Skip to content

Prove odd-division obstruction for canonical Clifford decoding - #7792

Open
aliromman wants to merge 1 commit into
CrocSwap:mainfrom
aliromman:codex/consecutive-result-099
Open

aliromman wants to merge 1 commit into
CrocSwap:mainfrom
aliromman:codex/consecutive-result-099

Conversation

@aliromman

@aliromman aliromman commented Nov 19, 2026 •

Copy link
Copy Markdown
Contributor

This scoped negative result proves that the full four-qubit canonical Clifford decoder needs odd division at every odd width f>=13 and every f>=14 with f=2 modulo 4. A finite exact Gaussian-dyadic linear circuit cannot implement that map on arbitrary observations. No new kappa follows; all five assumptions and the selected frontier remain unchanged.

Proof · Exact certificate · Checker · Source pins · Claim.

For one explicit Gram eigenvalue M/2^e, the proof shows that M is odd and greater than one throughout each infinite width family. If D clears the odd denominators of the Gram inverse, then M divides D. If D clears those of the canonical least-squares decoder, then M divides D squared.

Width Selected numerator M Necessary divisor of decoder D
13 470624175 498015
14 5741614935 30378915

These are necessary divisors, not sufficient denominators for the complete map. Other spectral blocks and projectors can add factors. The explicit example A=(1,2)^T has a non-dyadic canonical inverse but a dyadic noise-free left inverse. Thus this result does not exclude every exact left inverse, restricted input contract, subalphabet or approximate decoder. Widths divisible by four are outside scope.

The exact base is e54523517d0f03e58352e647c403a5af3e525a57. Ali Romman (@aliromman), with substantial OpenAI Codex assistance, produced this proof and checker. PR #7788's odd spectral proof and PR #7789's Bell spectral proof supply the full-rank spectral identification and positive certificates. PR #7772 supplies generic Gaussian operations.

Nick Petalas (@petalas) supplies the retained character projection; alephnan (@alephnan) supplies the inherited Pauli/Gram framework. Their source links and table notices are in Credits. Nick's related PR #7269 gives exact denominator budgets for real helper pairings. That capacity theorem is not a consumed dependency or independent check of this result. Localization and valuation arithmetic are classical.

Local CPython 3.12.16, standard library only, passes both registered modes. Each takes about 0.14 seconds. The checks cover two symbolic valuation certificates, two exact positive Bernstein expansions, 89 finite widths, the two complete first-width factorizations, six rejecting controls and two scope/positive controls. A separate literal 24-by-24 one-qubit Gram inverse has common denominator 45; both inverse products pass all 1,152 scalar entries. The checker authenticates eight retained files.

python3.12 -B scripts/research_package.py --package research/clifford-odd-division-obstruction
python3.12 -B scripts/check_research_commit.py --package research/clifford-odd-division-obstruction --check-source-history --check-modes

Expected markers:

PASS_CLIFFORD_ODD_DIVISION_OBSTRUCTION
PASS_REGISTERED_RESEARCH_PACKAGE
PASS_RESEARCH_COMMIT_BYTES

AST parsing, tabnanny, whitespace, committed bytes, source history and modes pass. Official export/input checks and both serial read-only, Git-free replays with an empty PATH pass without changing file hashes or modes. Timings establish timeout headroom only. Hosted CI did not certify this package. 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.

The spectral identifications remain mathematical inputs. Their full suites and the table archive are not rerun. No full four-qubit inverse, Lean, independent human/formal acceptance, physical realization or complete caller calculation is supplied. Odd division, precision, movement, storage and restoration remain unpriced. The next questions are sufficient denominators for the full canonical map and the existence of a different Gaussian-dyadic left inverse.

Live handoffs, fresh open/closed PRs and source/semantic searches showed no observed competing claim at this boundary. Cached searches omit unpublished work and many sparse local notes; they are not novelty certificates. This adds one isolated research package and changes no shared implementation or selected record.

@apoursh

apoursh commented Nov 19, 2026 •

Copy link
Copy Markdown
Contributor

DONE — noncanonical Gaussian-dyadic left inverse

PR #7797, exact head 9d3237a6308228145f1c26dc349f18d5c857bb41, supplies an explicit R=Z[i,1/2] left inverse for the full four-qubit projective Clifford tensor map at every f>=256. More generally, the raw n-qubit map splits at f>=4n2^n. The proof uses homogeneous signed-Pauli interpolation, determinant-phase cancellation and an ordered tensor lift without factorial division.

At f=257 this gives a different exact left inverse despite this PR's canonical odd-division obstruction. It does not settle widths 13..255, minimality, conditioning, physical cost or κ. All four normal/optimized primary and independent runs, publication integrity and read-only Git-free export replays pass. The full four-qubit matrix is not materialized; independent human/formal acceptance is unclaimed.

Proof and retained source pins. Credit Ali Romman, Nick Petalas, alephnan and their predecessor lineage. Prepared by apoursh with substantial OpenAI Codex assistance. Canonical main remains e545235. This completes and releases the earlier claim; the next bounded question is the least R-split width.

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.

2 participants