BUILT-IN-CLAUSES

Parent topic:  RULE-CLASSES
Home

See rule-classes for a general discussion of rule classes and how they are used to build rules from formulas. A :built-in-clause rule can be built from any formula other than propositional tautologies. Roughly speaking, the system uses the list of built-in clauses as the first method of proof when attacking a new goal. Any goal that is subsumed by a built in clause is proved ``silently.''

ACL2 maintains a set of ``built-in'' clauses that are used to short-circuit certain theorem proving tasks. We discuss this at length below. When a theorem is given the rule class :built-in-clause ACL2 flattens the implies and and structure of the :corollary formula so as to obtain a set of formulas whose conjunction is equivalent to the given corollary. It then converts each of these to clausal form and adds each clause to the set of built-in clauses.

For example, the following :corollary (regardless of the definition of abl)

(and (implies (and (true-listp x)
                   (not (equal x nil)))
              (< (acl2-count (abl x))
                 (acl2-count x)))
     (implies (and (true-listp x)
                   (not (equal nil x)))
              (< (acl2-count (abl x))
                 (acl2-count x))))

{(not (true-listp x))
 (equal x nil)
 (< (acl2-count (abl x)) (acl2-count x))}

{(not (true-listp x))
 (equal nil x)
 (< (acl2-count (abl x)) (acl2-count x))}.

Recall that a clause is a set of terms, implicitly representing the disjunction of the terms. Clause c1 is ``subsumed'' by clause c2 if some instance of c2 is a subset c1.

For example, let c1 be

{(not (consp l))
 (equal a (car l))
 (< (acl2-count (cdr l)) (acl2-count l))}.

{(not (consp x))
 ; second term omitted here
 (< (acl2-count (cdr x)) (acl2-count x))}

Observe that c1 is the clausal form of

(implies (and (consp l)
              (not (equal a (car l))))
         (< (acl2-count (cdr l)) (acl2-count l))),

(implies (consp l)
         (< (acl2-count (cdr l)) (acl2-count l)))

The set of built-in clauses is just a set of known theorems in clausal form. Any formula that is subsumed by a built-in clause is thus a theorem. If the set of built-in theorems is reasonably small, this little theorem prover is fast. ACL2 uses the ``built-in clause check'' in three places: at the entrance to the general-purpose theorem prover, as a filter on the clauses generated to prove the termination of recursively defun'd functions and as a filter on the clauses generated to verify the guards of a function.

The latter two uses are the ones that most often motivate an extension to the set of built-in clauses. Frequently a given formalization problem requires the definition of many functions which require virtually identical termination and/or guard proofs. These proofs can be short-circuited by extending the set of built-in clauses to contain the most general forms of the clauses generated by the definitional schemes in use.

The attentive user might have noticed that there are some recursive schemes, e.g., recursion by cdr after testing consp, that ACL2 just seems to ``know'' are ok, while for others it generates measure clauses to prove. Actually, it always generates measure clauses but then filters out any that pass the built-in clause check. When ACL2 is initialized, the clause justifying cdr recursion after a consp test is added to the set of built-in clauses. (That clause is c2 above.)

Note that only a subsumptin check is made; no rewriting or simplification is done. Thus, if we want the system to ``know'' that cdr recursion is ok after a negative atom test (which, by the definition of atom, is the same as a consp test), we have to build in a second clause. The subsumption algorithm does not ``know'' about commutative functions. Thus, for predictability, we have built in commuted versions of each clause involving commutative functions. For example, we build in both

{(not (integerp x))
 (< 0 x)
 (= x 0)
 (< (acl2-count (+ -1 x)) (acl2-count x))}

{(not (integerp x))
 (< 0 x)
 (= 0 x)
 (< (acl2-count (+ -1 x)) (acl2-count x))}

:built-in-clause rules added by the user can be enabled and disabled.

