Skip to content

Recover all unknown Clifford terms from exact tensor trains - #7815

Open
aliromman wants to merge 1 commit into
CrocSwap:mainfrom
aliromman:codex/major-clifford-candidate-synthesis
Open

aliromman wants to merge 1 commit into
CrocSwap:mainfrom
aliromman:codex/major-clifford-candidate-synthesis

Conversation

@aliromman

@aliromman aliromman commented Nov 19, 2026 •

Copy link
Copy Markdown
Contributor

An exact decoder now finds every unknown four-qubit Clifford label and its complex coefficient in a supplied tensor train, with a full membership test. At width 27 it covers arbitrary support in the 12,128,668,876,800-label dictionary. Its cost is polynomial in the train parameters, explicit input bits and physical alphabet. It needs no candidate list or prior scan of the dictionary. This is a new algorithmic interface. The selected complete κ remains 0.00363005817. Recovery proof and costs · Support certificates · Decoder · Source pins

Four-qubit tensor width Support covered
13 Any 19 labels
23 Any 447,742 labels
25 Any 58,040,063 labels, or any support with at most 226,718 complete symplectic fibres
27 and above Every support

A small observed flattening rank can also make rejection exclude the entire Clifford span without an input sparsity assumption. At width 25 a twelve-site cut of rank at most 58,040,062 suffices. The algorithm samples spanning boundary rows and columns, recovers projective entry ratios, checks Pauli conjugation, solves for all signed weights and verifies the residual over the full tensor. The complete-fibre bound follows from a missing-representation projector that is constant on each 256-label fibre. The table lists sufficient guarantees. Exact cut conditions · Independent character calculation

The public example provides 6,144 sparse records for an observation with 2¹⁰⁴ entries at width 13. Its all-zero entry cancels. The decoder recovers 19 matrices in different symplectic fibres and all 19 normalized coefficients, including the common coefficient 1/16+i/4. A separately written dense oracle checked 280 decisions on 140 tensors. Three separate Codex reviewers tested gauges, partial-rank cuts, off-model false candidates, coefficient sizes and the character argument. Run the example · Parity producer · Adversarial evidence · Dense oracle

Sources and scope

@petalas supplied the quotient projection from the ATLAS ordinary table in Thomas Breuer’s CTblLib 1.3.11. @alephnan supplied the Pauli-sector framework. @aliromman supplied the width-11 and width-13 decompositions, the width-12 null projector and the conditioning bound. The original archive matched all 81 projected rows and 324 retained character values. PR #7747 · Projection · PR #6714 · PR #7766 · Odd-width proof · PR #7770 · Even-width proof · PR #7772

@apoursh supplied the sharp dyadic decoder and, most recently, all 256 identity-fibre outputs on compact affine-quadratic input. Those outputs have a specified linear extension even off span. This PR accepts general exact train cores, locates occupied fibres and certifies membership. Input construction, approximate-data identification, reversible tape execution and a complete caller that pays both supplier budgets remain open. All five frontier assumptions remain in force. The general two-independent-factor decomposition principle is credited to Leurgans, Ross and Abel. PR #7799 · PR #7813 · Compact identity-fibre proof · Classical and repository credit

The campaign inventory reconciles all 100 published results to their frozen heads and credits. It compares this direction with @Clear1nsight’s pending 48-call rational and 36-call binary construction and with the recursive tape/precision interfaces. The base is canonical main e54523517d0f03e58352e647c403a5af3e525a57, priority revision 8. Base · 100-result inventory · Direction comparison · PR #7786 · Research discussion

Reproduction

Use CPython 3.12 and the standard library from the repository root.

python3.12 -B scripts/research_package.py --package research/clifford-unknown-support-decoder

Expected final line

PASS_REGISTERED_RESEARCH_PACKAGE research/clifford-unknown-support-decoder outcome=completed

All eight registered commands passed at the submitted commit in the checkout and in a read-only, Git-free export. Committed bytes, modes, source history and export coverage passed. The large support counts are proved cardinalities. The largest executable fixture in the adversarial review had 96 terms. The full group, historical 100 suites, selected cold replay, Lean and Docker were not run. Independent scientific acceptance remains open. Prepared by @aliromman with substantial OpenAI Codex assistance. Registered checks

Hosted verification. The shared planning job ran 760 tests, reported one failure and 24 skips, and skipped package replay. Its CI guide still gives frontier 6030 while the selected record requires make frontier7000-verify. The same navigation test failed locally on the unchanged base files. The separate guide repair is already open. Failed job · PR #7276

Derive restricted flank certificates, reconstruct signed mixtures and verify the complete tensor. Includes the source-bound 100-result synthesis and independent exact checks.
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