def}    True  == False-->False
\tdx{iff_def}     P<->Q == (P-->Q) & (Q-->P)

\tdx{conjR}   [| $H|- $E, P, $F;  $H|- $E, Q, $F |] ==> $H|- $E, P&Q, $F
\tdx{conjL}   $H, P, Q, $G |- $E ==> $H, P & Q, $G |- $E

\tdx{disjR}   $H |- $E, P, Q, $F ==> $H |- $E, P|Q, $F
\tdx{disjL}   [| $H, P, $G |- $E;  $H, Q, $G |- $E |] ==> $H, P|Q, $G |- $E
            
\tdx{impR}    $H, P |- $E, Q, $F ==> $H |- $E, P-->Q, $F
\tdx{impL}    [| $H,$G |- $E,P;  $H, Q, $G |- $E |] ==> $H, P-->Q, $G |- $E
            
\tdx{notR}    $H, P |- $E, $F ==> $H |- $E, ~P, $F
\tdx{notL}    $H, $G |- $E, P ==> $H, ~P, $G |- $E

\tdx{FalseL}  $H, False, $G |- $E

\tdx{allR}    (!!x. $H|- $E, P(x), $F) ==> $H|- $E, ALL x. P(x), $F
\tdx{allL}    $H, P(x), $G, ALL x. P(x) |- $E ==> $H, ALL x. P(x), $G|- $E
            
\tdx{exR}     $H|- $E, P(x), $F, EX x. P(x) ==> $H|- $E, EX x. P(x), $F
\tdx{exL}     (!!x. $H, P(x), $G|- $E) ==> $H, EX x. P(x), $G|- $E

\tdx{The}     [| $H |- $E, P(a), $F;  !!x. $H, P(x) |- $E, x=a, $F |] ==>
        $H |- $E, P(THE x. P(x)), $F
\subcaption{Logical rules}
\end{ttbox}

\caption{Rules of {\tt LK}}  \label{lk-rules}
\end{figure}


\section{Syntax and rules of inference}
\index{*sobj type}

Figure~\ref{lk-syntax} gives the syntax for {\tt LK}, which is complicated
by the representation of sequents.  Type $sobj\To sobj$ represents a list
of formulae.

The \textbf{definite description} operator~$\iota x. P[x]$ stands for some~$a$
satisfying~$P[a]$, if one exists and is unique.  Since all terms in LK denote
something, a description is always meaningful, but we do not know its value
unless $P[x]$ defines it uniquely.  The Isabelle notation is \hbox{\tt THE
  $x$.\ $P[x]$}.  The corresponding rule (Fig.\ts\ref{lk-rules}) does not
entail the Axiom of Choice because it requires uniqueness.

Conditional expressions are available with the notation 
\[ \dquotes
   "if"~formula~"then"~term~"else"~term. \]

Figure~\ref{lk-grammar} presents the grammar of LK.  Traditionally,
\(\Gamma\) and \(\Delta\) are meta-variables for sequences.  In Isabelle's
notation, the prefix~\verb|$| on a term makes it range over sequences.
In a sequent, anything not prefixed by \verb|$| is taken as a formula.

The notation \texttt{<<$sequence$>>} stands for a sequence of formul\ae{}.
For example, you can declare the constant \texttt{imps} to consist of two
implications: 
\begin{ttbox}
consts     P,Q,R :: o
constdefs imps :: seq'=>seq'
         "imps == <<P --> Q, Q --> R>>"
\end{ttbox}
Then you can use it in axioms and goals, for example
\begin{ttbox}
Goalw [imps_def] "P, $imps |- R";  
{\out Level 0}
{\out P, $imps |- R}
{\out  1. P, P --> Q, Q --> R |- R}
by (Fast_tac 1);
{\out Level 1}
{\out P, $imps |- R}
{\out No subgoals!}
\end{ttbox}

