Theory Manual
theory Manual
imports Base "HOL-Eisbach.Eisbach_Tools"
begin
chapter ‹The method command›
text ‹
The @{command_def method} command provides the ability to write proof
methods by combining existing ones with their usual syntax. Specifically it
allows compound proof methods to be named, and to extend the name space of
basic methods accordingly. Method definitions may abstract over parameters:
terms, facts, or other methods. They may also provide an optional text
description for display in @{command_ref print_methods}.
┉
The syntax diagram below refers to some syntactic categories that are
further defined in \<^cite>‹"isabelle-isar-ref"›.
\<^rail>‹
@@{command method} name args description @'=' method
;
args: term_args? method_args? ⏎ fact_args? decl_args?
;
term_args: @'for' @{syntax "fixes"}
;
method_args: @'methods' (name+)
;
fact_args: @'uses' (name+)
;
decl_args: @'declares' (name+)
;
description: @{syntax text}?
›
›
section ‹Basic method definitions›
text ‹
Consider the following proof that makes use of usual Isar method
combinators.
›
lemma "P ∧ Q ⟶ P"
by ((rule impI, (erule conjE)?) | assumption)+
text ‹
It is clear that this compound method will be applicable in more cases than
this proof alone. With the @{command method} command we can define a proof
method that makes the above functionality available generally.
›
method prop_solver⇩1 =
((rule impI, (erule conjE)?) | assumption)+
lemma "P ∧ Q ∧ R ⟶ P"
by prop_solver⇩1
text ‹
In this example, the facts ‹impI› and ‹conjE› are static. They are evaluated
once when the method is defined and cannot be changed later. This makes the
method stable in the sense of ∗‹static scoping›: naming another fact ‹impI›
in a later context won't affect the behaviour of ‹prop_solver⇩1›.
The following example defines the same method and gives it a description for the
@{command_ref print_methods} command.
›
method prop_solver⇩2 ‹solver for propositional formulae› =
((rule impI, (erule conjE)?) | assumption)+
section ‹Term abstraction›
text ‹
Methods can also abstract over terms using the @{keyword_def "for"} keyword,
optionally providing type constraints. For instance, the following proof
method ‹intro_ex› takes a term \<^term>‹y› of any type, which it uses to
instantiate the \<^term>‹x›-variable of ‹exI› (existential introduction)
before applying the result as a rule. The instantiation is performed here by
Isar's @{attribute_ref "where"} attribute. If the current subgoal is to find
a witness for the given predicate \<^term>‹Q›, then this has the effect of
committing to \<^term>‹y›.
›
method intro_ex for Q :: "'a ⇒ bool" and y :: 'a =
(rule exI ["where" P = Q and x = y])
text ‹
The term parameters \<^term>‹y› and \<^term>‹Q› can be used arbitrarily inside
the method body, as part of attribute applications or arguments to other
methods. The expression is type-checked as far as possible when the method
is defined, however dynamic type errors can still occur when it is invoked
(e.g.\ when terms are instantiated in a parameterized fact). Actual term
arguments are supplied positionally, in the same order as in the method
definition.
›
lemma "P a ⟹ ∃x. P x"
by (intro_ex P a)
section ‹Fact abstraction›
subsection ‹Named theorems›
text ‹
A ∗‹named theorem› is a fact whose contents are produced dynamically within
the current proof context. The Isar command @{command_ref "named_theorems"}
declares a dynamic fact with a corresponding ∗‹attribute› of the same
name. This allows to maintain a collection of facts in the context as
follows:
›
named_theorems intros
text ‹
So far ‹intros› refers to the empty fact. Using the Isar command
@{command_ref "declare"} we may apply declaration attributes to the context.
Below we declare both ‹conjI› and ‹impI› as ‹intros›, adding them to the
named theorem slot.
›
declare conjI [intros] and impI [intros]
text ‹
We can refer to named theorems as dynamic facts within a particular proof
context, which are evaluated whenever the method is invoked. Instead of
having facts hard-coded into the method, as in ‹prop_solver⇩1›, we can
instead refer to these named theorems.
›
named_theorems elims
declare conjE [elims]
method prop_solver⇩3 =
((rule intros, (erule elims)?) | assumption)+
lemma "P ∧ Q ⟶ P"
by prop_solver⇩3
text ‹
Often these named theorems need to be augmented on the spot, when a method
is invoked. The @{keyword_def "declares"} keyword in the signature of
@{command method} adds the common method syntax ‹method decl: facts› for
each named theorem ‹decl›.