COMPOUND-RECOGNIZER

Parent topic:  RULE-CLASSES
Home

See rule-classes for a general discussion of rule classes and how they are used to build rules from formulas. An example :corollary formula from which a :compound-recognizer rule might be built is:

Example: 
(implies (alistp x)         When (alistp x) is assumed true,
         (true-listp x))    add the additional hypothesis that x
                            is of primitive type true-listp.

General Forms:
(implies (fn x) concl)
(implies (not (fn x)) concl)
(and (implies (fn x) concl1)
     (implies (not (fn x)) concl2))
(iff (fn x) concl)
(equal (fn x) concl)

The primitive ACL2 types and a suitable primitive recognizing expression for each are listed below.

  type                suitable primitive recognizer


  zero                (equal x 0)
  negative integers   (and (integerp x) (< x 0))
  positive integers   (and (integerp x) (> x 0))
  negative ratio      (and (rationalp x)
                           (not (integerp x))
                           (< x 0))
  positive ratio      (and (rationalp x)
                           (not (integerp x))
                           (> x 0))
  complex rational    (complex-rationalp x)
  nil                 (equal x nil)
  t                   (equal x t)
  other symbols       (and (symbolp x)
                           (not (equal x nil))
                           (not (equal x t)))
  proper conses       (and (consp x)
                           (true-listp x))
  improper conses     (and (consp x)
                           (not (true-listp x)))
  strings             (stringp x)
  characters          (characterp x)

(or (equal x nil) (and (consp x) (true-listp x)))

There are essentially four forms of :compound-recognizer rules, the last two general forms above being equivalent. We explain how such rules are used by considering the individual forms.

Consider a rule of the first form, (implies (fn x) concl). The effect of such a rule is that when the rewriter assumes (fn x) true, as it would while diving through (if (fn x) xxx ...) to rewrite xxx, it restricts the type of x as specified by concl. However, when assuming (fn x) false, as necessary in (if (fn x) ... xxx), the rule permits no additional assumptions about the type of x. For example, if fn is primep, i.e., the predicate that recognizes prime numbers, then (implies (primep x) (and (integerp x) (>= x 0))) is a compound recognizer rule of the first form. When (primep x) is assumed true, the rewriter gains the additional information that x is a natural number. When (primep x) is assumed false, no additional information is gained -- since x may be a non-prime natural or may not even be a natural.

The second general form addresses itself to the symmetric case, when assuming (fn x) false permits type restrictions on x but assuming (fn x) true permits no such restrictions. For example, if we defined exprp to be the recognizer for well-formed expressions for some language in which all symbols, numbers, character objects and strings were well-formed -- e.g., the well-formedness rules only put restrictions on expressions represented by consps -- then the theorem (implies (not (exprp x)) (consp x)) is a rule of the second form. Assuming (exprp x) true tells us nothing about the type of x; assuming it false tells us x is a consp.

The third general form addresses itself to the case where one type may be deduced from (fn x) and a generally unrelated type may be deduced from its negation. If we modified the expression recognizer above so that character objects are illegal, then a rule of the third form is

(and (implies (exprp x) (not (characterp x)))
     (implies (not (exprp x)) (or (consp x) (characterp x)))).

(iff (booleanp x) (or (equal x t) (equal x nil))).

Often it is best to disable fn after proving that it is a compound recognizer, since (fn x) will not be recognized as a compound recognizer once it has been expanded.

Every time you prove a new compound recognizer rule about fn it overrides all previously proved compound recognizer rules about fn. Thus, if you want to establish the type implied by (fn x) and you want to establish the type implied by (not (fn x)), you must prove a compound recognizer rule of the third or fourth forms. Proving a rule of the first form followed by one of the second only leaves the second fact in the data base.

Compound recognizer rules can be disabled with the effect that older rules about fn, if any, are exposed.