Figures~\ref{lk-basic-rules} and~\ref{lk-rules} present the rules of theory
\thydx{LK}.  The connective $\bimp$ is defined using $\conj$ and $\imp$.  The
axiom for basic sequents is expressed in a form that provides automatic
thinning: redundant formulae are simply ignored.  The other rules are
expressed in the form most suitable for backward proof; exchange and
contraction rules are not normally required, although they are provided
anyway. 


\begin{figure} 
\begin{ttbox}
\tdx{thinR}        $H |- $E, $F ==> $H |- $E, P, $F
\tdx{thinL}        $H, $G |- $E ==> $H, P, $G |- $E

\tdx{contR}        $H |- $E, P, P, $F ==> $H |- $E, P, $F
\tdx{contL}        $H, P, P, $G |- $E ==> $H, P, $G |- $E

\tdx{symR}         $H |- $E, $F, a=b ==> $H |- $E, b=a, $F
\tdx{symL}         $H, $G, b=a |- $E ==> $H, a=b, $G |- $E

\tdx{transR}       [| $H|- $E, $F, a=b;  $H|- $E, $F, b=c |] 
             ==> $H|- $E, a=c, $F

\tdx{TrueR}        $H |- $E, True, $F

\tdx{iffR}         [| $H, P |- $E, Q, $F;  $H, Q |- $E, P, $F |]
             ==> $H |- $E, P<->Q, $F

\tdx{iffL}         [| $H, $G |- $E, P, Q;  $H, Q, P, $G |- $E |]
             ==> $H, P<->Q, $G |- $E

\tdx{allL_thin}    $H, P(x), $G |- $E ==> $H, ALL x. P(x), $G |- $E
\tdx{exR_thin}     $H |- $E, P(x), $F ==> $H |- $E, EX x. P(x), $F

\tdx{the_equality} [| $H |- $E, P(a), $F;  
                !!x. $H, P(x) |- $E, x=a, $F |] 
             ==> $H |- $E, (THE x. P(x)) = a, $F
\end{ttbox}

\caption{Derived rules for {\tt LK}} \label{lk-derived}
\end{figure}

Figure~\ref{lk-derived} presents derived rules, including rules for
$\bimp$.  The weakened quantifier rules discard each quantification after a
single use; in an automatic proof procedure, they guarantee termination,
but are incomplete.  Multiple use of a quantifier can be obtained by a
contraction rule, which in backward proof duplicates a formula.  The tactic
{\tt res_inst_tac} can instantiate the variable~{\tt?P} in these rules,
specifying the formula to duplicate.
See theory {\tt Sequents/LK0} in the sources for complete listings of
the rules and derived rules.

To support the simplifier, hundreds of equivalences are proved for
the logical connectives and for if-then-else expressions.  See the file
\texttt{Sequents/simpdata.ML}.

\section{Automatic Proof}

LK instantiates Isabelle's simplifier.  Both equality ($=$) and the
biconditional ($\bimp$) may be used for rewriting.  The tactic
\texttt{Simp_tac} refers to the default simpset (\texttt{simpset()}).  With
sequents, the \texttt{full_} and \texttt{asm_} forms of the simplifier are not
required; all the formulae{} in the sequent will be simplified.  The left-hand
formulae{} are taken as rewrite rules.  (Thus, the behaviour is what you would
normally expect from calling \texttt{Asm_full_simp_tac}.)

For classical reasoning, several tactics are available:
\begin{ttbox} 
Safe_tac : int -> tactic
Step_tac : int -> tactic
Fast_tac : int -> tactic
Best_tac : int -> tactic
Pc_tac   : int -> tactic
\end{ttbox}
These refer not to the standard classical reasoner but to a separate one
provided for the sequent calculus.  Two commands are available for adding new
sequent calculus rules, safe or unsafe, to the default ``theorem pack'':
\begin{ttbox} 
Add_safes   : thm list -> unit
Add_unsafes : thm list -> unit
\end{ttbox}
To control the set of rules for individual invocations, lower-case versions of
all these primitives are available.  Sections~\ref{sec:thm-pack}
and~\ref{sec:sequent-provers} give full details.