›
method prop_solver⇩4 declares intros elims =
((rule intros, (erule elims)?) | assumption)+
lemma "P ∧ (P ⟶ Q) ⟶ Q ∧ P"
by (prop_solver⇩4 elims: impE intros: conjI)
subsection ‹Simple fact abstraction›
text ‹
The @{keyword "declares"} keyword requires that a corresponding dynamic fact
has been declared with @{command_ref named_theorems}. This is useful for
managing collections of facts which are to be augmented with declarations,
but is overkill if we simply want to pass a fact to a method.
We may use the @{keyword_def "uses"} keyword in the method header to provide
a simple fact parameter. In contrast to @{keyword "declares"}, these facts
are always implicitly empty unless augmented when the method is invoked.
›
method rule_twice uses my_rule =
(rule my_rule, rule my_rule)
lemma "P ⟹ Q ⟹ (P ∧ Q) ∧ Q"
by (rule_twice my_rule: conjI)
section ‹Higher-order methods›
text ‹
The ∗‹structured concatenation› combinator ``‹method⇩1 ; method⇩2›'' was
introduced in Isabelle2015, motivated by the development of Eisbach. It is
similar to ``‹method⇩1, method⇩2›'', but ‹method⇩2› is invoked on ∗‹all›
subgoals that have newly emerged from ‹method⇩1›. This is useful to handle
cases where the number of subgoals produced by a method is determined
dynamically at run-time.
›
method conj_with uses rule =
(intro conjI ; intro rule)
lemma
assumes A: "P"
shows "P ∧ P ∧ P"
by (conj_with rule: A)
text ‹
Method definitions may take other methods as arguments, and thus implement
method combinators with prefix syntax. For example, to more usefully exploit
Isabelle's backtracking, the explicit requirement that a method solve all
produced subgoals is frequently useful. This can easily be written as a
∗‹higher-order method› using ``‹;›''. The @{keyword "methods"} keyword
denotes method parameters that are other proof methods to be invoked by the
method being defined.
›
method solve methods m = (m ; fail)
text ‹
Given some method-argument ‹m›, ‹solve ‹m›› applies the method ‹m› and then
fails whenever ‹m› produces any new unsolved subgoals --- i.e. when ‹m›
fails to completely discharge the goal it was applied to.
›
section ‹Example›
text ‹
With these simple features we are ready to write our first non-trivial proof
method. Returning to the first-order logic example, the following method
definition applies various rules with their canonical methods.
›
named_theorems subst
method prop_solver declares intros elims subst =
(assumption |
(rule intros) | erule elims |
subst subst | subst (asm) subst |
(erule notE ; solve ‹prop_solver›))+
text ‹
The only non-trivial part above is the final alternative ‹(erule notE ;
solve ‹prop_solver›)›. Here, in the case that all other alternatives fail,
the method takes one of the assumptions \<^term>‹¬ P› of the current goal
and eliminates it with the rule ‹notE›, causing the goal to be proved to
become \<^term>‹P›. The method then recursively invokes itself on the
remaining goals. The job of the recursive call is to demonstrate that there
is a contradiction in the original assumptions (i.e.\ that \<^term>‹P› can be
derived from them). Note this recursive invocation is applied with the
@{method solve} method combinator to ensure that a contradiction will indeed
be shown. In the case where a contradiction cannot be found, backtracking
will occur and a different assumption \<^term>‹¬ Q› will be chosen for
elimination.
Note that the recursive call to @{method prop_solver} does not have any
parameters passed to it. Recall that fact parameters, e.g.\ ‹intros›,
‹elims›, and ‹subst›, are managed by declarations in the current proof
context. They will therefore be passed to any recursive call to @{method
prop_solver} and, more generally, any invocation of a method which declares
these named theorems.
┉
After declaring some standard rules to the context, the @{method
prop_solver} becomes capable of solving non-trivial propositional
tautologies.
›
lemmas [intros] =
conjI
impI
disjCI
iffI
notI
lemmas [elims] =
impCE
conjE
disjE
lemma "(A ∨ B) ∧ (A ⟶ C) ∧ (B ⟶ C) ⟶ C"
by prop_solver
chapter ‹The match method \label{s:matching}›
text ‹
So far we have seen methods defined as simple combinations of other methods.
Some familiar programming language concepts have been introduced (i.e.\
abstraction and recursion). The only control flow has been implicitly the
result of backtracking. When designing more sophisticated proof methods this
proves too restrictive and difficult to manage conceptually.
