Skip to content

Prove sharp carryless channel widths for an ordinary product bit - #7838

Open
Clear1nsight wants to merge 1 commit into
CrocSwap:mainfrom
Clear1nsight:codex/carryless-middle-bit-rank
Open

Clear1nsight wants to merge 1 commit into
CrocSwap:mainfrom
Clear1nsight:codex/carryless-middle-bit-rank

Conversation

@Clear1nsight

Copy link
Copy Markdown
Contributor

Change and mathematical scope

For separate encodings of two m-bit operands, recover bit m−1 of the ordinary integer product from binary polynomial products. The package proves the exact full-domain GF(2) matrix rank, floor(2^(m−1)/3) + floor(3m/2) − 1 for m ≥ 2. With odd operands, the exact input width per side falls from 2^(m−3)+1 for affine decoding to m−1 for nonlinear decoding. Asking for the next lower bit as well raises the minimum to m. A single jointly mixed width-(m+1) channel produces the full odd product modulo 2^m with an exponential lookup.

The explicit five-bit circuit attains width four with 18 AND and 35 XOR gates in total. Three decoder ANDs are necessary for its fixed encoding. The paper also gives matching all-position bounds for separated convolution digits. These quantities concern an interface to integer multiplication. The independently verified full-multiplier comparator from #7830 is 0.00387992803556; the target of 0.003884928046155 remains open. Claimed κ gain for this package is zero.

Task and overlap

Base e545235. Changed scope from the supplier-boundary-audit card to a separately encoded carryless front end. #6252 bounds the size of propagated carries, #6552 supplies finite-field slices, and #7834 computes a different Gaussian multiplication pipeline. This package derives a product-bit matrix rank and matching channel widths. The #7830 comparator is pinned at 87fbf4f.

Validation and review requested

Run python3 -B scripts/research_package.py --package research/carryless-middle-bit-rank from the repository root with Python 3.12 on Linux. The checker covers the literal full and odd rank tables through m=12, 2,867,152 pre-carry digit comparisons, exhaustive five-bit gates, and 349,520 odd-input pairs for each of two field constructions through m=10. A separately written checker excluded every two-AND decoder for the five-bit encoding. Please review the valuation-zero elimination and the distinction between per-side width and actual gate cost.

Credit and next question

Thathachar's 1997 work on ordered representations and Bryant's earlier work are antecedents for exponential rank arguments about middle-bit multiplication. Clear1nsight used OpenAI Codex for derivation, implementation, and independent agent review. Source pins, proof assumptions and reproducible controls are in the package.

For six-bit odd operands, can one width-six carryless product recover bits four and five together with a small table-free decoder? The exact width lower bound is six. The seven-coordinate field construction gives a control. A proposed width-six code can first be checked for conflicting returns on all 1,024 input pairs.

@Clear1nsight

Copy link
Copy Markdown
Contributor Author

The registered package passed locally on Python 3.12.3. All 22 files and their executable modes match the published tree at cd9ce5a555977b469bd848199bc36025bd82fdcb. The GitHub planning job failed the repository's selected-result navigation test and skipped the package replay. I reproduced the same assertion on canonical main and this head: the selected record expects make frontier7000-verify, while the CI guide points to #6030. The test, guide and selected record have identical Git blobs on both commits.

The public handoff has the next six-bit experiment. Independent mathematical review of the rank elimination and of the single-channel versus unrestricted channel-list width is welcome.

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