Soundness of answer reduction: from evaluations to polynomials #
Piece AR-5d of planning/answer-reduction.md, the Schwartz--Zippel step of lem:ar-ar
(eq:ld-strat-con-def): the evaluated relations of SoundRelations between the sixth copy's
J and copy i's G become relations between their polynomial outcomes. Copy i's polynomial,
in m variables, is placed on its block of the sixth copy's m' variables
(LowIndDegPoly.liftIdx along blockEmb i); two distinct polynomials in m' variables of
individual degree d agree at a uniform point with probability at most m' d / q
(sum_uniform_eval_eq_le), so the polynomial outcomes of J's i-th component and of the placed
G disagree with probability at most their evaluated disagreement plus m' d / q
(disPolyA_le, and its mirror image), in the model (BipartiteModel.dis_le_sum_dis_map_add).
This replaces the paper's giant sandwich Λ of claim:ar-5: the oracle's measurement here is
J itself, whose components are placed polynomials except with the probability this file bounds.
Schwartz--Zippel for coefficient vectors: two distinct polynomials of individual degree d
in n variables agree at a uniform point with probability at most n d / q.
A polynomial of copy i, placed on its block of the sixth copy's variables.
Equations
Instances For
The error of the Schwartz--Zippel step, m' d / q.
Equations
- MIPRE.AnswerReduction.errSZ P hk = ↑P.m' * ↑MIPRE.AnswerReduction.dPcp / ↑(Fintype.card (MIPRE.AnswerReduction.Fq P hk))
Instances For
The multiplicity of the sixth copy's point: a uniform PCP vector carries a uniform point, so averaging over one is averaging over the other.
The outcome-level disagreement of Alice's J, component i, against Bob's G of copy i
placed on its block, at oracle half x.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The mirror image: Alice's G of copy i placed on its block against Bob's J, component
i.
Equations
- One or more equations did not get rendered due to their size.
Instances For
From evaluations to polynomials, Alice's J: at each oracle half, the outcome-level
disagreement is at most the evaluated one, averaged over the PCP vectors, plus m' d / q.
From evaluations to polynomials, Bob's J.
Summed over the oracle halves, Alice's J against Bob's placed G.
Summed over the oracle halves, Alice's placed G against Bob's J.