To address this, we introduce the @{method_def match} method, which provides
more direct access to the higher-order matching facility at the core of
Isabelle. It is implemented as a separate proof method (in Isabelle/ML), and
thus can be directly applied to proofs, however it is most useful when
applied in the context of writing Eisbach method definitions.
┉
The syntax diagram below refers to some syntactic categories that are
further defined in \<^cite>‹"isabelle-isar-ref"›.
\<^rail>‹
@@{method match} kind @'in' (pattern '⇒' @{syntax text} + '¦')
;
kind:
(@'conclusion' | @'premises' ('(' 'local' ')')? |
'(' term ')' | @{syntax thms})
;
pattern: fact_name? term args? ⏎ (@'for' fixes)?
;
fact_name: @{syntax name} @{syntax attributes}? ':'
;
args: '(' (('multi' | 'cut' nat?) + ',') ')'
›
Matching allows methods to introspect the goal state, and to implement more
explicit control flow. In the basic case, a term or fact ‹ts› is given to
match against as a ∗‹match target›, along with a collection of
pattern-method pairs ‹(p, m)›: roughly speaking, when the pattern ‹p›
matches any member of ‹ts›, the ∗‹inner› method ‹m› will be executed.
›
lemma
assumes X:
"Q ⟶ P"
"Q"
shows P
by (match X in I: "Q ⟶ P" and I': "Q" ⇒ ‹insert mp [OF I I']›)
text ‹
In this example we have a structured Isar proof, with the named assumption
‹X› and a conclusion \<^term>‹P›. With the match method we can find the
local facts \<^term>‹Q ⟶ P› and \<^term>‹Q›, binding them to separately as
‹I› and ‹I'›. We then specialize the modus-ponens rule @{thm mp [of Q P]} to
these facts to solve the goal.
›
section ‹Subgoal focus›
text‹
In the previous example we were able to match against an assumption out of
the Isar proof state. In general, however, proof subgoals can be
∗‹unstructured›, with goal parameters and premises arising from rule
application. To address this, @{method match} uses ∗‹subgoal focusing› to
produce structured goals out of unstructured ones. In place of fact or term,
we may give the keyword @{keyword_def "premises"} as the match target. This
causes a subgoal focus on the first subgoal, lifting local goal parameters
to fixed term variables and premises into hypothetical theorems. The match
is performed against these theorems, naming them and binding them as
appropriate. Similarly giving the keyword @{keyword_def "conclusion"}
matches against the conclusion of the first subgoal.
An unstructured version of the previous example can then be similarly solved
through focusing.
›
lemma "Q ⟶ P ⟹ Q ⟹ P"
by (match premises in
I: "Q ⟶ P" and I': "Q" ⇒ ‹insert mp [OF I I']›)
text ‹
Match variables may be specified by giving a list of @{keyword_ref
"for"}-fixes after the pattern description. This marks those terms as bound
variables, which may be used in the method body.
›
lemma "Q ⟶ P ⟹ Q ⟹ P"
by (match premises in I: "Q ⟶ A" and I': "Q" for A ⇒
‹match conclusion in A ⇒ ‹insert mp [OF I I']››)
text ‹
In this example \<^term>‹A› is a match variable which is bound to \<^term>‹P›
upon a successful match. The inner @{method match} then matches the
now-bound \<^term>‹A› (bound to \<^term>‹P›) against the conclusion (also \<^term>‹P›), finally applying the specialized rule to solve the goal.
Schematic terms like ‹?P› may also be used to specify match variables, but
the result of the match is not bound, and thus cannot be used in the inner
method body.
┉
In the following example we extract the predicate of an existentially
quantified conclusion in the current subgoal and search the current premises
for a matching fact. If both matches are successful, we then instantiate the
existential introduction rule with both the witness and predicate, solving
with the matched premise.
›
method solve_ex =
(match conclusion in "∃x. Q x" for Q ⇒
‹match premises in U: "Q y" for y ⇒
‹rule exI [where P = Q and x = y, OF U]››)
text ‹
The first @{method match} matches the pattern \<^term>‹∃x. Q x› against the
current conclusion, binding the term \<^term>‹Q› in the inner match. Next
the pattern ‹Q y› is matched against all premises of the current subgoal. In
this case \<^term>‹Q› is fixed and \<^term>‹y› may be instantiated. Once a
match is found, the local fact ‹U› is bound to the matching premise and the
variable \<^term>‹y› is bound to the matching witness. The existential
introduction rule ‹exI:›~@{thm exI} is then instantiated with \<^term>‹y› as
the witness and \<^term>‹Q› as the predicate, with its proof obligation
solved by the local fact U (using the Isar attribute @{attribute OF}). The
following example is a trivial use of this method.