\section{Tactics for the cut rule}

According to the cut-elimination theorem, the cut rule can be eliminated
from proofs of sequents.  But the rule is still essential.  It can be used
to structure a proof into lemmas, avoiding repeated proofs of the same
formula.  More importantly, the cut rule can not be eliminated from
derivations of rules.  For example, there is a trivial cut-free proof of
the sequent \(P\conj Q\turn Q\conj P\).
Noting this, we might want to derive a rule for swapping the conjuncts
in a right-hand formula:
\[ \Gamma\turn \Delta, P\conj Q\over \Gamma\turn \Delta, Q\conj P \]
The cut rule must be used, for $P\conj Q$ is not a subformula of $Q\conj
P$.  Most cuts directly involve a premise of the rule being derived (a
meta-assumption).  In a few cases, the cut formula is not part of any
premise, but serves as a bridge between the premises and the conclusion.
In such proofs, the cut formula is specified by calling an appropriate
tactic.

\begin{ttbox} 
cutR_tac : string -> int -> tactic
cutL_tac : string -> int -> tactic
\end{ttbox}
These tactics refine a subgoal into two by applying the cut rule.  The cut
formula is given as a string, and replaces some other formula in the sequent.
\begin{ttdescription}
\item[\ttindexbold{cutR_tac} {\it P\/} {\it i}] reads an LK formula~$P$, and
  applies the cut rule to subgoal~$i$.  It then deletes some formula from the
  right side of subgoal~$i$, replacing that formula by~$P$.
  
\item[\ttindexbold{cutL_tac} {\it P\/} {\it i}] reads an LK formula~$P$, and
  applies the cut rule to subgoal~$i$.  It then deletes some formula from the
  left side of the new subgoal $i+1$, replacing that formula by~$P$.
\end{ttdescription}
All the structural rules --- cut, contraction, and thinning --- can be
applied to particular formulae using {\tt res_inst_tac}.


\section{Tactics for sequents}
\begin{ttbox} 
forms_of_seq       : term -> term list
could_res          : term * term -> bool
could_resolve_seq  : term * term -> bool
filseq_resolve_tac : thm list -> int -> int -> tactic
\end{ttbox}
Associative unification is not as efficient as it might be, in part because
the representation of lists defeats some of Isabelle's internal
optimisations.  The following operations implement faster rule application,
and may have other uses.
\begin{ttdescription}
\item[\ttindexbold{forms_of_seq} {\it t}] 
returns the list of all formulae in the sequent~$t$, removing sequence
variables.

\item[\ttindexbold{could_res} ($t$,$u$)] 
tests whether two formula lists could be resolved.  List $t$ is from a
premise or subgoal, while $u$ is from the conclusion of an object-rule.
Assuming that each formula in $u$ is surrounded by sequence variables, it
checks that each conclusion formula is unifiable (using {\tt could_unify})
with some subgoal formula.

\item[\ttindexbold{could_resolve_seq} ($t$,$u$)] 
  tests whether two sequents could be resolved.  Sequent $t$ is a premise
  or subgoal, while $u$ is the conclusion of an object-rule.  It simply
  calls {\tt could_res} twice to check that both the left and the right
  sides of the sequents are compatible.

\item[\ttindexbold{filseq_resolve_tac} {\it thms} {\it maxr} {\it i}] 
uses {\tt filter_thms could_resolve} to extract the {\it thms} that are
applicable to subgoal~$i$.  If more than {\it maxr\/} theorems are
applicable then the tactic fails.  Otherwise it calls {\tt resolve_tac}.
Thus, it is the sequent calculus analogue of \ttindex{filt_resolve_tac}.
\end{ttdescription}


