AI-Assisted Discovery and Construction of a Counterexample to the Convergence of Three-Block ADMM with the Identity Matrix as Its Third Constraint Block
English | 简体中文
This repository accompanies the manuscript “AI-Assisted Discovery and Construction of a Counterexample to the Convergence of Three-Block ADMM with the Identity Matrix as Its Third Constraint Block.” It provides exact, replayable evidence for two fixed convex quadratic programs on which the unmodified direct three-block ADMM has a bounded non-KKT periodic sequence, despite a unique KKT point.
| If you want to... | Start with |
|---|---|
| Read the mathematical argument | arXiv:2608.14396 · Repository PDF |
| Verify both counterexamples | Five-minute verification |
| Inspect the exact machine certificates | certificates/ |
| Understand what each checker proves | Reproducibility contract |
| Follow the Codex and Kimi discovery routes | Research-stage index |
| Read the privacy-redacted Codex and Kimi interactions | Consolidated transcript · Raw text |
| Review time, token, and agent accounting | Computational provenance |
| Run the independent MATLAB check | MATLAB instructions |
The root package is the acceptance layer. The
research-process/ archive is historical evidence about
how the results were found; it is not a second acceptance layer.
| Certificate ID | Fixed instance | Exact conclusion | Main evidence |
|---|---|---|---|
identity_slack_p66_short_v1 |
m = 2, A = B = I_2, beta = 1 |
A specified initialization generates a bounded non-KKT orbit of minimal period 66 | two independent Python representations and a MATLAB implementation |
identity_slack_p23_rational_v1 |
m = 3, rational QP data, beta = 1 |
An open invariant set of reduced initializations converges phasewise to a non-KKT orbit of minimal period 23 | exact rational replay and Lyapunov certificate |
- Projection word:
(00)^2(01)^64. - Exact closure and minimal period:
66. - Strict projection checks:
132/132. - Minimum signed margin:
0.0037105246944352910173... > 1/1000. - The orbit is bounded and non-KKT; the QP has a unique KKT point.
- This certificate concerns a deliberately specified initialization. It does not claim attraction, unbounded divergence, or failure of corrected ADMM variants.
- All primitive QP coefficients are reduced fractions with numerator
magnitude and denominator at most
100. - The frozen certificate contains the complete exact phase-zero state
(x^0,y^0,z^0,lambda^0). - Exact closure and minimal period:
23. - Strict projection checks:
69/69, with minimum margin greater than1/250. - A rational matrix
PsatisfiesP - M_per^T P M_per > 0exactly. - The ellipsoid
e^T P e < 1/4000is return-invariant in the canonical reduced(y,t)state, and every initialization in it converges phasewise to the period-23 sequence. - This is a neighborhood of initializations for one fixed QP, not a perturbation result for the QP data or a global attraction theorem.
The Kimi Code K3 route originally reached an exact dyadic period-23 replay and
an exact Jury local-attraction certificate. The denominator-100 instance and
explicit invariant ellipsoid above are later release strengthening; see the
route attestation
and recompute its public ledger with
python python/verify_kimi_provenance.py --check.
The exact relaxation certificate proves three separate statements:
- one rational Lyapunov matrix works on the strict KKT branch for every
tau in [49/100, 51/100]; - the former period-66 initialization follows a certified strict prefix for
232 steps and enters the invariant ellipsoid for every
tau in [1/2 - 10^-10, 1/2 + 10^-10]; - the strict KKT branch has a unique Schur boundary satisfying
0.9366061114 < tau_c < 0.9366061115.
These are local or fixed-initialization results. They do not establish global convergence from arbitrary initial points.
The frozen Python environment is 3.13.5 with SymPy 1.13.3, NumPy 2.1.3, and pytest 8.3.4.
git clone https://github.com/ConanXu-math/identity-slack-admm-cycle-certificate.git
cd identity-slack-admm-cycle-certificate
python3.13 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python python/verify_all.py
python python/verify_research_process_archive.pyThe certificate command should end with:
{"checks": [{"name": "period66", "returncode": 0, "status": "passed"}, {"name": "period23", "returncode": 0, "status": "passed"}], "valid": true}For the full release check used by GitHub Actions:
python python/verify_all.py
python python/verify_research_process_archive.py
python python/verify_kimi_provenance.py --check
python python/export_orbit_66.py
python python/certify_relaxed_multiplier_interval_theory.py
python python/verify_universal_step_obstruction.py --check
python -m pytest -q python/tests/test_relaxed_multiplier_interval_theory.py
python -m unittest discover -s tests -p "test_*.py"
python python/verify_matlab_certificate.py
git diff --exit-code -- certificates/The last command is part of acceptance: regeneration must leave every tracked certificate byte-for-byte unchanged.