›
lemma "halts p ⟹ ∃x. halts x"
by solve_ex
subsection ‹Operating within a focus›
text ‹
Subgoal focusing provides a structured form of a subgoal, allowing for more
expressive introspection of the goal state. This requires some consideration
in order to be used effectively. When the keyword @{keyword "premises"} is
given as the match target, the premises of the subgoal are lifted into
hypothetical theorems, which can be found and named via match patterns.
Additionally these premises are stripped from the subgoal, leaving only the
conclusion. This renders them inaccessible to standard proof methods which
operate on the premises, such as @{method frule} or @{method erule}. Naive
usage of these methods within a match will most likely not function as the
method author intended.
›
method my_allE_bad for y :: 'a =
(match premises in I: "∀x :: 'a. ?Q x" ⇒
‹erule allE [where x = y]›)
text ‹
Here we take a single parameter \<^term>‹y› and specialize the universal
elimination rule (@{thm allE}) to it, then attempt to apply this specialized
rule with @{method erule}. The method @{method erule} will attempt to unify
with a universal quantifier in the premises that matches the type of \<^term>‹y›. Since @{keyword "premises"} causes a focus, however, there are no
subgoal premises to be found and thus @{method my_allE_bad} will always
fail. If focusing instead left the premises in place, using methods like
@{method erule} would lead to unintended behaviour, specifically during
backtracking. In our example, @{method erule} could choose an alternate
premise while backtracking, while leaving ‹I› bound to the original match.
In the case of more complex inner methods, where either ‹I› or bound terms
are used, this would almost certainly not be the intended behaviour.
An alternative implementation would be to specialize the elimination rule to
the bound term and apply it directly.
›
method my_allE_almost for y :: 'a =
(match premises in I: "∀x :: 'a. ?Q x" ⇒
‹rule allE [where x = y, OF I]›)
lemma "∀x. P x ⟹ P y"
by (my_allE_almost y)
text ‹
This method will insert a specialized duplicate of a universally quantified
premise. Although this will successfully apply in the presence of such a
premise, it is not likely the intended behaviour. Repeated application of
this method will produce an infinite stream of duplicate specialized
premises, due to the original premise never being removed. To address this,
matched premises may be declared with the @{attribute thin} attribute. This
will hide the premise from subsequent inner matches, and remove it from the
list of premises when the inner method has finished and the subgoal is
unfocused. It can be considered analogous to the existing ‹thin_tac›.
To complete our example, the correct implementation of the method will
@{attribute thin} the premise from the match and then apply it to the
specialized elimination rule.›
method my_allE for y :: 'a =
(match premises in I [thin]: "∀x :: 'a. ?Q x" ⇒
‹rule allE [where x = y, OF I]›)
lemma "∀x. P x ⟹ ∀x. Q x ⟹ P y ∧ Q y"
by (my_allE y)+ (rule conjI)
subsubsection ‹Inner focusing›
text ‹
Premises are ∗‹accumulated› for the purposes of subgoal focusing. In
contrast to using standard methods like @{method frule} within focused
match, another @{method match} will have access to all the premises of the
outer focus.
›
lemma "A ⟹ B ⟹ A ∧ B"
by (match premises in H: A ⇒ ‹intro conjI, rule H,
match premises in H': B ⇒ ‹rule H'››)
text ‹
In this example, the inner @{method match} can find the focused premise
\<^term>‹B›. In contrast, the @{method assumption} method would fail here due
to \<^term>‹B› not being logically accessible.
›
lemma "A ⟹ A ∧ (B ⟶ B)"
by (match premises in H: A ⇒ ‹intro conjI, rule H, rule impI,
match premises (local) in A ⇒ ‹fail›
¦ H': B ⇒ ‹rule H'››)
text ‹
In this example, the only premise that exists in the first focus is \<^term>‹A›. Prior to the inner match, the rule ‹impI› changes the goal \<^term>‹B ⟶
B› into \<^term>‹B ⟹ B›. A standard premise match would also include \<^term>‹A› as an original premise of the outer match. The ‹local› argument limits
the match to newly focused premises.
›
section ‹Attributes›
text ‹
Attributes may throw errors when applied to a given fact. For example, rule
instantiation will fail if there is a type mismatch or if a given variable
doesn't exist. Within a match or a method definition, it isn't generally
possible to guarantee that applied attributes won't fail. For example, in
the following method there is no guarantee that the two provided facts will
necessarily compose.