\section{A simple example of classical reasoning} 
The theorem $\turn\ex{y}\all{x}P(y)\imp P(x)$ is a standard example of the
classical treatment of the existential quantifier.  Classical reasoning is
easy using~LK, as you can see by comparing this proof with the one given in
the FOL manual~\cite{isabelle-ZF}.  From a logical point of view, the proofs
are essentially the same; the key step here is to use \tdx{exR} rather than
the weaker~\tdx{exR_thin}.
\begin{ttbox}
Goal "|- EX y. ALL x. P(y)-->P(x)";
{\out Level 0}
{\out  |- EX y. ALL x. P(y) --> P(x)}
{\out  1.  |- EX y. ALL x. P(y) --> P(x)}
by (resolve_tac [exR] 1);
{\out Level 1}
{\out  |- EX y. ALL x. P(y) --> P(x)}
{\out  1.  |- ALL x. P(?x) --> P(x), EX x. ALL xa. P(x) --> P(xa)}
\end{ttbox}
There are now two formulae on the right side.  Keeping the existential one
in reserve, we break down the universal one.
\begin{ttbox}
by (resolve_tac [allR] 1);
{\out Level 2}
{\out  |- EX y. ALL x. P(y) --> P(x)}
{\out  1. !!x.  |- P(?x) --> P(x), EX x. ALL xa. P(x) --> P(xa)}
by (resolve_tac [impR] 1);
{\out Level 3}
{\out  |- EX y. ALL x. P(y) --> P(x)}
{\out  1. !!x. P(?x) |- P(x), EX x. ALL xa. P(x) --> P(xa)}
\end{ttbox}
Because LK is a sequent calculus, the formula~$P(\Var{x})$ does not become an
assumption; instead, it moves to the left side.  The resulting subgoal cannot
be instantiated to a basic sequent: the bound variable~$x$ is not unifiable
with the unknown~$\Var{x}$.
\begin{ttbox}
by (resolve_tac [basic] 1);
{\out by: tactic failed}
\end{ttbox}
We reuse the existential formula using~\tdx{exR_thin}, which discards
it; we shall not need it a third time.  We again break down the resulting
formula.
\begin{ttbox}
by (resolve_tac [exR_thin] 1);
{\out Level 4}
{\out  |- EX y. ALL x. P(y) --> P(x)}
{\out  1. !!x. P(?x) |- P(x), ALL xa. P(?x7(x)) --> P(xa)}
by (resolve_tac [allR] 1);
{\out Level 5}
{\out  |- EX y. ALL x. P(y) --> P(x)}
{\out  1. !!x xa. P(?x) |- P(x), P(?x7(x)) --> P(xa)}
by (resolve_tac [impR] 1);
{\out Level 6}
{\out  |- EX y. ALL x. P(y) --> P(x)}
{\out  1. !!x xa. P(?x), P(?x7(x)) |- P(x), P(xa)}
\end{ttbox}
Subgoal~1 seems to offer lots of possibilities.  Actually the only useful
step is instantiating~$\Var{x@7}$ to $\lambda x. x$,
transforming~$\Var{x@7}(x)$ into~$x$.
\begin{ttbox}
by (resolve_tac [basic] 1);
{\out Level 7}
{\out  |- EX y. ALL x. P(y) --> P(x)}
{\out No subgoals!}
\end{ttbox}
This theorem can be proved automatically.  Because it involves quantifier
duplication, we employ best-first search:
\begin{ttbox}
Goal "|- EX y. ALL x. P(y)-->P(x)";
{\out Level 0}
{\out  |- EX y. ALL x. P(y) --> P(x)}
{\out  1.  |- EX y. ALL x. P(y) --> P(x)}
by (best_tac LK_dup_pack 1);
{\out Level 1}
{\out  |- EX y. ALL x. P(y) --> P(x)}
{\out No subgoals!}
\end{ttbox}



\section{A more complex proof}
Many of Pelletier's test problems for theorem provers \cite{pelletier86}
can be solved automatically.  Problem~39 concerns set theory, asserting
that there is no Russell set --- a set consisting of those sets that are
not members of themselves:
\[  \turn \neg (\exists x. \forall y. y\in x \bimp y\not\in y) \]
This does not require special properties of membership; we may generalize
$x\in y$ to an arbitrary predicate~$F(x,y)$.  The theorem, which is trivial
for \texttt{Fast_tac}, has a short manual proof.  See the directory {\tt
  Sequents/LK} for many more examples.

