Skip to content

Raise the exact-sum Gaussian rounding slack floor to sqrt(13)-3 (no kappa) - #7673

Open
aliromman wants to merge 3 commits into
CrocSwap:mainfrom
aliromman:codex/consecutive-result-019
Open

aliromman wants to merge 3 commits into
CrocSwap:mainfrom
aliromman:codex/consecutive-result-019

Conversation

@aliromman

@aliromman aliromman commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

Result and scope

Every injective, exact-sum Gaussian-pair rounder with Euclidean error at most one needs forward norm slack at least sqrt(13)-3, about 0.605551275464. This raises the necessary bound sqrt(5)-sqrt(3), about 0.504017169931, in PR #7652. Eighteen inputs have only sixteen legal outputs below the new threshold. A matching attains equality on all 22 sum-one inputs with squared norm at most 13.

This is a scoped proof with an exact finite certificate. It assumes no period, phase covariance, square-to-swap identity, inverse contract or surjectivity. It does not prove global attainment. No new kappa, supplier admission, selected-frontier change or selected-assumption discharge is claimed. All five selected assumptions remain.

Read the proof, complete matching, exact receipt and verifier. This answers the larger-neighborhood question in the pinned predecessor and completes the advance claim.

Evidence

Normal and optimized CPython 3.12.16 replay passed on macOS arm64. The checks cover all 88 legal edges, the complete 18-to-16 Hall neighborhood, all 22 matching edges, a separate four-coordinate Gaussian calculation, exact rational root separation, independent augmenting-path sizes 20 and 22, and four rejection controls. The written Hall argument and explicit matching suffice without trusting the matching search.

Metadata, committed bytes, executable modes, source ancestry and the official 11-file export passed. The optimized export replay passed in a read-only directory with no Git tree and PATH set to /nonexistent; all 12 payload/manifest file hashes and modes stayed unchanged. The exact predecessor README and its license are retained and hashed, so no unmerged external Git object is needed.

From the repository root, using CPython 3.12:

python3.12 -I -W error -B research/gaussian-exact-sum-next-obstruction/verify.py
python3.12 -I -W error -B -O research/gaussian-exact-sum-next-obstruction/verify.py
python3.12 -B scripts/research_package.py --package research/gaussian-exact-sum-next-obstruction
python3.12 -B scripts/check_research_commit.py --package research/gaussian-exact-sum-next-obstruction --check-source-history --check-modes

Expected status markers:

PASS_GAUSSIAN_EXACT_SUM_NEXT_OBSTRUCTION
PASS_REGISTERED_RESEARCH_PACKAGE
PASS_RESEARCH_COMMIT_BYTES

The earlier hosted run at 424273b0c82886d93526548faa0cbdd56ca3216b did not certify this package. The latest head changes only proof formatting. Its plan job failed at the same shared guide test; package replay was skipped. The plan job failed in the shared test_current_ci_section_matches_selected_record: the base guide names frontier6030 while the selected record requires make frontier7000-verify. The log reports 760 tests, one failure and 24 skipped; package replay was skipped. This package changes no shared CI file. Local replay remains the scientific evidence. No predecessor scientific replay, selected cold replay, Lean or Docker run was performed. Independent human and formal-kernel review are unclaimed.

Credit, relation and remaining work

Prepared by @aliromman with substantial OpenAI Codex assistance. Nick Petalas (@petalas) supplied the candidate formula, earlier lower bound and open question in PR #7652, source at 317fecee. Jerry Chen (@CubeJerry) supplied the fixed-sum matching context in PR #7570. Ali Romman's PR #6509 supplies the ideal pair. Exact source identity and attribution are retained. No source code was copied.

PR #7651 uses an unrestricted rounding graph; this result requires exact sum. PR #7572 already handles affine-period slack; that idea was discarded. The overlap refresh covered 972 open PR bodies, 6175 closed PR bodies and 7603 Discussion records; the sparse local note search was incomplete and is not a novelty certificate.

The next useful step is an all-lattice matching or structural Hall argument. Larger passing finite censuses alone do not settle global attainment. Physical execution, signed range, precision, inverse/caller behavior and every implementation cost remain open. The task record states the boundary and stop rule. Canonical base is e54523517d0f03e58352e647c403a5af3e525a57; selected PR #7000 is unchanged.

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