The universal-step check keeps the JSON structure and every exact symbolic
field byte-strict. Its six NumPy/LAPACK spectral-radius values are explicitly
numerical sanity checks and use absolute tolerance 1e-12 across platforms;
the smallest stored excess above one is greater than 3.87e-7.
| Layer | Purpose | Authoritative entry point |
|---|---|---|
| Exact acceptance | Rebuilds the fixed QPs and checks every finite proof obligation | python/verify_all.py |
| Frozen data | Canonical inputs, exact orbits, verdicts, hashes, and Lyapunov data | certificates/ |
| Implementation cross-checks | Independent signed-state, full-state, and MATLAB implementations for period 66 | python/README.md, matlab/README.md |
| Process archive | Selected theory, experiments, failures, state files, and internal reviews | research-process/INDEX.md |
| Route accounting | Scope and telemetry for the Codex/Kimi comparison | provenance/README.md |
| Continuous integration | Clean-environment regeneration and artifact-stability gate | Exact certificate workflow |
Agreement between implementations is an internal reproducibility cross-check, not external peer review.
Use research-process/INDEX.md rather than
browsing the archive chronologically:
- Codex / period 66: begin with the persistent task state and algebraic reductions, then follow Stages 43–46 from obstruction to numerical discovery, rationalization, exact replay, and precision audit.
- Kimi Code K3 / period 23: begin with
START_GOAL.txtandRESEARCH_LOG.md, then follow the withdrawn routes, targeted instability experiments, period locking, and the exact certificate.
Archive labels matter:
numerical_screenandproof_attemptare exploratory;withdrawnrecords a route that was rejected;theoremandexact_certificateare accepted only within their stated scope;reviewmeans an internal check, not external peer review.
Raw native session files, credentials, private configuration, local absolute
paths, caches, and repetitive bulk outputs are intentionally excluded. A
single privacy-redacted transcript of the user-visible Codex and Kimi Code K3
interactions is published separately as
INTERACTION_TRANSCRIPT.md.
The 168 retained files in research-process/ are covered by
research-process/manifest.json and checked
in CI.
The MATLAB verifier covers the period-66 instance and requires MATLAB R2025a, Symbolic Math Toolbox, and a valid license:
addpath("matlab")
result = verify_exact_cycle_matlab();
assert(result.valid)Then compare the generated MATLAB JSON against the frozen Python fields:
python python/verify_matlab_certificate.pySee matlab/README.md for the class-based test and licensed
GitHub Actions instructions.
.
├── python/ exact Python verifiers and comparison drivers
├── matlab/ independent period-66 MATLAB verifier and tests
├── certificates/ frozen inputs and machine-readable certificates
├── research-process/ curated Codex and Kimi discovery archives
├── provenance/ comparison scope, accounting, and evidence boundaries
├── docs/ detailed reproducibility contract
├── paper/ compiled manuscript PDF
└── .github/ CI workflows and ownership rules
For script-level descriptions, see python/README.md. For
the exact predicates, artifact meanings, runtime contract, and release
checklist, see docs/REPRODUCIBILITY.md.
The two routes independently reached exact counterexamples to the same proposed convergence principle, but they did not solve the same QP and were not matched in compute, tools, telemetry, stopping policy, or human intervention. The comparison is descriptive and endpoint-aligned; it cannot support a causal ranking of model speed, cost, or capability.
The software source code and machine-readable certificate data in this
repository are licensed under the MIT License. The manuscript PDF
under paper/ and prose research materials are excluded from the MIT License
and remain all rights reserved unless a file states otherwise.
Before creating an immutable archival release, the maintainers must freeze the
author list, add CITATION.cff, create a tagged archive, assign a DOI, and make
the manuscript and code-availability statement point to that same release.
Repository owner: ConanXu-math.