pub fn reduction_point<F: Clone>(
r_x_star: &[F],
log_zero_constraints: usize,
sample: impl FnMut() -> F,
) -> Vec<F>Expand description
Builds the constraint point the Zero reduction closes at.
The reduction runs directly after the BitAnd sumcheck and evaluates at the constraint half of
the point that sumcheck has just produced, r_x_star: rho is its
length-log_zero_constraints prefix. The sumcheck runs over the longest of the AND, IMUL and
BMUL arrays, so fresh challenges from sample extend r_x_star only when the ZERO array is
longer still. The prover and verifier derive it identically, so sample draws the same
challenges at the same point in both transcripts.
Asked for max(r_x_star.len(), log_zero_constraints) coordinates instead, it returns the
unified constraint point every operation is claimed at a prefix of, rho included. It draws the
same challenges either way.
sample stands in for the channel: the prover and verifier channel traits are unrelated, so one
function can only reach both through a closure.
Every coordinate of rho is uniform and independent — those inherited from the BitAnd sumcheck
are its per-round verifier challenges — so the soundness bound is plain Schwartz-Zippel on a
polynomial the witness commitment pinned down before the first challenge was drawn. The
reduction’s input zerocheck challenges would serve too, but they open with the deterministic
Rijndael coordinates, which carry no randomness and would need absorbing by a separate
F_2-independence argument first.
Sharing the point with the BitAnd reduction is sound despite the resulting correlation: the Zero bound appeals only to the marginal distribution of the challenges it consumes, which reuse leaves untouched. What it does require is that every challenge be drawn after the witness is committed, which the phase ordering guarantees.