FREE-VARIABLES

free variables in rules 
Major Section:  RULE-CLASSES

As described elsewhere (see rule-classes), ACL2 rules are treated as implications for which there are zero or more hypotheses hj to prove. In particular, rules of class :rewrite may look like this: 

(implies (and h1 ... hn)
         (fn lhs rhs))
Variables of hi are said to occur free in the above :rewrite rule if they do not occur in lhs or in any hj with j<i. (To be precise, here we are only discussing those variables that are not in the scope of a let/let*/lambda that binds them.) We also refer to these as the free variables of the rule. ACL2 may issue a warning or error when there are free variables in a rule, as described below. (Variables of rhs may be considered free if they do not occur in lhs or in any hj. But we do not consider those in this discussion.)

In general, the free variables of rules are those variables occurring in their hypotheses (not let/let*/lambda-bound) that are not bound when the rule is applied. For rules of class :linear and :forward-chaining, variables are bound by a trigger term. (See rule-classes for a discussion of the :trigger-terms field). For rules of class :type-prescription, variables are bound by the :typed-term field.

Let us discuss the method for relieving hypotheses of rewrite rules with free variables. Similar considerations apply to linear and forward-chaining rules, and type-prescription rules.

See free-variables-examples for more examples of how this all works, including illustration of how the user can exercise some control over it. In particular, see free-variables-examples-rewrite for an explanation of output from the break-rewrite facility in the presence of rewriting failures involving free variables, as well as an example exploring ``binding hypotheses'' as described below.

Note that the :match-free mechanism discussed below does not apply to type-prescription rules. See free-variables-type-prescription for a discussion of how to control free-variable matchng for type-prescription rules.

Some Related Topics

We begin with an example. Does the proof of the thm below succeed? 

(defstub p2 (x y) t)

(defaxiom p2-trans
  (implies (and (p2 x y)
                (p2 y z))
           (equal (p2 x z) t))
  :rule-classes ((:rewrite :match-free :all)))

(thm (implies (and (p2 a c)
                   (p2 a b)
                   (p2 c d))
              (p2 a d)))
Consider what happens when the proof of the thm is attempted. The ACL2 rewriter attempts to apply rule p2-trans to the conclusion, (p2 a d). So, it binds variables x and z from the left-hand side of the conclusion of p2-trans to terms a and d, respectively, and then attempts to relieve the hypotheses of p2-trans. The first hypothesis of p2-trans, (p2 x y), is considered first. Variable y is free in that hypothesis, i.e., it has not yet been bound. Since x is bound to a, the rewriter looks through the context for a binding of y such that (p2 a y) is true, and it so happens that it first finds the term (p2 a b), thus binding y to b. Now it goes on to the next hypothesis, (p2 y z). At this point y and z have already been bound to b and d; but (p2 b d) cannot be proved.

So, in order for the proof of the thm to succeed, the rewriter needs to backtrack and look for another way to instantiate the first hypothesis of p2-trans. Because :match-free :all has been specified, backtracking does take place. This time y is bound to c, and the subsequent instantiated hypothesis becomes (p2 c d), which is true. The application of rule (p2-trans) succeeds and the theorem is proved.

If instead :match-free :all had been replaced by :match-free :once in rule p2-trans, then backtracking would not occur, and the proof of the thm would fail.

Next we describe in detail the steps used by the rewriter in dealing with free variables.

ACL2 uses the following sequence of steps to relieve a hypothesis with free variables, except that steps (1) and (3) are skipped for :forward-chaining rules and step (3) is skipped for :type-prescription rules. First, if the hypothesis is of the form (force hyp0) or (case-split hyp0), then replace it with hyp0.

(1) Suppose the hypothesis has the form (equiv var term) where var is free and no variable of term is free, and either equiv is equal or else equiv is a known equivalence relation and term is a call of double-rewrite. We call this a ``binding hypothesis.'' Then bind var to the result of rewriting term in the current context.