We set the main goal and move the negated formula to the left.
\begin{ttbox}
Goal "|- ~ (EX x. ALL y. F(y,x) <-> ~F(y,y))";
{\out Level 0}
{\out  |- ~ (EX x. ALL y. F(y,x) <-> ~ F(y,y))}
{\out  1.  |- ~ (EX x. ALL y. F(y,x) <-> ~ F(y,y))}
by (resolve_tac [notR] 1);
{\out Level 1}
{\out  |- ~ (EX x. ALL y. F(y,x) <-> ~ F(y,y))}
{\out  1. EX x. ALL y. F(y,x) <-> ~ F(y,y) |-}
\end{ttbox}
The right side is empty; we strip both quantifiers from the formula on the
left.
\begin{ttbox}
by (resolve_tac [exL] 1);
{\out Level 2}
{\out  |- ~ (EX x. ALL y. F(y,x) <-> ~ F(y,y))}
{\out  1. !!x. ALL y. F(y,x) <-> ~ F(y,y) |-}
by (resolve_tac [allL_thin] 1);
{\out Level 3}
{\out  |- ~ (EX x. ALL y. F(y,x) <-> ~ F(y,y))}
{\out  1. !!x. F(?x2(x),x) <-> ~ F(?x2(x),?x2(x)) |-}
\end{ttbox}
The rule \tdx{iffL} says, if $P\bimp Q$ then $P$ and~$Q$ are either
both true or both false.  It yields two subgoals.
\begin{ttbox}
by (resolve_tac [iffL] 1);
{\out Level 4}
{\out  |- ~ (EX x. ALL y. F(y,x) <-> ~ F(y,y))}
{\out  1. !!x.  |- F(?x2(x),x), ~ F(?x2(x),?x2(x))}
{\out  2. !!x. ~ F(?x2(x),?x2(x)), F(?x2(x),x) |-}
\end{ttbox}
We must instantiate~$\Var{x@2}$, the shared unknown, to satisfy both
subgoals.  Beginning with subgoal~2, we move a negated formula to the left
and create a basic sequent.
\begin{ttbox}
by (resolve_tac [notL] 2);
{\out Level 5}
{\out  |- ~ (EX x. ALL y. F(y,x) <-> ~ F(y,y))}
{\out  1. !!x.  |- F(?x2(x),x), ~ F(?x2(x),?x2(x))}
{\out  2. !!x. F(?x2(x),x) |- F(?x2(x),?x2(x))}
by (resolve_tac [basic] 2);
{\out Level 6}
{\out  |- ~ (EX x. ALL y. F(y,x) <-> ~ F(y,y))}
{\out  1. !!x.  |- F(x,x), ~ F(x,x)}
\end{ttbox}
Thanks to the instantiation of~$\Var{x@2}$, subgoal~1 is obviously true.
\begin{ttbox}
by (resolve_tac [notR] 1);
{\out Level 7}
{\out  |- ~ (EX x. ALL y. F(y,x) <-> ~ F(y,y))}
{\out  1. !!x. F(x,x) |- F(x,x)}
by (resolve_tac [basic] 1);
{\out Level 8}
{\out  |- ~ (EX x. ALL y. F(y,x) <-> ~ F(y,y))}
{\out No subgoals!}
\end{ttbox}

\section{*Unification for lists}\label{sec:assoc-unification}

Higher-order unification includes associative unification as a special
case, by an encoding that involves function composition
\cite[page~37]{huet78}.  To represent lists, let $C$ be a new constant.
The empty list is $\lambda x. x$, while $[t@1,t@2,\ldots,t@n]$ is
represented by
\[ \lambda x. C(t@1,C(t@2,\ldots,C(t@n,x))).  \]
The unifiers of this with $\lambda x.\Var{f}(\Var{g}(x))$ give all the ways
of expressing $[t@1,t@2,\ldots,t@n]$ as the concatenation of two lists.

