This directory is the public code supplement for the manuscript:
Robust MIMO Channel Identification under Matrix-Variate Heavy-Tailed Noise via Sign-Preserving Fractional Polynomial Maximization
It is intentionally separated from the manuscript sources. It contains the open implementation code, deterministic result snapshots, R reproduction scripts, rendered figures, and a compact Lean 4 formal supplement.
python/- reference Python package, experiment drivers, configs, data download helpers, and tests. Includes the Middleton Class A DGP (src/mv_patp/data_gen/middleton.py), the complex-baseband demo (experiments/mv_patp_mimo/run_complex_demo.py), and the symbolic Recovery Theorem check (theory/recovery_symbolic_check.py).data/results/- CSV and manifest snapshots for the result directories used by the manuscript.R/- lightweight R scripts that summarize the CSV snapshots and render independent PDF checks.figures/- manuscript PDF figures copied from the release build.Lean/,lakefile.lean,lean-toolchain- Lean 4 project with symbolic definitions and structural checks for the oPMM score and recovery anchor.REPRODUCIBILITY.md- exact commands for Python, R, and Lean checks.
cd code_supplement/python
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,real_data]"
pytest -q -m "not slow and not gate"Regenerate manuscript figures from the bundled result snapshots:
python -m experiments.mv_patp_mimo.plot \
--synthetic-dir ../data/results/synthetic_full_v1 \
--deepmimo-dir ../data/results/deepmimo_v2 \
--output-dir ../figures \
--format pdfcd code_supplement
Rscript R/run_all.RThe R scripts regenerate summary CSVs under R/results/ and independent PDF
checks under figures/.
cd code_supplement
lake buildThe Lean files are not a full mechanization of the statistical proof. They are
a compact formal supplement for symbolic definitions, admissible alpha ranges,
the trace-normalization invariant, and the alpha = 3 PMM3 recovery anchor.