Skip to content

Prove sharp polynomial budgets for acyclic interaction errors (no κ) - #7507

Open
petalas wants to merge 2 commits into
CrocSwap:mainfrom
petalas:research/w7-1113a-helper-geometry
Open

petalas wants to merge 2 commits into
CrocSwap:mainfrom
petalas:research/w7-1113a-helper-geometry

Conversation

@petalas

@petalas petalas commented Nov 13, 2026 •

Copy link
Copy Markdown
Contributor

Summary

Scoped written proof with exact finite corroboration, produced by Nick Petalas. For interaction errors supported on one common acyclic coordinate graph of path depth at most h, the sharp forward envelope is the degree-h sum of elementary symmetric polynomials in the individual norm budgets. The sharp inverse envelope uses complete homogeneous symmetric polynomials. Explicit shift matrices attain each bound, with separate sign choices. The same proof gives sharp operator-defect and suffix-weighted arithmetic-noise bounds.

Common depth one, individual budgets d_i:
  sufficient shear product: product_i(1+d_i)
  new exact envelope:       1+sum_i d_i

Twenty errors of norm 1/10 have exact depth-one forward and inverse envelope 3. At depth two the sharp envelopes for the same list are 49/10 and 51/10. The common acyclic interaction-support hypothesis is essential and is not established for a selected caller. No new kappa or selected-frontier change.

Credit ywz65's PR2701 for the interaction-coordinate contract and aliromman's PR5288 for the distinct sufficient square-zero-shear product comparison. Classical nilpotent expansions and symmetric polynomials retain their mathematical credit. No predecessor executable or scientific replay is used. The card explicitly changes scope to this structured precision-interface result.

Advance claim. Canonical main was refreshed before submission and remains e545235. Live claims, open PRs, all-main README terms and targeted merged searches were checked. Full open/Discussion pagination completed; sparse local notes were incomplete and a separate complete closed-PR refresh timed out. Those limits do not affect the self-contained theorem or replay.

Evidence

Before, the cited structure-dependent contract supplied a sufficient scalar product and explicitly left sharpness open. After imposing common acyclic support, the new proof gives exact envelopes over every graph of bounded depth, with all-dimensional attaining families.

The standalone replay reports PASS_ACYCLIC_INTERACTION_BUDGETS and errors=0. It checks 326 polynomial coefficients, all 729 three-dimensional matrix pairs over the stated entry alphabet, 15 permuted words, 746 support graphs, 1,553 prefixes and noisy steps, 51 attaining families, 20 malformed-input rejections and seven mathematical controls. Matrix inverses use independent rational Gaussian elimination; graph depths use an adjacency-power oracle. Finite checks corroborate the written universal proof.

Local CPython 3.12.15 normal and optimized checks pass in about 0.5 seconds each, below the registered 120-second limits. Registered package replay, committed bytes/modes and full source-history checks pass at head e637e0e. Official export and declared-input checks pass. Both direct modes pass from the read-only Git-free export with an empty executable search path and identical receipts. Each receipt is generated and binds all seven package-file hashes. Syntax and function-length checks pass. No local mathematical check is deferred to CI.

Exact reproduction from the repository root:

/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -W error -B research/acyclic-interaction-budgets/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -O -W error -B research/acyclic-interaction-budgets/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/research_package.py --package research/acyclic-interaction-budgets
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/check_research_commit.py --package research/acyclic-interaction-budgets --check-source-history --check-modes

Another CPython 3.12 executable may replace that absolute path. Only the standard library is needed.

Limitations and unrun obligations

No actual selected word has been bound to a common acyclic interaction graph. Physical gate support can change under the required prefix conjugations, as the included exact control demonstrates. Sharpness is over all graphs of depth at most h, rather than each prescribed graph. Physical frame-norm factors are sufficient, with no optimality claim.

An inverse of a realized linear product is not an inverse of nonlinear grid rounding. Actual coefficient storage, arithmetic error, overflow, basis-change costs, native operations, ownership/restoration, precision/storage/tape costs, complete signed invoice and supplier admission remain open. All five framework assumptions remain. Independent human and formal proof-kernel review are unclaimed. No selected cold replay, predecessor scientific replay, Lean, Docker or native build was run.

Open next questions

  1. Bind an actual local tile and coefficient-storage policy to the interaction-support hypothesis, including basis norms, arithmetic noise and physical costs.
  2. Find the sharper envelope for one prescribed graph that excludes the common-shift witness, after a real consumer and error budget are supplied.

Related open work

  • PR7474 bounds a nonlinear rounded half-shear between Walsh axes; this theorem concerns structured perturbations of linear products and does not compare its error constant.

  • PR7471 characterizes affine-lattice output hulls; this result bounds norms and does not certify lattice closure.

  • PR7247 studies rounded tensor-image block minima; no tensor-image or quantizer minimum is claimed here.

  • PR7211 reduces rounded-MOVE component width; this package does not bind its interaction errors or change a MOVE width.

  • PR7192 studies residue-assisted rounded inverse injectivity; the present linear-product inverse is explicitly a different contract.

  • PR7268 reduces rounded-ADD control size; this package supplies no control implementation or physical savings.

  • PR371 accounts for Gaussian final-grid rounding; these structured interaction bounds do not certify a Gaussian caller or its final rounding.

  • PR7505 supplies an exact banded-dyadic inverse fallback retaining residual/error certificates; this theorem bounds norms of structured interaction products and supplies no inverse-preparation algorithm or Gaussian caller. No shared number is compared.

The own-open-PR list was refreshed after the advance CLAIM and before DONE. No open PR on the same common-acyclic interaction envelope was found in the checked corpus. No extra computation was performed to validate another contribution.

CI notes

Hosted CI did not certify this package. At unchanged head e637e0e158814f2b588fbfc91e5ebaac3650c520, run 43437748958, plan job 131927117881, failed in Onboarding helper tests. The downloaded job log reports exactly one failure among 760 tests, with 24 skipped:

test_research_current_status.CurrentCiNavigationTests.test_current_ci_section_matches_selected_record
AssertionError: '`make frontier7000-verify`' not found in the CI verification doc

At base e54523517d0f03e58352e647c403a5af3e525a57, certificates/selected-result.json names make frontier7000-verify, but the current section of docs/ci-verification.md still describes #6030 and make frontier6030-verify. I inspected the test, base document and this run's failure log. This PR only adds its seven isolated package files and changes neither shared file. The replay job was skipped because plan failed, so there is no hosted package replay or scientific certification.

I downloaded and verified the same sole test failure, missing command and 760-test/24-skip totals in run 42905094314, from rohanarun, head d14f787dc0948188f8f5485115d92f3ac1b0688c, and run 43437279341, from petalas, head 82d677390993e28e12a0203169a69d9d929aa5a2. These are the other runs verified here; no claim is made to have inspected every affected PR.

All required local package checks, committed-byte/source-history checks and read-only Git-free exported replays passed as recorded above. The universal theorem rests on the written proof, with finite corroboration. No retrigger, new push, shared-file repair or CI rerun was made. Main's documentation repair remains with the maintainers.

Merge Danger

Door: two-way.

Seven new files under research/acyclic-interaction-budgets. Frozen sources, shared code and selected-frontier files are unchanged.

Blast Radius: isolated.

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