Unlike orthodox associative unification, this technique can represent certain
infinite sets of unifiers by flex-flex equations.   But note that the term
$\lambda x. C(t,\Var{a})$ does not represent any list.  Flex-flex constraints
containing such garbage terms may accumulate during a proof.
\index{flex-flex constraints}

This technique lets Isabelle formalize sequent calculus rules,
where the comma is the associative operator:
\[ \infer[(\conj\hbox{-left})]
         {\Gamma,P\conj Q,\Delta \turn \Theta}
         {\Gamma,P,Q,\Delta \turn \Theta}  \] 
Multiple unifiers occur whenever this is resolved against a goal containing
more than one conjunction on the left.  

LK exploits this representation of lists.  As an alternative, the sequent
calculus can be formalized using an ordinary representation of lists, with a
logic program for removing a formula from a list.  Amy Felty has applied this
technique using the language $\lambda$Prolog~\cite{felty91a}.

Explicit formalization of sequents can be tiresome.  But it gives precise
control over contraction and weakening, and is essential to handle relevant
and linear logics.


\section{*Packaging sequent rules}\label{sec:thm-pack}

The sequent calculi come with simple proof procedures.  These are incomplete
but are reasonably powerful for interactive use.  They expect rules to be
classified as \textbf{safe} or \textbf{unsafe}.  A rule is safe if applying it to a
provable goal always yields provable subgoals.  If a rule is safe then it can
be applied automatically to a goal without destroying our chances of finding a
proof.  For instance, all the standard rules of the classical sequent calculus
{\sc lk} are safe.  An unsafe rule may render the goal unprovable; typical
examples are the weakened quantifier rules {\tt allL_thin} and {\tt exR_thin}.

Proof procedures use safe rules whenever possible, using an unsafe rule as a
last resort.  Those safe rules are preferred that generate the fewest
subgoals.  Safe rules are (by definition) deterministic, while the unsafe
rules require a search strategy, such as backtracking.

A \textbf{pack} is a pair whose first component is a list of safe rules and
whose second is a list of unsafe rules.  Packs can be extended in an obvious
way to allow reasoning with various collections of rules.  For clarity, LK
declares \mltydx{pack} as an \ML{} datatype, although is essentially a type
synonym:
\begin{ttbox}
datatype pack = Pack of thm list * thm list;
\end{ttbox}
Pattern-matching using constructor {\tt Pack} can inspect a pack's
contents.  Packs support the following operations:
\begin{ttbox} 
pack        : unit -> pack
pack_of     : theory -> pack
empty_pack  : pack
prop_pack   : pack
LK_pack     : pack
LK_dup_pack : pack
add_safes   : pack * thm list -> pack               \hfill\textbf{infix 4}
add_unsafes : pack * thm list -> pack               \hfill\textbf{infix 4}
\end{ttbox}
\begin{ttdescription}
\item[\ttindexbold{pack}] returns the pack attached to the current theory.

\item[\ttindexbold{pack_of $thy$}] returns the pack attached to theory $thy$.

\item[\ttindexbold{empty_pack}] is the empty pack.

\item[\ttindexbold{prop_pack}] contains the propositional rules, namely
those for $\conj$, $\disj$, $\neg$, $\imp$ and~$\bimp$, along with the
rules {\tt basic} and {\tt refl}.  These are all safe.

\item[\ttindexbold{LK_pack}] 
extends {\tt prop_pack} with the safe rules {\tt allR}
and~{\tt exL} and the unsafe rules {\tt allL_thin} and
{\tt exR_thin}.  Search using this is incomplete since quantified
formulae are used at most once.

\item[\ttindexbold{LK_dup_pack}] 
extends {\tt prop_pack} with the safe rules {\tt allR}
and~{\tt exL} and the unsafe rules \tdx{allL} and~\tdx{exR}.
Search using this is complete, since quantified formulae may be reused, but
frequently fails to terminate.  It is generally unsuitable for depth-first
search.

