I am interested in the following "improvement" of the Valiant-Vazirani reduction. As pointed out here, under the right derandomization assumptions one can obtain a deterministic polynomial-time algorithm $R$ such that $R(\varphi) = (\psi_1, \dots, \psi_k)$ with $k$ polynomial in the size of $\varphi$, such that
- if $\varphi$ is satisfiable, then at least one of the $\psi_i$ has exactly one satisfying assignment;
- if $\varphi$ is unsatisfiable, then every $\psi_i$ is unsatisfiable.
Looking at the existing proofs of the Valiant-Vazirani lemma (Arora-Barak, and some lecture notes online), they all use the pairwise independent hashing over either $GF(2)^n$ or $GF(2^n)$. My problem is that, as far as I can tell, in both of these setups when $\varphi$ is satisfiable, some of the formulas in $\psi_i$ might have multiple satisfying assignments -- it's just that one particular $\psi_i$ is guaranteed to have exactly one. I want to avoid this.
My question: Is it possible to obtain (possibly with a different family of hashing functions) a reduction where
- if $\varphi$ is satisfiable, then all the $\psi_i$ have at most one satisfying assignment, and at least one of the $\psi_i$ is satisfiable;
- if $\varphi$ is unsatisfiable, then every $\psi_i$ is unsatisfiable.
I know that different improvements on Valiant-Vazirani are unlikely, but this variation doesn't seem to be ruled out by those results, or at least I couldn't see that. Is it known whether this reduction is possible?