If you prove multiple compound recognizer rules for a function, you may see a warning message to the effect that the new rule is not as ``restrictive'' as the old. That is, the new rules do not give the rewriter strictly more type information than it already had. The new rule is stored anyway, overriding the old, if enabled. You may be playing subtle games with enabling or rewriting. But two other interpretions are more likely, we think. One is that you have forgotten about an earlier rule and should merely print it out to make sure it says what you know and then forget your new rule. The other is that you meant to give the system more information and the system has simply been unable to extract the intended type information from the term you placed in the conclusion of the new rule. Given our lack of specificity in saying how type information is extracted from rules, you can hardly blame yourself for this problem. Sorry. If you suspect you've been burned this way, you should rephrase the new rule in terms of the primitive recognizing expressions above and see if the warning is still given. It would also be helpful to let us see your example so we can consider it as we redesign this stuff.

Compound recognizer rules are similar to :forward-chaining rules in that the system deduces new information from the act of assuming something true or false. If a compound recognizer rule were stored as a forward chaining rule it would have essentially the same effect as described, when it has any effect at all. The important point is that :forward-chaining rules, because of their more general and expensive form, are used ``at the top level'' of the simplification process: we forward chain from assumptions in the goal being proved. But compound recognizer rules are built in at the bottom-most level of the simplifier, where type reasoning is done.

All that said, compound recognizer rules are a rather fancy, specialized mechanism. It may be more appropriate to create :forward-chaining rules instead of :compound-recognizer rules.

CONGRUENCE

Parent topic:  RULE-CLASSES
Home

See rule-classes for a general discussion of rule classes and how they are used to build rules from formulas. An example :corollary formula from which a :congruence rule might be built is:

Example:
(implies (set-equal x y)
         (iff (memb e x) (memb e y))).

General Form:
(implies (equiv1 xk xk-equiv)
         (equiv2 (fn x1... xk       ...xn)
                 (fn x1... xk-equiv ...xn)))

The macro form (defcong equiv1 equiv2 (fn x1 ... x1) k) is an abbreviation for a defthm of rule-class :congruence that attempts to establish that equiv2 is maintained by maintaining equiv1 in fn's kth argument. The defcong macro automatically generates the general formula shown above. See defcong.

The memb example above tells us that (memb e x) is propositionally equivalent to (memb e y), provided x and y are set-equal. The outside equivalence is iff and the inside equivalence for the second argument is set-equal. If we see a memb expression in a propositional context, e.g., as a literal of a clause or test of an if (but not, for example, as an argument to cons), we can rewrite its second argument maintaining set-equality. For example, a rule stating the commutativity of append (modulo set-equality) could be applied in this context. Since equality is a refinement of all equivalence relations, all equality rules are always available. See refinement.

All known :congruence rules about a given outside equivalence and fn can be used independently. That is, consider two :congruence rules with the same outside equivalence, equiv, and about the same function fn. Suppose one says that equiv1 is the inside equivalence for the first argument and the other says equiv2 is the inside equivalence for the second argument. Then (fn a b) is equiv (fn a' b') provided a is equiv1 to a' and b is equiv2 to b'. This is an easy consequence of the transitivity of equiv. It permits you to think independently about the inside equivalences.

Furthermore, it is possible that more than one inside equivalence for a given argument slot will maintain a given outside equivalence. For example, (length a) is equal to (length a') if a and a' are related either by list-equal or by string-equal. You may prove two (or more) :congruence rules for the same slot of a function. The result is that the system uses a new, ``generated'' equivalence relation for that slot with the result that rules of both (or all) kinds are available while rewriting.

:congruence rules can be disabled. For example, if you have two different inside equivalences for a given argument position and you find that the :rewrite rules for one are unexpectedly preventing the application of the desired rule, you can disable the rule that introduced the unwanted inside equivalence.

More will be written about this as we develop the techniques.

