Skip to content

Execute fine-prime Gaussian multiplication from binary operands - #7834

Open
aliromman wants to merge 2 commits into
CrocSwap:mainfrom
aliromman:codex/fine-prime-product
Open

aliromman wants to merge 2 commits into
CrocSwap:mainfrom
aliromman:codex/fine-prime-product

Conversation

@aliromman

@aliromman aliromman commented Nov 20, 2026 •

Copy link
Copy Markdown
Contributor

This construction takes two unsigned binary strings, generates and decodes its Gaussian coefficient catalogue, and returns their exact binary product. The previous physical sandwich returned zero on its diagnostic packet and rejected it in production. The new path passes its own admission test and executes both CRT ingresses, two forward transforms, the pointwise product, the opposite transform, original-coordinate recovery, inverse CRT and carry. Product implementation · All-size proof and cost · Execution evidence

Numerical and movement result. The producer uses short gauged circulants on dyadic blocks. It stores the exceptional rows at holes and block edges, and applies the inverse correction with a bordered Schur solve on sequential tapes. The reference matrix agrees exactly with the rounded block factors away from those rows. At the original fine α, the old concentration test rejects every precision for sufficiently large p₀. Even after retuning α, it requires more than (p₀²/4)(1+2 log₂ p₀) fractional bits. The new producer bounds absolute operator error. Its dense butterflies use only block-local bits, and only exceptional rows require long axis movement. Both source-size factors and all three dyadic normalizers are restored before rounding. Coefficient proof · Sparse inverse and movement · Error through rounding · Normalized polynomial convolution

For arbitrary operands the fine controller uses p₀ = 6 ceil(log₂ n), an external grid of 8p₀ fractional bits and O(p₀)-bit full-field components. It constructs all Gaussian setup in 2^o(p₀) bit operations. The proof admits all sufficiently large lengths and gives O(n) backing plus a fixed source constant. Its largest log-length power before polylogarithmic factors is 0.998268211. This fits the complete fine-prime target of @1337isnot1337 at κ = 0.00173097911567079247276057. The selected conditional κ = 0.00363005817 is stronger. PR #4100 · Selected record · Physical ordinary interchange · Grouped butterflies

Executed Gaussian product Source → target Output Pre-carry checks Uniform radius
741503 × 466013 5 × 7 → 8 × 8 345550037539 35 < 2⁻⁶³
Same operands, padded to 145 bits 31 × 29 → 32 × 32 345550037539 899 < 0.000247
Two distinct 1,800-bit operands 31 × 29 → 32 × 32 3,600 bits 899, all positive < 0.000247

The first run recovered 11811, 10996, 4152, 868, 77 followed by 30 zeroes. The independent dense Gaussian implementation by @mason-hale agreed at every coordinate. The second run exercised two strict block interiors and a different digit width. At the 1,800-bit endpoint, all 899 nonzero coefficients and all 3,600 output bits agreed with independent integer computations. Its largest observed residual was below 4.14 × 10⁻⁹. Each run used the Gaussian path for the complete product. Full-size operands, reproduction and digest Adversarial review also found that the original fitted high-field exchange exhausted 512 scratch roles at the actual source prime. The corrected primitive reaches the same payload read with nine roles. Reduced complete root calls verify the permutations and restoration with and without a remainder. Product verifier · Interior product verifier · Physical negative control

Dependencies and limits. The all-size comparison uses the periodic Gaussian identity, Trudgian’s prime-count estimate, the all-column dyadic word existence result, the compact436 logical endpoint and stock/refinement bounds, and the physical short-word FFT theorem with its numerical proof. The finite product runs use bounded scalar and Walsh branches. The complete production-prime GL₆₄ orbit and eventual funded C word exceed practical replay sizes. Independent scientific review of the combined theorem remains open. Source bindings and credit · Input hashes

Predecessors. Harvey and van der Hoeven supplied periodic resampling. @JoyboyBrian and @tekuuu supplied the physical sandwich and CRT lineage. @petalas supplied the arbitrary-bit boundary. @CubeJerry supplied the cached chirp, deferred recovery, original-coordinate extractor and expansion analysis. @jmbarrancoml supplied physical word and ordinary implementations. @ai-nanachi supplied compact436. The funded C-word construction is due to @zxv and its credited integer-flow predecessors. The asymmetric error and two-target CRT proposals enter the comparison. The base is canonical main at e545235. Prepared for @aliromman with substantial OpenAI Codex assistance. PR #4024 · PR #5095 · PR #2313 · PR #6185 · PR #6982 · PR #6987 · PR #6991 · PR #6993 · PR #6995 · PR #3787 · PR #2899 · PR #4326 · PR #4359 · PR #7267 · PR #7390

Reproduce with Python 3.12 and the standard library from the repository root.

python3.12 -I -B research/fine-prime-gaussian-product/product.py 10110101000001111111 1110001110001011101
python3.12 -B scripts/research_package.py --package research/fine-prime-gaussian-product

The first command prints the exact 40-bit result. The last line of the registered replay follows it below.

0101000001110100011000110010001000100011
PASS_REGISTERED_RESEARCH_PACKAGE research/fine-prime-gaussian-product outcome=completed

All eight registered commands passed locally and from the read-only Git-free export. The export included all 151 declared external input files. Committed byte, executable mode and source-history checks passed. Replay manifest · Coordination and review scope

GitHub’s hosted planner ran 760 onboarding tests. Its sole failure also reproduces on the canonical base. The CI guide still names result 6030, while the selected record requires make frontier7000-verify. GitHub therefore skipped the package job. Hosted planner · Canonical CI guide · Existing repair in PR #7276

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.

1 participant