Major Section: MISCELLANEOUS
Occasionally the ACL2 theorem prover reports that the current goal simplifies to itself or to a set including itself. Such simplifications are said to be ``specious'' and are ignored in the sense that the theorem prover acts as though no simplification were possible and tries the next available proof technique. Specious simplifications are almost always caused by forcing.
The simplification of a formula proceeds primarily by the local
type-prescription, and other rules to its
various subterms. If no rewrite rules apply, the formula cannot be
simplified and is passed to the next ACL2 proof technique, which is
generally the elimination of destructors. The experienced ACL2 user
pays special attention to such ``maximally simplified'' formulas;
the presence of unexpected terms in them indicates the need for
additional rules or the presence of some conflict that prevents
existing rules from working harmoniously together.
However, consider the following interesting possibility: local
rewrite rules apply but, when applied, reproduce the goal as one of
its own subgoals. How can rewrite rules apply and reproduce the
goal? Of course, one way is for one rule application to undo the
effect of another, as when commutativity is applied twice in
succession to the same term. Another kind of example is when two rules
conflict and undermine each other. For example, under suitable
(length x) might be rewritten to
(+ 1 (length (cdr x)))
length and then a
rewrite rule might be used
to ``fold'' that back to
(length x). Generally speaking the
presence of such ``looping'' rewrite rules causes ACL2's simplifier
either to stop gracefully because of heuristics such as that
described in the documentation for
loop-stopper or to cause a
stack overflow because of indefinite recursion.
A more insidious kind of loop can be imagined: two rewrites in different parts of the formula undo each other's effects ``at a distance,'' that is, without ever being applied to one another's output. For example, perhaps the first hypothesis of the formula is simplified to the second, but then the second is simplified to the first, so that the end result is a formula propositionally equivalent to the original one but with the two hypotheses commuted. This is thought to be impossible unless forcing or case-splitting occurs, but if those features are exploited (see force and see case-split) it can be made to happen relatively easily.
Here is a simple example. Declare
foo to be a function of one
argument returning one result:
(defstub p1 (x) t)Prove the following silly rule:
(defthm bad (implies (case-split (p1 x)) (p1 x)))Now suppose we try the following.
(thm (p1 x))The rewrite rule
t, but it will be unable to prove the hypothesis
(case-split (p1 x)). As a result, the prover will spawn a new goal, to prove
(p1 x). However, since this new goal is the same as the original goal, ACL2 will recognize the simplification as specious and consider the attempted simplification to have failed.
In other words, despite the rewriting, no progress was made! In more common cases, the original goal may simplify to a set of subgoals, one of which includes the original goal.
If ACL2 were to adopt the new set of subgoals, it would loop indefinitely. Therefore, it checks whether the input goal is a member of the output subgoals. If so, it announces that the simplification is ``specious'' and pretends that no simplification occurred.
``Maximally simplified'' formulas that produce specious simplifications are maximally simplified in a very technical sense: were ACL2 to apply every applicable rule to them, no progress would be made. Since ACL2 can only apply every applicable rule, it cannot make further progress with the formula. But the informed user can perhaps identify some rule that should not be applied and make it inapplicable by disabling it, allowing the simplifier to apply all the others and thus make progress.
When specious simplifications are a problem it might be helpful to disable all forcing (including case-splits) and resubmit the formula to observe whether forcing is involved in the loop or not. See force. The commands
ACL2 !>:disable-forcing and ACL2 !>:enable-forcingdisable and enable the pragmatic effects of both
case-split. If the loop is broken when forcing is disabled, then it is very likely some forced hypothesis of some rule is ``undoing'' a prior simplification. The most common cause of this is when we force a hypothesis that is actually false but whose falsity is somehow temporarily hidden (more below). To find the offending rule, compare the specious simplification with its non-specious counterpart and look for rules that were speciously applied that are not applied in the non-specious case. Most likely you will find at least one such rule and it will have a
forced hypothesis. By disabling that rule, at least for the subgoal in question, you may allow the simplifier to make progress on the subgoal.