DEFINITION

Parent topic:  RULE-CLASSES
Home

Example:
(implies (true-listp x)
         (equal (len x)
                (if (null x)
                    0
                    (if (null (cdr x))
                        1
                        (len (cddr x))))))

General Form:
(implies hyp (equiv (fn a1 ... an) body))

Consider the general form above. Generalized definitions are stored among the :rewrite rules for the function ``defined,'' fn above, but the procedure for applying them is a little different. During rewriting, instances of (fn a1 ... an) are replaced by corresponding instances of body provided the hyps can be established as for a :rewrite rule and the result of rewriting body satisfies the criteria for function expansion. There are two primary criteria, either of which permits expansion. The first is that the ``recursive'' calls of fn in the rewritten body have arguments that already occur in the goal conjecture. The second is that the ``controlling'' arguments to fn are simpler in the rewritten body.

The notions of ``recursive call'' and ``controllers'' are complicated by the provisions for mutually recursive definitions. Consider a ``clique'' of mutually recursive definitions. Then a ``recursive call'' is a call to any function defined in the clique and an argument is a ``controller'' if it is involved in the measure that decreases in all recursive calls. These notions are precisely defined by the definitional principle and do not necessarily make sense in the context of generalized definitional equations as implemented here.

But because the heuristics governing the use of generalized definitions require these notions, it is generally up to the user to specify which calls in body are to be considered recursive and what the controlling arguments are. This information is specified in the :clique and :controller-alist fields of the :definition rule class.

The :clique field is the list of function symbols to be considered recursive calls of fn. In the case of a non-recursive definition, the :clique field is empty; in a singly recursive definition, it should consist of the singleton list containing fn; otherwise it should be a list of all of the functions in the mutually recursive clique with this definition of fn.

If the :clique field is not provided it defaults to nil if fn does not occur as a function symbol in body and it defaults to the singleton list containing fn otherwise. Thus, :clique must be supplied by the user only when the generalized definition rule is to be treated as one of several in a mutually recursive clique.

The :controller-alist is an alist that maps each function symbol in the :clique to a mask specifying which arguments are considered controllers. The mask for a given member of the clique, fn, must be a list of t's and nil's of length equal to the arity of fn. A t should be in each argument position that is considered a ``controller'' of the recursion. For a function admitted under the principle of definition, an argument controls the recursion if it is one of the arguments measured in the termination argument for the function. But in generalized definition rules, the user is free to designate any subset of the arguments as controllers. Failure to choose wisely may result in the ``infinite expansion'' of definitional rules but cannot render ACL2 unsound since the rule being misused is a theorem.

If the :controller-alist is omitted it can sometimes be defaulted automatically by the system. If the :clique is nil, the :controller-alist defaults to nil. If the :clique is a singleton containing fn, the :controller-alist defaults to the controller alist computed by (defun fn args body). If the :clique contains more than one function, the user must supply the :controller-alist specifying the controllers for each function in the clique. This is necessary since the system cannot determine and thus cannot analyze the other definitional equations to be included in the clique.

For example, suppose fn1 and fn2 have been defined one way and it is desired to make ``alternative'' mutually recursive definitions available to the rewriter. Then one would prove two theorems and store each as a :definition rule. These two theorems would exhibit equations ``defining'' fn1 and fn2 in terms of each other. No provision is here made for exhibiting these two equations as a system of equations. One is proved and then the other. It just so happens that the user intends them to be treated as mutually recursive definitions. To achieve this end, both :definition rules should specify the :clique (fn1 fn2) and should specify a suitable :controller-alist. If, for example, the new definition of fn1 is controlled by its first argument and the new definition of fn2 is controlled by its second and third (and they each take three arguments) then a suitable :controller-alist would be ((fn1 t nil nil) (fn2 nil t t)). The order of the pairs in the alist is unimportant, but there must be a pair for each function in the clique.

Inappropriate heuristic advice via :clique and :controller-alist can cause ``infinite expansion'' of generalized definitions, but cannot render ACL2 unsound.

Note that the actual definition of fn1 has the runic name (:definition fn1). The runic name of the alternative definition is (:definition lemma), where lemma is the name given to the event that created the generalized :definition rule. This allows theories to switch between various ``definitions'' of the functions.

