Skip to content

Prove the exact threshold for fusing two congruent products - #7839

Open
Clear1nsight wants to merge 2 commits into
CrocSwap:mainfrom
Clear1nsight:codex/recursive-carry-sharing-0856
Open

Clear1nsight wants to merge 2 commits into
CrocSwap:mainfrom
Clear1nsight:codex/recursive-carry-sharing-0856

Conversation

@Clear1nsight

Copy link
Copy Markdown
Contributor

Change and mathematical scope

When can a parent obtain the low m bits of two products' sum, including its incoming carry, from one multiplication? If the siblings agree on a low bits on the left and b on the right, the exact threshold is a+b ≥ m+1 in the direct-product model, even with arbitrary separate encoders and one-sided additive corrections. One bit below the threshold, a single AND of difference parities repairs the high bit. For independent odd pairs, the high-bit-only carryless channel model needs exactly 2m−2 input coordinates per side.

At m=8 and a=b=4, the two-product control uses 170 AND and 170 XOR gates. The corrected circuit uses 91 AND and 107 XOR, including the incoming carry. Its clean reversible translation uses 182 Toffoli and 436 CNOT gates. I request review of these bounds, the circuits and their caller applicability. The κ target of 0.003884928046155 remains open.

Task and overlap

Changed scope on the supplier-boundary-audit card, based on canonical main e545235. Single-product predecessor #7838 has a different observable. Gaussian source #7834 has genuine sum-only parents. Mixed-rank reference #7830 retains separate outputs. Exact heads, complete reference histograms, source links and reproduction commands are in the package.

Validation

The primary verifier checked 1,044,944 admissible residues through five bits, all 65,536 four-bit circuit inputs with carry, and the literal six-bit odd row classes. Independent replay checked both clean gate programs and 20,480 additional eight-bit cases. A complete 20-by-19-bit Gaussian product retained its coefficients, residual, bits and cleanup with 74,655 → 62,284 ordinary word-product dispatches using a width-preserving three-product control. A separate phase control reduced an emitted middle fixture from 1,728 to 1,664 scalar calls. Both are counts of ordinary work. The 145-bit complete replay reached its CPU limit. The separate 8,200-bit diagnostic substituted an exact integer at the wide-worker boundary.

python3 -I -B research/recursive-dot-carry/verify.py
python3 -B scripts/research_package.py --package research/recursive-dot-carry
python3 -B scripts/check_research_commit.py --package research/recursive-dot-carry

Credit

Clear1nsight, with substantial OpenAI Codex assistance. Separate workers traced the two callers and independently checked the proofs, circuits and source replays. The optional Gaussian control uses the classical three-product identity. Credit for the underlying Gaussian program belongs to aliromman and its cited suppliers. The original mixed-rank and single-product contributors are credited in the package.

Open next questions

Find a charged parent where the sibling congruence holds on every allowed input. Pin both first consumers and charge retention or reconstruction of any separately observed difference. The current signed deficits are at least 4.88039108995 rational and 1.92561529070 binary units; the ordinary savings here supply zero on those invoices.

Include the direct residue and channel proofs, reversible circuit prices and pinned Gaussian caller controls.
@Clear1nsight

Copy link
Copy Markdown
Contributor Author

The exact submitted head is ad13267. A fresh standard-library replay of all 23 files from that Git tree passed in 7.8 seconds with Python 3.12. The independent review covers the shared-carry threshold, the clean circuits, the signed caller controls and the scope of the Gaussian dispatch savings. Independent caller review · Canonical research handoff

Hosted CI stopped in the onboarding planner before package selection. One of 760 tests failed. The navigation test expects the selected command make frontier7000-verify while the current CI guide lists make frontier6030-verify. I reproduced the same assertion on canonical base e545235 and on this PR head. The hosted package replay was skipped. Planner log · Hosted replay job

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