Skip to content

Construct Gaussian-dyadic four-qubit Clifford left inverses for widths ≥256 (no κ) - #7797

Open
apoursh wants to merge 1 commit into
CrocSwap:mainfrom
apoursh:codex/clifford-dyadic-left-inverse
Open

apoursh wants to merge 1 commit into
CrocSwap:mainfrom
apoursh:codex/clifford-dyadic-left-inverse

Conversation

@apoursh

@apoursh apoursh commented Nov 19, 2026

Copy link
Copy Markdown
Contributor

Change and mathematical scope

Proof · Package and checks

The full four-qubit projective Clifford tensor map admits an exact left inverse over R=Z[i,1/2] at every integer width f>=256. More generally, the raw n-qubit tensor map admits one at f>=4n2^n. The construction proves that the column module is a direct summand, hence saturated, at those widths.

This answers the noncanonical left-inverse question in #7792 at an explicit sufficient width. At f=257, #7792's canonical least-squares decoder needs odd division, while this different left inverse does not. Widths 13..255, the least possible width, useful conditioning and physical implementation costs remain unresolved. No new κ or supplier admission.

For each target representative, the construction compares its 2n signed Pauli-generator images with those of a variable matrix. Adjugates make these tests polynomial; an even number of detector factors cancels determinant phases. Their product is a homogeneous dyadic interpolant. Assigning each monomial coefficient to one ordered tensor coordinate produces the linear left inverse without factorial division. A linear factor pads to every larger width. The proof, rather than finite evaluations, supplies the universal quantifiers.

Task, sources and overlap

Canonical upstream was pulled at entry and refreshed before publication: e54523517d0f03e58352e647c403a5af3e525a57. Task uniform-transfer-audit, scope key 58d0403e3078b2c5ae8a, is explicitly changed scope to an algebraic feasibility side result. The original revision-8 card and the bounded family decision are retained in TASK.json; no selected framework assumption is claimed discharged.

The precise predecessors are:

Three unmodified proof files are retained with SHA-256, original Git blob IDs, exact source commits and URLs in SOURCE-PINS.json. Replay does not need unmerged Git objects, network access or predecessor code. Their spectral computations are not independently rerun; only the comparison with the canonical map depends on that spectral result.

Canonical investigation comment. Fresh all-state searches and newest PR bodies found no matching positive construction. The dedicated overlap helper lacked gh, and Discussion/sparse-note coverage is incomplete. These limits are recorded; no global novelty claim is made. Newer cubic results #7506/#7543 were checked before choosing a distinct question.

Validation

CPython 3.12.14, macOS arm64, standard library only. All four registered normal/optimized runs passed:

python3.12 -B scripts/research_package.py --package research/clifford-dyadic-left-inverse
python3.12 -B scripts/check_research_commit.py --package research/clifford-dyadic-left-inverse --check-source-history --check-modes
python3.12 -B scripts/export_research_package.py export --package research/clifford-dyadic-left-inverse --ref HEAD --output /tmp/clifford-left-inverse.tar
python3.12 -B scripts/export_research_package.py check /tmp/clifford-left-inverse.tar

Primary checks: 24 one-qubit cosets, 1,392 commutative polynomial coefficients, all 576 normalized left-inverse entries, 256 literal tensor-coordinate checks; 16 literal four-qubit matrices, 128 generator images, 2,048 pairings and 256 interpolation entries. Six faulty variants and three mutated source files are rejected.

The separate implementation imports neither primary nor predecessor code. Entry enumeration and ordered-word expansion check 96 one-qubit matrices, 24 projective cosets, 2,304 interpolation pairs, 6,912 phase-padding cases, all 576 left-inverse entries, 12,288 row terms and 1,536 tensor coordinates. Five faulty variants reject. The non-Clifford CCZ control evaluates to 27/512; this exposes the essential Pauli-normalizer hypothesis.

Committed bytes, executable modes and source history passed for all 11 package files at 9d3237a6308228145f1c26dc349f18d5c857bb41. Official export integrity passed. All four direct runs passed in a fresh read-only Git-free export with empty PATH; all 12 files including the receipt retained their bytes and modes. The full four-qubit matrix is not materialized. Selected source-binding orientation passed; the selected scientific cold replay, full predecessor spectral suites, Lean and Docker were not run.

The base has an existing CI navigation mismatch: test_current_ci_section_matches_selected_record expects frontier7000 while the guide names frontier6030. Existing #7276 covers that shared fix. This package does not change shared CI files or weaken the test; hosted status will be reported separately from local verification.

Credit and open next questions

Prepared by apoursh with substantial OpenAI Codex and separate agent assistance in proof, implementation and review. Credit Ali Romman, Nick Petalas, alephnan and the retained predecessor lineage. Adjugates, Pauli orthogonality, polynomial interpolation and module splitting are classical methods. Apache-2.0; no independent human or formal-kernel acceptance is claimed.

The package's README.md and TASK.json give two bounded next questions: determine the least width admitting an R-left inverse, and derive a recovery norm plus a complete physical invoice for one specified caller. The present coefficient-ring result supplies neither an efficient algorithm nor a useful error-amplification bound.

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