The definitional principle, defun, actually adds :definition rules. Thus the handling of generalized definitions is exactly the same as for ``real'' definitions because no distinction is made in the implementation. Suppose (fn x y) is defun'd to be body. Note that defun (or defuns or mutual-recursion) can compute the clique for fn from the syntactic presentation and it can compute the controllers from the termination analysis. Provided the definition is admissible, defun adds the :definition rule (equal (fn x y) body). Thus, (fn a b) will not be expanded unless (g a b) can be established.

ELIM

Parent topic:  RULE-CLASSES
Home

See rule-classes for a general discussion of rule classes and how they are used to build rules from formulas. An example :corollary formula from which an :elim rule might be built is:

Example:
(implies (consp x)                      when (car v) or (cdr v) appears
         (equal (cons (car x) (cdr x))  in a conjecture, and v is a
                x))                     variable, consider replacing v by
                                        (cons a b), for two new variables
                                        a and b.

General Form:
(implies hyp (equal lhs x))

To use an :elim rule, the theorem prover waits until a conjecture has been maximally simplified. If it then finds an instance of some destructor term (fn v1 ... vn) in the conjecture, where the instance for x is some variable symbol, vi, it instantiates the :elim formula as indicated by the destructor term matched, splits the conjecture into two goals, according to whether the instantiated hypothesis, hyp, holds, and in the case that it does, generalizes all the instantiated destructor terms in the conjecture to new variables and then replaces vi in the conjecture by the generalized instantiated lhs.

At the moment, the best description of how ACL2 :elim rules are used may be found in the discussion of ``ELIM Rules,'' pp 247 of A Computational Logic Handbook.

EQUIVALENCE

Parent topic:  RULE-CLASSES
Home

See rule-classes for a general discussion of rule classes and how they are used to build rules from formulas. An example :corollary formula from which a :equivalence rule might be built is as follows. (We assume that r-equal has been defined.)

Example:
(and (booleanp (r-equal x y))
     (r-equal x x)
     (implies (r-equal x y) (r-equal y x))
     (implies (and (r-equal x y)
                   (r-equal y z))
              (r-equal x z))).

General Form:
(and (booleanp (equiv x y))
     (equiv x x)
     (implies (equiv x y) (equiv y x))
     (implies (and (equiv x y)
                   (equiv y z))
              (equiv x z)))

The macro form (defequiv equiv) is an abbreviation for a defthm of rule-class :equivalence that establishes that equiv is an equivalence relation. It generates the formula shown above. See defequiv.

When equiv is marked as an equivalence relation, its reflexivity, symmetry, and transitivity are built into the system in a deeper way than via :rewrite rules. More importantly, after equiv has been shown to be an equivalence relation, lemmas about equiv, e.g.,

(implies hyps (equiv lhs rhs)),

For example, suppose that r-equal is known to be an equivalence relation. The :congruence lemma

(implies (r-equal s1 s2)
         (equal (fn s1 n) (fn s2 n)))

(implies (and (fn x1 y1)
              (fn x2 y2))
         (iff (fn x1 x2) (fn y1 y2)))

The system initially has knowledge of two equivalence relations, equality, denoted by the symbol equal, and propositional equivalence, denoted by iff. Equal is known to be a refinement of all equivalence relations and to preserve equality across all arguments of all functions.

Typically there are five steps involved in introducing and using a new equivalence relation, equiv.

(1) Define equiv,

(2) prove the :equivalence lemma about equiv,

(3) prove the :congruence lemmas that show where equiv can be used to maintain known relations,

(4) prove the :refinement lemmas that relate equiv to known relations other than equal, and

(5) develop the theory of conditional :rewrite rules that drive equiv rewriting.