\item[$pack$ \ttindexbold{add_safes} $rules$] 
adds some safe~$rules$ to the pack~$pack$.

\item[$pack$ \ttindexbold{add_unsafes} $rules$] 
adds some unsafe~$rules$ to the pack~$pack$.
\end{ttdescription}


\section{*Proof procedures}\label{sec:sequent-provers}

The LK proof procedure is similar to the classical reasoner described in
\iflabelundefined{chap:classical}{the {\em Reference Manual\/}}%
            {Chap.\ts\ref{chap:classical}}.  
%
In fact it is simpler, since it works directly with sequents rather than
simulating them.  There is no need to distinguish introduction rules from
elimination rules, and of course there is no swap rule.  As always,
Isabelle's classical proof procedures are less powerful than resolution
theorem provers.  But they are more natural and flexible, working with an
open-ended set of rules.

Backtracking over the choice of a safe rule accomplishes nothing: applying
them in any order leads to essentially the same result.  Backtracking may
be necessary over basic sequents when they perform unification.  Suppose
that~0, 1, 2,~3 are constants in the subgoals
\[  \begin{array}{c}
      P(0), P(1), P(2) \turn P(\Var{a})  \\
      P(0), P(2), P(3) \turn P(\Var{a})  \\
      P(1), P(3), P(2) \turn P(\Var{a})  
    \end{array}
\]
The only assignment that satisfies all three subgoals is $\Var{a}\mapsto 2$,
and this can only be discovered by search.  The tactics given below permit
backtracking only over axioms, such as {\tt basic} and {\tt refl};
otherwise they are deterministic.


\subsection{Method A}
\begin{ttbox} 
reresolve_tac   : thm list -> int -> tactic
repeat_goal_tac : pack -> int -> tactic
pc_tac          : pack -> int -> tactic
\end{ttbox}
These tactics use a method developed by Philippe de Groote.  A subgoal is
refined and the resulting subgoals are attempted in reverse order.  For
some reason, this is much faster than attempting the subgoals in order.
The method is inherently depth-first.

At present, these tactics only work for rules that have no more than two
premises.  They fail --- return no next state --- if they can do nothing.
\begin{ttdescription}
\item[\ttindexbold{reresolve_tac} $thms$ $i$] 
repeatedly applies the $thms$ to subgoal $i$ and the resulting subgoals.

\item[\ttindexbold{repeat_goal_tac} $pack$ $i$] 
applies the safe rules in the pack to a goal and the resulting subgoals.
If no safe rule is applicable then it applies an unsafe rule and continues.

\item[\ttindexbold{pc_tac} $pack$ $i$] 
applies {\tt repeat_goal_tac} using depth-first search to solve subgoal~$i$.
\end{ttdescription}


\subsection{Method B}
\begin{ttbox} 
safe_tac : pack -> int -> tactic
step_tac : pack -> int -> tactic
fast_tac : pack -> int -> tactic
best_tac : pack -> int -> tactic
\end{ttbox}
These tactics are analogous to those of the generic classical
reasoner.  They use `Method~A' only on safe rules.  They fail if they
can do nothing.
\begin{ttdescription}
\item[\ttindexbold{safe_goal_tac} $pack$ $i$] 
applies the safe rules in the pack to a goal and the resulting subgoals.
It ignores the unsafe rules.  

\item[\ttindexbold{step_tac} $pack$ $i$] 
either applies safe rules (using {\tt safe_goal_tac}) or applies one unsafe
rule.

\item[\ttindexbold{fast_tac} $pack$ $i$] 
applies {\tt step_tac} using depth-first search to solve subgoal~$i$.
Despite its name, it is frequently slower than {\tt pc_tac}.

\item[\ttindexbold{best_tac} $pack$ $i$] 
applies {\tt step_tac} using best-first search to solve subgoal~$i$.  It is
particularly useful for quantifier duplication (using \ttindex{LK_dup_pack}).
\end{ttdescription}



\index{sequent calculus|)}