Repository navigation
feat(Probability): Holenstein's Corollary 6, conditioning with side information (#886) - #955
Conversation
|
Review claimed by reviewer-claude-c4b0c164-01 (agent review) at d57cb49. |
|
reviewer-claude-c4b0c164-01 — APPROVED at d57cb49 Agent review, not human approval. I reviewed this PR's own payload over #942's Against Holenstein (ToC 2009), Corollary 6.
Nonblocking:
|
8ab69f2 to
5122f72
Compare
d57cb49 to
1de301e
Compare
|
|
1de301e to
3b85fe2
Compare
…nformation (#886) Extend Raz's conditioning lemma to events involving side information, and prove Holenstein's Corollary 6: with coordinates independent given t and side information v, sum_j ||P_{TVU_j|W} - P_{TV|W} P_{U_j|T}|| is at most sqrt(k ln(|V|/Pr[W]) / 2). Co-authored-by: Jetski (worker-jetski-17d3679e) <noreply@google.com>
… issue (#950) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dvi6MU29zU9t8zYfeU51sB
3b85fe2 to
8cd877a
Compare
|
Restack ( coordinator-claude-c4b0c164-20261023 — APPROVED at 8cd877a (restack; reviewer-claude-c4b0c164-01's approval at |
Part of #886 (M2b). Stacked on #942 (which is on #940); retarget to
mainonce those land.This PR proves Holenstein's Corollary 6 [Holenstein2009]: the conditioning bound that Lemma 14 applies to select a good coordinate. It is in
Complexity/Probability/ProductConditioning/SideInformation.lean.The statement. The coordinates of
uneed only be independent given somet, and side informationvmay depend ontandu.sideJoint μ Jis the law of((t, v), u), witht ~ μ,(u, v) ~ J tand(J t).map Prod.fst = pi (K · t).sum_dist_filter_sideJoint_le: conditioning on any eventWgives∑ⱼ ‖P_{TV Uⱼ | W} − P_{TV | W} P_{Uⱼ | T}‖ ≤ √(k · ln(|V| / Pr[W]) / 2).The steps, following Holenstein:
sum_sq_dist_map_filter_leis Raz's conditioning lemma with side information. The first marginal of the joint is a product andWis an event about the pair. The conditioned marginal is dominated bypi p / Pr[W](sum_mul_log_div_le_log_of_le), so its divergence is at mostln(1/Pr[W]); superadditivity and Pinsker then finish the proof as in feat(Probability): Raz's conditioning lemma for product distributions (#886) #942.dist_map_pair_bind_eq: revealing(t, v)turns each distance into theP_{TV|W}-average of conditional distances.conditionalSnd_filter_sideJointandmap_fst_filter_sideJoint_toReal: given(t, v), the conditioned law ofuisJ tconditioned on the slicesliceEvent W t v. The slice has probabilityμ(t)·Pr_{J t}[W_{t,v}]/Pr[W]under the conditioned joint.√andlncombines the slices. The weights satisfy∑ P_{TV|W}(t,v)/Pr_{J t}[W_{t,v}] ≤ |V|/Pr[W].Faithfulness.
|V*|, the support ofvunderW. This statement uses the full alphabet|V|, which is weaker in general.1/2is stronger than Holenstein'slog(·), sinceln x/2 ≤ log₂ x.|V*| ≤ |V| = (|A||B|)^{n−k}, so the|V|form suffices there.Not in this PR. Fact 11, Claims 12–13, Lemma 10, Corollary 9, Lemmas 14–15 and Theorem 4 are still open on #886.
Tests.
ComplexityTests/SideInformation.leanchecks the trivial-T, trivial-Vinstance (Raz's setting) and adds axiom guards for the two main results. The ROADMAP clause is updated.Validation.
scripts/check.shpasses all six gates locally at d57cb49, andcheck_hygiene.py --check-simppasses.AI disclosure: written by Jetski, Google's AI coding agent (session
worker-jetski-17d3679e). Samuel has not reviewed it yet.