Converting first-order logic to CNF without exponential blowup - first-order-logic

Question

When attempting to solve logic problems on a computer, it is usual to first convert them to CNF, because the best solving algorithms expect CNF as input.
For propositional logic, the textbook rules for this conversion are simple, but if you apply them as is, the result is one of the very rare cases where a program encounters double exponential resource consumption without being specifically constructed to do so:
a <=> (b <=> (c <=> ...))
with N variables, generates 2^2^N clauses, one exponential blowup in the conversion of equivalence to AND/OR, and another in the distribution of OR into AND.
The solution to this is to rename subterms. If we rewrite the above as something like
r <=> (c <=> ...)
a <=> (b <=> r)
where r is a fresh symbol that is being defined to be equal to a subterm - in general, we may need O(N) such symbols - the exponential blowups can be avoided.
Unfortunately, this runs into a problem when we try to extend it to first-order logic. Using TPTP notation where ? means 'there exists' and variables begin with capital letters, consider
a <=> ?[X]:p(X)
Admittedly this case is simple enough that there is no actual need to rename the subterm, but it's necessary to use a simple case for illustration, so suppose we are using an algorithm that just automatically renames arguments of the equivalence operator; the point generalizes to more complex cases.
If we try the above trick and rename the ? subterm, we get
r <=> ?[X]:p(X)
Existential variables are converted to Skolem symbols, so that ends up as
r <=> p(s)
The original formula then expands to
(~a | r) & (a | ~r)
Which is by construction equivalent to
(~a | p(s)) & (a | ~p(s))
But this is not correct! Suppose we had not done the renaming, but just expanded the original formula as it was, we would get
(~a | ?[X]:p(X)) & (a | ~?[X]:p(X))
(~a | ?[X]:p(X)) & (a | ![X]:~p(X))
(~a | p(s)) & (a | ~p(X))
which is critically different from the version we got with the renaming.
The problem is that equivalence needs both the positive and negative versions of each argument, but applying negation to terms that contain universal or existential quantifiers, structurally changes those terms; you cannot just encapsulate them in a definition, then apply the negation to the defined symbol.
The upshot of this is that when you have equivalence and the arguments may contain such quantifiers, you actually need to recur through each argument twice, once for the positive version, once for the negative. This suffices to bring back the existential blowup we hoped to avoid by doing the renaming. As far as I can see, this problem is not caused by the way a particular algorithm works, but by the nature of the task.
So my question:
Given an input formula that may contain arbitrary nesting of equivalence and quantifiers, is there any algorithm that will correctly turn this to CNF with a polynomial rather than exponential number of clauses?

Answer 1

As you observed, an existential such as ∃X.p(X) is not in fact equivalent to a Skolemized expression p(S). Its negation ¬∃X.p(X) is not equivalent to ¬p(S), but to ∀Y.¬p(Y).
Possible approaches that avoid the exponential blow-up:
Convert existentials such as ∃X.p(X) to universals such as ¬∀Y.p(Y), or vice versa, so you have a canonical form. Skolemize at a later step.
Remember when you convert that your p(S) is a Skolemized existential, and that its negation is ∀Y.¬p(Y).
Define terms equivalent to universals and existentials, such that a represents ∀Y.p(Y) and ¬a then represents ¬∀Y.p(Y), or equivalently, ∃X.¬p(X).
Use the symmetry of Boolean duals, so that the same transformations apply with AND and OR swapped, De Morgan’s Laws, and the equivalence between existentials and negated universals, to restore the symmetry between the expansions of r and ~r. The negations in the conversion between universals and existentials and in De Morgan's Laws cancel each other out, and the duality of switching AND and OR means you can re-use the result on the left to generate the one on the right mechanically again?
Given that you need to support ALL and NOT ALL statements anyway, this should not create any new problems. Just canonicalize and use the same approach you would for a universal.
If you’re solving by converting to SAT, your terms can represent universals, too. So, in your example, you’re trying to replace a with r, but you can still use ~a, equivalent to the negative universal.
In your expressions. you’d still use (~a | r) & (a | ~r), but expand ~r to its correct rather than the incorrect value. That example is trivial, since that’s just ~a, but you’d normally define r as equivalent to a more complex transformation, and in that case you need to remember what both r and ~r represent. It is not really a simple mechanical transformation of the Skolemized expression.
In this example, I’m not sure why it’s a problem that (~a | r) & (a | ~r) is equivalent to (~a | r) & (a | ~a), which simplifies to (~a | r). That’s not going to give you exponential blow-up? When you translate back to first-order predicate logic, make the correct translation.
Update
Thanks for clarifying what the problem was in chat. As I currently think I understand it, what you have is an equivalence with a left and a right side, which contains other nested equivalences, and you want to expand both the equivalence and its negation. The problem is that, because the negation does not have symmetrical form, you need to recurse twice for each nested equivalence in the tree, once when expanding the equivalence and once when expanding its negation?
You should define a transformation that generates the negative expansion from the positive expansion in linear time, and divide-and-conquer the expressions containing nested equivalences using that. This seems to be what you were after with the ~p(S) transformation.
To do this, you recall that ¬∃X.p(X) is equivalent to ∀X.¬p(X), and vice versa. Then if you’ve expanded p(x) into normal form as conjunctions and disjunctions, De Morgan’s Laws lets you turn an expression like ¬(a ∨ ¬b) into ¬a ∧ b. The inner ¬ on the quantifier transformation and the outer ¬ on the De Morgan transformation cancel each other out. Finally, the dual of any Boolean equivalence remains valid when you replace each ∨ and ∧ with the other and any atom a or ¬a with its inverse.
So, while I might be making an error, especially at 1 AM, it looks to me like what you want is the dual transformation that substitutes:
An outer ∃ for ∀ and vice versa
∧ for ∨ and vice versa
Each term t with ¬t and vice versa
Apply this to the expansion of the positive equivalence to generate the negative dual in time proportional to its length, without further recursion.