Among the theorems proved below is

(defthm insert-sort-is-id 
  (perm (insert-sort x) x))

Then we prove the lemma

(defthm insert-is-cons
  (perm (insert a x) (cons a x)))

Now prove (perm (insert-sort x) x). The base case is trivial. The induction step is

   (consp x)
 & (perm (insert-sort (cdr x)) (cdr x))


-> (perm (insert-sort x) x).

   (perm (insert (car x) (insert-sort (cdr x))) x).

(perm (insert (car x) (cdr x)) x)

Here are the events.

(encapsulate ((lt (x y) t))
  (local (defun lt (x y) (declare (ignore x y)) nil))
  (defthm lt-non-symmetric (implies (lt x y) (not (lt y x)))))


(defun insert (x lst)
  (cond ((atom lst) (list x))
        ((lt x (car lst)) (cons x lst))
        (t (cons (car lst) (insert x (cdr lst))))))


(defun insert-sort (lst)
  (cond ((atom lst) nil)
        (t (insert (car lst) (insert-sort (cdr lst))))))


(defun del (x lst)
  (cond ((atom lst) nil)
        ((equal x (car lst)) (cdr lst))
        (t (cons (car lst) (del x (cdr lst))))))


(defun mem (x lst)
  (cond ((atom lst) nil)
        ((equal x (car lst)) t)
        (t (mem x (cdr lst)))))


(defun perm (lst1 lst2)
  (cond ((atom lst1) (atom lst2))
        ((mem (car lst1) lst2)
         (perm (cdr lst1) (del (car lst1) lst2)))
        (t nil)))


(defthm perm-reflexive
  (perm x x))


(defthm perm-cons
  (implies (mem a x)
           (equal (perm x (cons a y))
                  (perm (del a x) y)))
  :hints (("Goal" :induct (perm x y))))


(defthm perm-symmetric
  (implies (perm x y) (perm y x)))


(defthm mem-del
  (implies (mem a (del b x)) (mem a x)))


(defthm perm-mem
  (implies (and (perm x y)
                (mem a x))
           (mem a y)))


(defthm mem-del2
  (implies (and (mem a x)
                (not (equal a b)))
           (mem a (del b x))))


(defthm comm-del
  (equal (del a (del b x)) (del b (del a x))))


(defthm perm-del
  (implies (perm x y)
           (perm (del a x) (del a y))))


(defthm perm-transitive
  (implies (and (perm x y) (perm y z)) (perm x z)))


(defequiv perm)


(in-theory (disable perm
                    perm-reflexive
                    perm-symmetric
                    perm-transitive))


(defcong perm perm (cons x y) 2)


(defcong perm iff (mem x y) 2)


(defthm atom-perm
  (implies (not (consp x)) (perm x nil))
  :rule-classes :forward-chaining
  :hints (("Goal" :in-theory (enable perm))))


(defthm insert-is-cons
  (perm (insert a x) (cons a x)))


(defthm insert-sort-is-id 
  (perm (insert-sort x) x))


(defun app (x y) (if (consp x) (cons (car x) (app (cdr x) y)) y))


(defun rev (x)
  (if (consp x) (app (rev (cdr x)) (list (car x))) nil))


(defcong perm perm (app x y) 2)


(defthm app-cons
  (perm (app a (cons b c)) (cons b (app a c))))


(defthm app-commutes
  (perm (app a b) (app b a)))


(defcong perm perm (app x y) 1
  :hints (("Goal" :induct (app y x))))


(defthm rev-is-id (perm (rev x) x))


(defun == (x y)
  (if (consp x)
      (if (consp y)
          (and (equal (car x) (car y))
               (== (cdr x) (cdr y)))
          nil)
      (not (consp y))))


(defthm ==-symmetric (== x x))


(defthm ==-reflexive (implies (== x y) (== y x)))


(defequiv ==)


(in-theory (disable ==-symmetric ==-reflexive))


(defcong == == (cons x y) 2)


(defcong == iff (consp x) 1)


(defcong == == (app x y) 2)


(defcong == == (app x y) 1)


(defthm rev-rev (== (rev (rev x)) x))