›
method my_compose uses rule1 rule2 =
(rule rule1 [OF rule2])
text ‹
Some attributes (like @{attribute OF}) have been made partially
Eisbach-aware. This means that they are able to form a closure despite not
necessarily always being applicable. In the case of @{attribute OF}, it is
up to the proof author to guard attribute application with an appropriate
@{method match}, but there are still no static guarantees.
In contrast to @{attribute OF}, the @{attribute "where"} and @{attribute of}
attributes attempt to provide static guarantees that they will apply
whenever possible.
Within a match pattern for a fact, each outermost quantifier specifies the
requirement that a matching fact must have a schematic variable at that
point. This gives a corresponding name to this ``slot'' for the purposes of
forming a static closure, allowing the @{attribute "where"} attribute to
perform an instantiation at run-time.
›
lemma
assumes A: "Q ⟹ False"
shows "¬ Q"
by (match intros in X: "⋀P. (P ⟹ False) ⟹ ¬ P" ⇒
‹rule X [where P = Q, OF A]›)
text ‹
Subgoal focusing converts the outermost quantifiers of premises into
schematics when lifting them to hypothetical facts. This allows us to
instantiate them with @{attribute "where"} when using an appropriate match
pattern.
›
lemma "(⋀x :: 'a. A x ⟹ B x) ⟹ A y ⟹ B y"
by (match premises in I: "⋀x :: 'a. ?P x ⟹ ?Q x" ⇒
‹rule I [where x = y]›)
text ‹
The @{attribute of} attribute behaves similarly. It is worth noting,
however, that the positional instantiation of @{attribute of} occurs against
the position of the variables as they are declared ∗‹in the match pattern›.
›
lemma
fixes A B and x :: 'a and y :: 'b
assumes asm: "(⋀x y. A y x ⟹ B x y )"
shows "A y x ⟹ B x y"
by (match asm in I: "⋀(x :: 'a) (y :: 'b). ?P x y ⟹ ?Q x y" ⇒
‹rule I [of x y]›)
text ‹
In this example, the order of schematics in ‹asm› is actually ‹?y ?x›, but
we instantiate our matched rule in the opposite order. This is because the
effective rule \<^term>‹I› was bound from the match, which declared the \<^typ>‹'a› slot first and the \<^typ>‹'b› slot second.
To get the dynamic behaviour of @{attribute of} we can choose to invoke it
∗‹unchecked›. This avoids trying to do any type inference for the provided
parameters, instead storing them as their most general type and doing type
matching at run-time. This, like @{attribute OF}, will throw errors if the
expected slots don't exist or there is a type mismatch.
›
lemma
fixes A B and x :: 'a and y :: 'b
assumes asm: "⋀x y. A y x ⟹ B x y"
shows "A y x ⟹ B x y"
by (match asm in I: "PROP ?P" ⇒ ‹rule I [of (unchecked) y x]›)
text ‹
Attributes may be applied to matched facts directly as they are matched. Any
declarations will therefore be applied in the context of the inner method,
as well as any transformations to the rule.
›
lemma "(⋀x :: 'a. A x ⟹ B x) ⟹ A y ⟶ B y"
by (match premises in I [of y, intros]: "⋀x :: 'a. ?P x ⟹ ?Q x" ⇒
‹prop_solver›)
text ‹
In this example, the pattern ‹⋀x :: 'a. ?P x ⟹ ?Q x› matches against the
only premise, giving an appropriately typed slot for \<^term>‹y›. After the
match, the resulting rule is instantiated to \<^term>‹y› and then declared as
an @{attribute intros} rule. This is then picked up by @{method prop_solver}
to solve the goal.
›
section ‹Multi-match \label{sec:multi}›
text ‹
In all previous examples, @{method match} was only ever searching for a
single rule or premise. Each local fact would therefore always have a length
of exactly one. We may, however, wish to find ∗‹all› matching results. To
achieve this, we can simply mark a given pattern with the ‹(multi)›
argument.