(2) Look for a binding of the free variables of the hypothesis so that the corresponding instance of the hypothesis is known to be true in the current context.

(3) Search all enabled, hypothesis-free rewrite rules of the form (equiv lhs rhs), where lhs has no variables (other than those bound by let, let*, or lambda), rhs is known to be true in the current context, and equiv is typically equal but can be any equivalence relation appropriate for the current context (see congruence); then attempt to bind the free variables so that the instantiated hypothesis is lhs.

If all attempts fail and the original hypothesis is a call of force or case-split, where forcing is enabled (see force) then the hypothesis is relieved, but in the split-off goals, all free variables are bound to unusual names that call attention to this odd situation.

When a rewrite or linear rule has free variables in the hypotheses, the user generally needs to specify whether to consider only the first instance found in steps (2) and (3) above, or instead to consider them all. Below we discuss how to specify these two options as ``:once'' or ``:all'' (the default), respectively.

Is it better to specify :once or :all? We believe that :all is generally the better choice because of its greater power, provided the user does not introduce a large number of rules with free variables, which has been known to slow down the prover due to combinatorial explosion in the search (Steps (2) and (3) above).

Either way, it is good practice to put the ``more substantial'' hypotheses first, so that the most likely bindings of free variables will be found first (in the case of :all) or found at all (in the case of :once). For example, a rewrite rule like 

(implies (and (p1 x y)
              (p2 x y))
         (equal (bar x) (bar-prime x)))
may never succeed if p1 is nonrecursive and enabled, since we may well not find calls of p1 in the current context. If however p2 is disabled or recursive, then the above rule may apply if the two hypotheses are switched. For in that case, we can hope for a match of (p2 x y) in the current context that therefore binds x and y; then the rewriter's full power may be brought to bear to prove (p1 x y) for that x and y.

Moreover, the ordering of hypotheses can affect the efficiency of the rewriter. For example, the rule 

(implies (and (rationalp y)
              (foo x y))
         (equal (bar x) (bar-prime x)))
may well be sub-optimal. Presumably the intention is to rewrite (bar x) to (bar-prime x) in a context where (foo x y) is explicitly known to be true for some rational number y. But y will be bound first to the first term found in the current context that is known to represent a rational number. If the 100th such y that is found is the first one for which (foo x y) is known to be true, then wasted work will have been done on behalf of the first 99 such terms y -- unless :once has been specified, in which case the rule will simply fail after the first binding of y for which (rationalp y) is known to be true. Thus, a better form of the above rule is almost certainly the following. 
(implies (and (foo x y)
              (rationalp y))
         (equal (bar x) (bar-prime x)))

Specifying `once' or `all'.

As noted above, the following discussion applies only to rewrite, linear, and forward-chaining rules. See free-variables-type-prescription for a discussion of analogous considerations for type-prescription rules.

One method for specifying :once or :all for free-variable matching is to provide the :match-free field of the :rule-classes of the rule, for example, (:rewrite :match-free :all). See rule-classes. However, there are global events that can be used to specify :once or :all; see set-match-free-default and see add-match-free-override. Here are some examples. 

(set-match-free-default :once)    ; future rules without a :match-free field
                                  ; are stored as :match-free :once (but this
                                  ; behavior is local to a book)
(add-match-free-override :once t) ; existing rules are treated as
                                  ; :match-free :once regardless of their
                                  ; original :match-free fields
(add-match-free-override :once (:rewrite foo) (:rewrite bar . 2))
                                  ; the two indicated rules are treated as
                                  ; :match-free :once regardless of their
                                  ; original :match-free fields

Some history. Before Version 2.7 the ACL2 rewriter performed Step (2) above first. More significantly, it always acted as though :once had been specified. That is, if Step (2) did not apply, then the rewriter took the first binding it found using either Steps (1) or (3), in that order, and proceeded to relieve the remaining hypotheses without trying any other bindings of the free variables of that hypothesis.