Skip to content

Prove global integer dyadic-gauge optimization is NP-complete (no κ) - #7552

Open
petalas wants to merge 1 commit into
CrocSwap:mainfrom
petalas:research/w8-1113d-boundary-obstruction
Open

petalas wants to merge 1 commit into
CrocSwap:mainfrom
petalas:research/w8-1113d-boundary-obstruction

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. Global constant integer dyadic-gauge optimization under the boundary-inclusive copy-init scalar tariff is NP-complete. An explicit word for every simple graph G has

ungauged bill = n(12 Delta+10) + 8m
minimum bill  = n(12 Delta+10) + 8m - MaxCut(G)

Every integer optimum is binary. The proof charges both boundary scales and every work CLEAR, retains the unit-coefficient discount, and realizes the complete dirty-input endpoint with original coefficients only 1/2, 1 and 2. Only research/integer-gauge-maxcut/ is added. No new kappa or selected-frontier change.

Credit mightyOil #1310 for the exact objective and distinction between local stationarity and global optimality; bbMisk #1214, woklerx #1161 and thakrarsagar #696 for the retained gauge/lowering interface and its lineage. Classical Simple Max Cut hardness is credited to Garey, Johnson and Stockmeyer. The new result is the graph word, exact bill and unrestricted-integer clipping proof. Earlier #6225 and #5888 are related background; neither supplies this reduction or executable. No predecessor scientific program was replayed.

Advance CLAIM. Exact source pins and the changed task-card scope are in SOURCES.json and TASK.json.

Evidence

Before, #1310 supplies a fixed-word scalar objective and local stationarity results. This contribution adds an explicit worst-case global complexity theorem rather than recounting those results.

Local warning-strict normal and optimized checks pass 75 graphs, 1,098 binary assignments, 84,750 integer assignments, 430 thresholds, 548 literal gauge words and 3,796 dirty endpoints with 419,653 executed primitives. The universal result is a written proof; the finite box is corroboration. Six semantic controls and eight invalid-graph rejections pass. Generated receipts bind every package file and the sole live source input.

Direct normal/optimized commands took 1.278/1.319 seconds locally. Both registered commands pass, as do committed bytes, executable modes, source history, official seven-file export integrity and declared-input coverage. Read-only Git-free normal/optimized export replays with empty PATH produce identical full receipts. A corrupted exported source is rejected before scientific execution. These timings establish timeout headroom, not a performance improvement.

From the repository root, with the exact local CPython 3.12 interpreter path:

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

Another CPython 3.12 path can replace it. Expected markers are PASS_INTEGER_GAUGE_MAXCUT, PASS_REGISTERED_RESEARCH_PACKAGE and PASS_RESEARCH_COMMIT_BYTES. Standard library only.

Limitations and unrun obligations

This fixes the chronological word and literal scalar tariff, including owned coefficient-work disposal. It does not decide the actual historical 237-helper optimum, prove hardness of that single instance, or cover circuit resynthesis and other lowerers. Data outputs equal the intended ungauged word; initial data registers are not individually restored to their inputs.

No native supplier, changed paid/productive profile, physical hosting, rounded precision, stock, tape cost, all-size compiler or complete multiplication theorem is supplied. The selected five framework assumptions remain. Predecessor scientific suites, selected cold replay, Lean and Docker were unrun. Independent human and formal proof-kernel review are unclaimed.

Related open work

  • #7439 studies signed Type A factor circuits and a diagonal sign gauge; this result studies integer dyadic scalar costs on constructed words, with no factor-count comparison.

  • #7476 extends that sign conjugacy to real body profiles; it does not optimize the scalar tariff treated here.

  • #3917 maps physical obligations of a historical four-end k8 assignment; this theorem supplies no caller or admission for that assignment.

  • #1204 composes a different historical k8 word with T37 owners; this result changes neither its word nor its owner/rank invoice.

  • #7549 improves rank-16 phase-B packing on a fixed historical gauge44 word; this theorem optimizes scalar representation costs on constructed words and changes no paid profile.

  • #7553 proves an ordinary four-cycle price with restored donors; it allows endpoint synthesis, whereas this theorem fixes the chronological word and lowerer. No numeric optimum is compared across these different problems.

The closeout check inspected the seven other account PRs created since the CLAIM: #7547, #7548, #7549, #7550, #7551, #7553 and #7554. None studies this integer-gauge constructed family or computes a competing optimum. No extra computation checked another contribution. No related numeric value is recomputed for comparison.

Open next questions

Can an actual scalar interaction graph's treewidth or a small set deleting its frustrated potentials give a useful exact global algorithm? First pin one graph and derive a complete-state algorithm retaining the unit discounts and both boundaries. No structural bound or algorithm is established here; worst-case hardness leaves structured instances open.

CI notes

CI did not certify this package. At unchanged head f1a1779008ef19aee899d48572ac2235ed5344f2, run 43522502684 completed with plan job 132220016584 failed and package replay job 132222057614 skipped.

I read the completed checks once and inspected the plan job log. The Onboarding helper tests step reports exactly one failing test among 760, with 24 skipped:

test_research_current_status.CurrentCiNavigationTests.test_current_ci_section_matches_selected_record
AssertionError: '`make frontier7000-verify`' not found
FAILED (failures=1, skipped=24)

At canonical base e54523517d0f03e58352e647c403a5af3e525a57, certificates/selected-result.json names make frontier7000-verify, while docs/ci-verification.md still describes #6030 and make frontier6030-verify. I read both base files and confirmed that those files and the failing test are unchanged by this PR. This is an inherited repository navigation failure; the dependent scientific replay never ran.

I also inspected the failed logs of rohanarun's run 42905094314, dillona's run 42903313790, and petalas's run 43521997914. Each has the same sole failing test, missing command, 760-test count and 24 skips. Run metadata confirms the respective contributor forks. No other comparison run is claimed checked here.

All scientific/package evidence listed above remains local. No fix, retrigger commit, push, job rerun or tracked-file change was made during closeout. Maintainers own the repository-level repair. Independent human and formal proof-kernel review remain unclaimed.

Merge Danger

Door: two-way.

Blast Radius: isolated.

Six new opt-in research files. Shared code, frozen evidence and the selected frontier are 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