›
lemma
assumes asms: "A ⟹ B" "A ⟹ D"
shows "(A ⟶ B) ∧ (A ⟶ D)"
apply (match asms in I [intros]: "?P ⟹ ?Q" ⇒ ‹solves ‹prop_solver››)?
apply (match asms in I [intros]: "?P ⟹ ?Q" (multi) ⇒ ‹prop_solver›)
done
text ‹
In the first @{method match}, without the ‹(multi)› argument, \<^term>‹I› is
only ever be bound to one of the members of ‹asms›. This backtracks over
both possibilities (see next section), however neither assumption in
isolation is sufficient to solve to goal. The use of the @{method solves}
combinator ensures that @{method prop_solver} has no effect on the goal when
it doesn't solve it, and so the first match leaves the goal unchanged. In
the second @{method match}, ‹I› is bound to all of ‹asms›, declaring both
results as ‹intros›. With these rules @{method prop_solver} is capable of
solving the goal.
Using for-fixed variables in patterns imposes additional constraints on the
results. In all previous examples, the choice of using ‹?P› or a for-fixed
\<^term>‹P› only depended on whether or not \<^term>‹P› was mentioned in another
pattern or the inner method. When using a multi-match, however, all
for-fixed terms must agree in the results.
›
lemma
assumes asms: "A ⟹ B" "A ⟹ D" "D ⟹ B"
shows "(A ⟶ B) ∧ (A ⟶ D)"
apply (match asms in I [intros]: "?P ⟹ Q" (multi) for Q ⇒
‹solves ‹prop_solver››)?
apply (match asms in I [intros]: "P ⟹ ?Q" (multi) for P ⇒
‹prop_solver›)
done
text ‹
Here we have two seemingly-equivalent applications of @{method match},
however only the second one is capable of solving the goal. The first
@{method match} selects the first and third members of ‹asms› (those that
agree on their conclusion), which is not sufficient. The second @{method
match} selects the first and second members of ‹asms› (those that agree on
their assumption), which is enough for @{method prop_solver} to solve the
goal.
›
section ‹Dummy patterns›
text ‹
Dummy patterns may be given as placeholders for unique schematics in
patterns. They implicitly receive all currently bound variables as
arguments, and are coerced into the \<^typ>‹prop› type whenever possible. For
example, the trivial dummy pattern ‹_› will match any proposition. In
contrast, by default the pattern ‹?P› is considered to have type \<^typ>‹bool›. It will not bind anything with meta-logical connectives (e.g. ‹_ ⟹ _›
or ‹_ &&& _›).
›
lemma
assumes asms: "A &&& B ⟹ D"
shows "(A ∧ B ⟶ D)"
by (match asms in I: _ ⇒ ‹prop_solver intros: I conjunctionI›)
section ‹Backtracking›
text ‹
Patterns are considered top-down, executing the inner method ‹m› of the
first pattern which is satisfied by the current match target. By default,
matching performs extensive backtracking by attempting all valid variable
and fact bindings according to the given pattern. In particular, all
unifiers for a given pattern will be explored, as well as each matching
fact. The inner method ‹m› will be re-executed for each different
variable/fact binding during backtracking. A successful match is considered
a cut-point for backtracking. Specifically, once a match is made no other
pattern-method pairs will be considered.
The method ‹foo› below fails for all goals that are conjunctions. Any such
goal will match the first pattern, causing the second pattern (that would
otherwise match all goals) to never be considered.
›
method foo =
(match conclusion in "?P ∧ ?Q" ⇒ ‹fail› ¦ "?R" ⇒ ‹prop_solver›)
text ‹
The failure of an inner method that is executed after a successful match
will cause the entire match to fail. This distinction is important due to
the pervasive use of backtracking. When a method is used in a combinator
chain, its failure becomes significant because it signals previously applied
methods to move to the next result. Therefore, it is necessary for @{method
match} to not mask such failure. One can always rewrite a match using the
combinators ``‹?›'' and ``‹|›'' to try subsequent patterns in the case of an
inner-method failure. The following proof method, for example, always
invokes @{method prop_solver} for all goals because its first alternative
either never matches or (if it does match) always fails.
›
method foo⇩1 =
(match conclusion in "?P ∧ ?Q" ⇒ ‹fail›) |
(match conclusion in "?R" ⇒ ‹prop_solver›)
subsection ‹Cut›
text ‹
Backtracking may be controlled more precisely by marking individual patterns
as ‹cut›. This causes backtracking to not progress beyond this pattern: once
a match is found no others will be considered.
›
method foo⇩2 =
(match premises in I: "P ∧ Q" (cut) and I': "P ⟶ ?U" for P Q ⇒
‹rule mp [OF I' I [THEN conjunct1]]›)
text ‹
In this example, once a conjunction is found (\<^term>‹P ∧ Q›), all possible
implications of \<^term>‹P› in the premises are considered, evaluating the
inner @{method rule} with each consequent. No other conjunctions will be
considered, with method failure occurring once all implications of the form
‹P ⟶ ?U› have been explored. Here the left-right processing of individual
patterns is important, as all patterns after of the cut will maintain their
usual backtracking behaviour.
