Skip to content

Research: general odd-kernel exclusion pattern for native DATA shear pairs (no new kappa) - #2687

Merged
Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
VictorLiwentao:cursor/odd-kernel-query-pair-60b9
Oct 16, 2026
Merged

Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
VictorLiwentao:cursor/odd-kernel-query-pair-60b9

Conversation

@VictorLiwentao

Copy link
Copy Markdown
Contributor

Change and scope

Completes open question 2 of #2593 (research/triadic-query-pair, kernel-3 specialization of #840): for every positive odd kernel q, the same three-query fusion at a=±q·2^j attains the typed invoice below on every odd modulus coprime to q. Even moduli and odd moduli sharing a prime with q are excluded. Typed costs do not improve over the dyadic (q=1) case.

x <- h^-a x
x ^= y
x <- h^(2a) x
x ^= y
x <- h^-a x
Complete cost item Separate pair Fused odd-kernel pair
Ordinary recursive children 4 3
Ordered affine events 24 18
Aligned XOR scans 4 2
Extra full private rows 0 0

Certificate kernels: {1,3,5,7,9,15} (dyadic recovery, triadic cross-check, and primary new {5,7,9,15}). Pattern: admitted ⇔ odd and gcd(modulus,q)=1.

Research construction / recursive interface; no new kappa. Leave-alone through #2686 (including #2677, #2643, #2593) was not reopened.

Result

Package: research/odd-kernel-query-pair/.
Tip: c319093b5b96211e89165f4ad9dd0d20d5039414.

Exact invoice children 4→3; affine 24→18; XOR 4→2 uniformly across certified odd kernels; no typed gain over dyadic.

Validation

python3.12 -I -B research/odd-kernel-query-pair/verify.py
python3.12 -I -O -B research/odd-kernel-query-pair/verify.py
python3.12 -I -B research/odd-kernel-query-pair/independent.py
python3.12 -I -B research/odd-kernel-query-pair/test_package.py
python3.12 -B scripts/research_package.py --package research/odd-kernel-query-pair

Expected:

PASS_ODD_KERNEL_QUERY_PAIR kernels=[1, 3, 5, 7, 9, 15] primary_new=[5, 7, 9, 15] children=4->3 affine=24->18 xor=4->2 no_typed_gain_over_dyadic ... no_new_kappa
PASS_INDEPENDENT_ODD_KERNEL_QUERY_PAIR kernels=[1, 3, 5, 7, 9, 15] ... no_new_kappa
PASS_ODD_KERNEL_QUERY_PAIR_PACKAGE_TESTS groups=6
PASS_REGISTERED_RESEARCH_PACKAGE research/odd-kernel-query-pair outcome=completed

Default verify is read-only. Admission uses need(); -O also passes.

Overlap and credit

Closest priors: #840 dyadic; #2593 kernel-3 (OQ2). Freestanding generalization; does not edit those packages. No competing open PR on this multi-kernel pattern at refresh.

Credit Wentao Li / PR840 / #2593 lineage. Prepared by Wentao Li (VictorLiwentao, @Is__Victor) with substantial AI assistance, under Apache-2.0.

…pairs

Answer triadic OQ2 / PR840 non-dyadic note: for every odd kernel q the fused
three-query shear at a=±q·2^j admits odd moduli coprime to q with typed costs
4→3 / 24→18 / 4→2 and no improvement over dyadic. Certificate kernels
{1,3,5,7,9,15}. No new kappa.

Prepared by Wentao Li with substantial AI assistance.
@VictorLiwentao

Copy link
Copy Markdown
Contributor Author

OQ1 follow-up attempt (deferred; OQ1 still open)

A candidate package research/odd-kernel-source-return-host/ @ db3deecf6 named an immutable-source / exclusive-target “admitted-screen” host bill on top of the #3702 coeff-bit T_word: I=0 recovers zero odd fused beats on j=0..32; I=1 yields a unique beat at (q=3,j=0) (674<677) with critical I*=14/17.

Soft-overlap / weak-need deferred: the odd-vs-dyadic comparison is ΔT + I·ΔA, so fixed units C,X,D,R_SR cancel from that order; R_SR never binds the proved inequalities; screen ODD_MODULI_3_99 / unit routing I are author-chosen (near-minimal upper bound with I*<1). This is not a grounded admitted-BIT-supplier hosting lemma with a pinned source-return boundary that charges routing/precision/dispatch non-cancellably.

No separate CrocSwap PR opened. OQ1 remains open.

@VictorLiwentao

Copy link
Copy Markdown
Contributor Author

Odd-kernel run-batching attempt (deferred; OQ1 still open)

A candidate package research/odd-kernel-query-runs/ @ 327262ebe on branch cursor/odd-kernel-query-runs-60b9 lifts the PR953 dyadic-run schedule through this PR’s signed q-dyadic admission: for certified kernels and all 63 nonempty J⊆{0,…,5}, the batch walk attains children 2r+c / affine 6(2r+c) / XOR 2r with no typed gain over dyadic.

Soft-overlap / weak-need deferred: soft-overlap ≈0.90–0.95 vs PR953 (same schedule/bound/seed-55, powers×q) and ≈0.55–0.70 vs this PR’s admission census; combined compose-and-retarget ≈0.9. The 2r+c invoice is forced once the dyadic-run walk is scaled into the odd-kernel alphabet; machine “lower bound” equals the constructed program. Does not answer OQ1 (admitted BIT-supplier / non-cancellable source-return bill).

No separate CrocSwap PR opened. Prefer a grounded OQ1 hosting lemma or a disjoint non-odd-kernel-run card.

Mister-Meeseeks pushed a commit that referenced this pull request Oct 20, 2026
… (answers #2687 OQ2; no new kappa)

Attach a named ADDRESS affine coefficient-bit tariff to the PR840/odd-kernel
address_word lowering. Every certified odd kernel strictly exceeds the dyadic
fused and baseline bills on j=0..32 (165+165), so larger-q tables never pay
for themselves despite identical typed child/XOR counts.

Prepared by Wentao Li with substantial AI assistance.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants