GRPCA-GD

Graph-smooth sparse orthogonal PCA with a split-variable solver and a synthetic-first evaluation. This repository contains the Track-A manuscript, code, and reproducible synthetic experiments.

Methods Compared (Paper-1)

• PCA (dense baseline)
• Minimal A-ManPG (external sparse orthogonal baseline, no graph)
• SparseNoGraph (in-family ablation with (\lambda_2=0))
• Proposed (sparse + graph-smooth + joint orthogonality)

Experiments Included

• Chain robustness (seeds 0–4)
• SBM robustness (seeds 0–4)
• SBM (\lambda_2) sweep (single seed)

Quickstart

Run a smoke test (r=1, rho=5.0, eta_A=0.05):

uv run python main.py configs/smoke_r1.yaml

Run the small r=3 config:

uv run python main.py configs/small_r3.yaml

Reproduce Paper Results

1. Run robustness configs (chain + SBM, seeds 0–4):

for s in 0 1 2 3 4; do
  uv run python main.py configs/robust_chain_seed${s}.yaml
  uv run python main.py configs/robust_sbm_seed${s}.yaml
 done

2. Run SBM (\lambda_2) sweep:

for val in 0p00 0p05 0p10 0p20 0p50; do
  uv run python main.py configs/sbm_lambda2_${val}.yaml
 done

3. Regenerate the sweep panel figure:

uv run python - <<'PY'
from pathlib import Path
import json
import numpy as np
import matplotlib.pyplot as plt

root = Path('outputs')
order = ['0p00','0p05','0p10','0p20','0p50']
lams = [0.0,0.05,0.1,0.2,0.5]

support_f1 = []
smooth_norm = []
expl_var = []

for tag in order:
    metrics = json.loads((root / f'sbm_lambda2_{tag}' / 'metrics.json').read_text())
    proposed = metrics['Proposed']
    support_f1.append(proposed['support_metrics']['union']['f1'])
    smooth_norm.append(proposed['graph_smoothness_norm_trueL'])
    expl_var.append(proposed['shared_explained_variance'])

fig, axes = plt.subplots(1, 3, figsize=(10, 3))
axes[0].plot(lams, support_f1, marker='o')
axes[0].set_title('Support F1')
axes[0].set_xlabel('lambda2')

axes[1].plot(lams, smooth_norm, marker='o')
axes[1].set_title('Graph Smoothness (norm)')
axes[1].set_xlabel('lambda2')

axes[2].plot(lams, expl_var, marker='o')
axes[2].set_title('Shared Explained Variance')
axes[2].set_xlabel('lambda2')

plt.tight_layout()
Path('figures').mkdir(exist_ok=True)
plt.savefig('figures/sbm_lambda2_sweep_panel.png', dpi=200)
PY

4. Update tables in latex/manuscript_sample.tex using the newly generated outputs. The tables in the manuscript should match the means/stds computed from the metrics.json files in outputs/.

Manuscript Build

Compile from the repo root (not from inside latex/):

pdflatex latex/manuscript_sample.tex
BIBINPUTS=latex: bibtex manuscript_sample
pdflatex latex/manuscript_sample.tex
pdflatex latex/manuscript_sample.tex

Outputs

Each run writes to outputs/<run_name>/:

• artifacts.npz and artifacts.json
• manifest.json
• metrics.json and metrics.csv
• plots/ convergence traces

Paper Snapshot

The frozen Track-A manuscript PDF is stored at:

• paper/paper1-trackA-v1.pdf