You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Justify indifferentiability of mapHashOutputsToCurve from an RO wrt the hash_to_field outputs #25
That is, prove the concrete indifferentiability relation $\mathsf{mapHashOutputsToCurve} \circ \mathsf{hash\_to\_field}(·, 2) {\large ⊏}_c \mathsf{hash\_to\_field}(·, 2)$, where $\mathsf{hash\_to\_field}(·, 2) ⦂ M \rightarrow (\mathbb{F}_{q_{\mathbb{G}}})^2$ for $M$ a suitable input type (domain separator and byte sequence), and $\mathsf{mapHashOutputsToCurve}$ as formalized in #19. This follows Brier et al.'s Theorem 1; it will require the form of the Weil bound defined in Curves.Hashing.WellDistributed as a named hypothesis, and consume TwoTermUniformity as the regularity half.
Re-scoped after PR #19 merged: the proof is split across the two repos so that CompElliptic contains no probability and ironwood contains no new counting. The game-theoretic side — the simulator, the single-query bias lemma, the adaptive-hybrid headline over OracleComp query trees, the costed simulator, and the book page — now lives at zcash/ironwood#198, built on ironwood's existing oracle machinery.
This issue now tracks the counting-side deliverables, in this repo's exact-Finset register:
a per-point fibre bound for the single-term mapToCurve (the isogeny is injective on rational points, so this is a bound on SSWU fibres: per target abscissa, a linear and a cubic branch equation in t = Z·u² give at most 4 values of t, each with at most two roots u, and oddness pairs ±u between P and -P);
the unsquared L¹ corollary of sq_sum_abs_prob_dev_le, with denominator-cleared exact forms convenient for ℝ≥0∞ consumption, and the pre-Cauchy–Schwarz L² bound (sum_sq_dev_le) exported for bad-set counting (regularity does not lower-bound individual fibres, so the sampler's acceptance constant holds only outside a small bad set);
the deployed-vs-zero-repaired transport as pure pair counting (the two two-term sums differ on at most 2·#F - 1 of (#F)² pairs).
These ship as one PR against main, iterated in parallel with the ironwood side so that the exported interface is confirmed sufficient before it merges.
The CompElliptic side of this landed in #27 (fibre counting and the two-term regularity budget) and #29 (the Weil-bound reduction discharging WeilBounded at the deployed mappings). The indifferentiability endpoints that consume them — pallas_indiffFromRO/vesta_indiffFromRO, at advantage q/2^{120} after q queries with hash_to_field modelled as a random oracle — are on zcash/ironwood#199.
Activity
feat(Hashing): mapHashOutputsToCurve, the deployed construction after…
docs(Hashing): sharpen the analysis-boundary claims
daira commentedon Aug 13, 2026
Re-scoped after PR #19 merged: the proof is split across the two repos so that CompElliptic contains no probability and ironwood contains no new counting. The game-theoretic side — the simulator, the single-query bias lemma, the adaptive-hybrid headline over
OracleCompquery trees, the costed simulator, and the book page — now lives at zcash/ironwood#198, built on ironwood's existing oracle machinery.This issue now tracks the counting-side deliverables, in this repo's exact-
Finsetregister:mapToCurve(the isogeny is injective on rational points, so this is a bound on SSWU fibres: per target abscissa, a linear and a cubic branch equation int = Z·u²give at most 4 values oft, each with at most two rootsu, and oddness pairs±ubetweenPand-P);sq_sum_abs_prob_dev_le, with denominator-cleared exact forms convenient forℝ≥0∞consumption, and the pre-Cauchy–Schwarz L² bound (sum_sq_dev_le) exported for bad-set counting (regularity does not lower-bound individual fibres, so the sampler's acceptance constant holds only outside a small bad set);2·#F - 1of(#F)²pairs).These ship as one PR against main, iterated in parallel with the ironwood side so that the exported interface is confirmed sufficient before it merges.
Posted with the assistance of Claude Fable 5
daira commentedon Aug 20, 2026
The CompElliptic side of this landed in #27 (fibre counting and the two-term regularity budget) and #29 (the Weil-bound reduction discharging
WeilBoundedat the deployed mappings). The indifferentiability endpoints that consume them —pallas_indiffFromRO/vesta_indiffFromRO, at advantage q/2^{120} after q queries withhash_to_fieldmodelled as a random oracle — are on zcash/ironwood#199.Posted with the assistance of Claude Fable 5