BACKCHAIN-LIMIT

limiting the effort expended on relieving hypotheses
Major Section:  MISCELLANEOUS

Before ACL2 can apply a rule with hypotheses, it must establish that the hypotheses are true. (We ignore the relaxing of this requirement afforded by case-splits and forced hypotheses.) ACL2 typically establishes each hypothesis by backchaining -- instantiating the hypothesis and then rewriting it recursively. Here we describe how ACL2 allows the user to limit backchaining. At the end, below, we describe the function backchain-limit.

Each hypothesis of a rewrite, meta, linear, or type-prescription rule is assigned a backchain-limit when the rule is stored. By default, this limit is nil, denoting infinity (no limit). However, the value used for the default may be set to a non-negative integer (or to nil) by the user; see set-default-backchain-limit. The default is overridden when a :backchain-limit-lst is supplied explicitly with the rule; see rule-classes. The number of recursive applications of backchaining starting with the hypothesis of a rule is limited to the backchain-limit associated with that hypothesis.

Moreover, the user may set global backchain-limits that limit the total backchaining depth. See set-backchain-limit. One limit is for the use of rewrite, meta, and linear rules, while the other limit is for so-called ``type-set reasoning'', which uses rules of class type-prescription rules. The two limits operate independently. Below, we discuss the first kind of backchain limits, i.e., for other than type-prescription rules, except as otherwise indicated; but the mechanism for those rules is similar.

Below we lay out the precise sense in which a global backchain-limit interacts with the backchain-limits of individual rules in order to limit backchaining. But first we note that when further backchaining is disallowed, ACL2 can still prove a hypothesis in a given context by using that contextual information. In fact, type-set reasoning may be used (except that a weaker version of it is used in the second case above, i.e., where we are already doing type-set reasoning). Thus, the relieving of hypotheses may be limited to the use of contextual information (without backchaining, i.e., without recursively rewriting hypotheses) by executing :set-backchain-limit 0.

Recall that there are two sorts of backchain limits: those applied to hypotheses of individual rules, as assigned by their :rule-classes or else taken from the default (see set-default-backchain-limit); and the global limit, initially nil (no limit) but settable with :set-backchain-limit. Here is how these two types of limits interact to limit backchaining, i.e., recursive rewriting of hypotheses. ACL2 maintains a current backchain limit, which is the limit on the depth of recursive calls to the rewriter, as well as a current backchain depth, which is initially 0 and is incremented each time ACL2 backchains (and is decremented when a backchain completes). When ACL2 begins to rewrite a literal (crudely, one of the ``top-level'' terms of the goal currently being worked on), it sets the current backchain-limit to the global value, which is initially nil but can be set using :set-backchain-limit. When ACL2 is preparing to relieve a hypothesis by backchaining (hence, after it has already tried type-set reasoning), it first makes sure that the current backchain limit is greater than the current backchain depth. If not, then it refuses to relieve that hypothesis. Otherwise, it increments the current backchain depth and calculates a new current backchain-limit by taking the minimum of two values: the existing current backchain-limit, and the sum of the current backchain depth and the backchain-limit associated with the hypothesis. Thus, ACL2 only modifies the current backchain-limit if it is necessary to decrease that limit in order to respect the backchain limit associated with the hypothesis.

We illustrate with the following examples.

; We stub out some functions so that we can reason about them.

(defstub p0 (x) t)
(defstub p1 (x) t)
(defstub p2 (x) t)
(defstub p3 (x) t)

; Initially, the default-backchain-limit is nil, or infinite.

(defaxiom p2-implies-p1-limitless
  (implies (p2 x)
           (p1 x)))

; The following rule will have a backchain-limit of 0.

(defaxiom p1-implies-p0-limit-0
  (implies (p1 x)
           (p0 x))
  :rule-classes ((:rewrite :backchain-limit-lst 0)))

; We have (p2 x) ==> (p1 x) ==> (p0 x).  We wish to establish that
; (p2 x) ==> (p0 x).  Normally, this would be no problem, but here
; we fail because ACL2 cannot establish (p0 x) by type-set reasoning
; alone.

(thm
  (implies (p2 x)
           (p0 x)))

; We set the default-backchain-limit (for rewriting) to 1.

:set-default-backchain-limit 1

; The following is more powerful than p1-implies-p0-limit-0
; because it can use rewrite rules to establish (p1 x).

(defaxiom p1-implies-p0-limit-1
  (implies (p1 x)
           (p0 x)))

; This theorem will succeed:

(thm
  (implies (p2 x)
           (p0 x)))

; We return the default-backchain-limit to its initial value.

:set-default-backchain-limit nil

; Here is our last axiom.

(defaxiom p3-implies-p2-limitless
  (implies (p3 x)
           (p2 x)))

; We now have (p3 x) ==> (p2 x) ==> (p1 x) ==> (p0 x).  However the
; rule p1-implies-p0-limit-1 has a backchain-limit of 1; hence we
; are not allowed to backchain far enough back to use
; p3-implies-p2-limitless.  We therefore lose.

(defthm will-fail
  (implies (p3 x)
           (p0 x)))

Finally, we remark that to see the current global backchain-limits, issue the following commands.

(backchain-limit wrld :ts) ; backchain limit for type-set reasoning
(backchain-limit wrld :rewrite) ; backchain limit for rewriting