›
lemma "A ∧ B ⟹ A ⟶ D ⟹ A ⟶ C ⟹ C"
by foo⇩2
lemma "C ∧ D ⟹ A ∧ B ⟹ A ⟶ C ⟹ C"
by (foo⇩2 | prop_solver)
text ‹
In this example, the first lemma is solved by ‹foo⇩2›, by first picking
\<^term>‹A ⟶ D› for ‹I'›, then backtracking and ultimately succeeding after
picking \<^term>‹A ⟶ C›. In the second lemma, however, \<^term>‹C ∧ D› is
matched first, the second pattern in the match cannot be found and so the
method fails, falling through to @{method prop_solver}.
More precise control is also possible by giving a positive number ‹n› as an
argument to ‹cut›. This will limit the number of backtracking results of
that match to be at most ‹n›. The match argument ‹(cut 1)› is the same as
simply ‹(cut)›.
›
subsection ‹Multi-match revisited›
text ‹
A multi-match will produce a sequence of potential bindings for for-fixed
variables, where each binding environment is the result of matching against
at least one element from the match target. For each environment, the match
result will be all elements of the match target which agree with the pattern
under that environment. This can result in unexpected behaviour when giving
very general patterns.
›
lemma
assumes asms: "⋀x. A x ∧ B x" "⋀y. A y ∧ C y" "⋀z. B z ∧ C z"
shows "A x ∧ C x"
by (match asms in I: "⋀x. P x ∧ ?Q x" (multi) for P ⇒
‹match (P) in "A" ⇒ ‹fail›
¦ _ ⇒ ‹match I in "⋀x. A x ∧ B x" ⇒ ‹fail›
¦ _ ⇒ ‹rule I›››)
text ‹
Intuitively it seems like this proof should fail to check. The first match
result, which binds \<^term>‹I› to the first two members of ‹asms›, fails the
second inner match due to binding \<^term>‹P› to \<^term>‹A›. Backtracking then
attempts to bind \<^term>‹I› to the third member of ‹asms›. This passes all
inner matches, but fails when @{method rule} cannot successfully apply this
to the current goal. After this, a valid match that is produced by the
unifier is one which binds \<^term>‹P› to simply ‹λa. A ?x›. The first inner
match succeeds because ‹λa. A ?x› does not match \<^term>‹A›. The next inner
match succeeds because \<^term>‹I› has only been bound to the first member of
‹asms›. This is due to @{method match} considering ‹λa. A ?x› and ‹λa. A ?y›
as distinct terms.
The simplest way to address this is to explicitly disallow term bindings
which we would consider invalid.
›
method abs_used for P =
(match (P) in "λa. ?P" ⇒ ‹fail› ¦ _ ⇒ ‹-›)
text ‹
This method has no effect on the goal state, but instead serves as a filter
on the environment produced from match.
›
section ‹Uncurrying›
text ‹
The @{method match} method is not aware of the logical content of match
targets. Each pattern is simply matched against the shallow structure of a
fact or term. Most facts are in ∗‹normal form›, which curries premises via
meta-implication ‹_ ⟹ _›.
›
lemma
assumes asms: "D ⟹ B ⟹ C" "D ⟹ A"
shows "D ⟹ B ⟹ C ∧ A"
by (match asms in H: "D ⟹ _" (multi) ⇒ ‹prop_solver elims: H›)
text ‹
For the first member of ‹asms› the dummy pattern successfully matches
against \<^term>‹B ⟹ C› and so the proof is successful.
›
lemma
assumes asms: "A ⟹ B ⟹ C" "D ⟹ C"
shows "D ∨ (A ∧ B) ⟹ C"
apply (match asms in H: "_ ⟹ C" (multi) ⇒ ‹prop_solver elims: H›)?
apply (prop_solver elims: asms)
done
text ‹
This proof will fail to solve the goal. Our match pattern will only match
rules which have a single premise, and conclusion \<^term>‹C›, so the first
member of ‹asms› is not bound and thus the proof fails. Matching a pattern
of the form \<^term>‹P ⟹ Q› against this fact will bind \<^term>‹P› to
\<^term>‹A› and \<^term>‹Q› to \<^term>‹B ⟹ C›. Our pattern, with a concrete
\<^term>‹C› in the conclusion, will fail to match this fact.
To express our desired match, we may ∗‹uncurry› our rules before matching
against them. This forms a meta-conjunction of all premises in a fact, so
that only one implication remains. For example the uncurried version of
\<^term>‹A ⟹ B ⟹ C› is \<^term>‹A &&& B ⟹ C›. This will now match our
desired pattern ‹_ ⟹ C›, and can be ∗‹curried› after the match to put it
back into normal form.
›
lemma
assumes asms: "A ⟹ B ⟹ C" "D ⟹ C"
shows "D ∨ (A ∧ B) ⟹ C"
by (match asms [uncurry] in H [curry]: "_ ⟹ C" (multi) ⇒
‹prop_solver elims: H›)
section ‹Reverse matching›
text ‹
The @{method match} method only attempts to perform matching of the pattern
against the match target. Specifically this means that it will not
instantiate schematic terms in the match target.
›
lemma
assumes asms: "⋀x :: 'a. A x"
shows "A y"
apply (match asms in H: "A y" ⇒ ‹rule H›)?
apply (match asms in H: P for P ⇒
‹match ("A y") in P ⇒ ‹rule H››)
done
text ‹
In the first @{method match} we attempt to find a member of ‹asms› which
matches our goal precisely. This fails due to no such member existing. The
second match reverses the role of the fact in the match, by first giving a
general pattern \<^term>‹P›. This bound pattern is then matched against \<^term>‹A y›. In this case, \<^term>‹P› is bound to ‹A ?x› and so it successfully
matches.
›
section ‹Type matching›
text ‹
The rule instantiation attributes @{attribute "where"} and @{attribute "of"}
attempt to guarantee type-correctness wherever possible. This can require
additional invocations of @{method match} in order to statically ensure that
instantiation will succeed.
›
lemma
assumes asms: "⋀x :: 'a. A x"
shows "A y"
by (match asms in H: "⋀z :: 'b. P z" for P ⇒
‹match (y) in "y :: 'b" for y ⇒ ‹rule H [where z = y]››)
text ‹
In this example the type ‹'b› is matched to ‹'a›, however statically they
are formally distinct types. The first match binds ‹'b› while the inner
match serves to coerce \<^term>‹y› into having the type ‹'b›. This allows the
rule instantiation to successfully apply.
›
chapter ‹Method development›
section ‹Tracing methods›
text ‹
Method tracing is supported by auxiliary print methods provided by \<^theory>‹HOL-Eisbach.Eisbach_Tools›. These include @{method print_fact}, @{method
print_term} and @{method print_type}. Whenever a print method is evaluated
it leaves the goal unchanged and writes its argument as tracing output.
Print methods can be combined with the @{method fail} method to investigate
the backtracking behaviour of a method.
›
lemma
assumes asms: A B C D
shows D
apply (match asms in H: _ ⇒ ‹print_fact H, fail›)?
apply (simp add: asms)
done
text ‹
This proof will fail, but the tracing output will show the order that the
assumptions are attempted.
›
section ‹Integrating with Isabelle/ML›
subsubsection ‹Attributes›
text ‹
A custom rule attribute is a simple way to extend the functionality of
Eisbach methods. The dummy rule attribute notation (‹[[ _ ]]›) invokes the
given attribute against a dummy fact and evaluates to the result of that
attribute. When used as a match target, this can serve as an effective
auxiliary function.
›
attribute_setup get_split_rule =
‹Args.term >> (fn t =>
Thm.rule_attribute [] (fn context => fn _ =>
(case get_split_rule (Context.proof_of context) t of
SOME thm => thm
| NONE => Drule.dummy_thm)))›
text ‹
In this example, the new attribute @{attribute get_split_rule} lifts the ML
function of the same name into an attribute. When applied to a case
distinction over a datatype, it retrieves its corresponding split rule.
We can then integrate this into a method that applies the split rule, first
matching to ensure that fetching the rule was successful.
›
declare TrueI [intros]
method splits =
(match conclusion in "?P f" for f ⇒
‹match [[get_split_rule f]] in U: "(_ :: bool) = _" ⇒
‹rule U [THEN iffD2]››)
lemma "L ≠ [] ⟹ case L of [] ⇒ False | _ ⇒ True"
apply splits
apply (prop_solver intros: allI)
done
text ‹
Here the new @{method splits} method transforms the goal to use only logical
connectives: \<^term>‹L = [] ⟶ False ∧ (∀x y. L = x # y ⟶ True)›. This goal
is then in a form solvable by @{method prop_solver} when given the universal
quantifier introduction rule ‹allI›.
›
end