Add debug.md

debug.md

@@ -0,0 +1,33 @@
+# Debug Log
+
+## Numerical Issues
+
+### CAVI Convergence
+- Some configurations with weak priors (a0=0.1, b0=0.1) and high count scales show slow convergence
+- Many runs hit max_iter=200 without convergence (tol=1e-5)
+- This is expected behavior: weak priors with high counts create strong coupling, making convergence harder
+- Results are still usable as the parameters have stabilized significantly
+
+### Gaussian-Gamma MAP
+- Gradient-based optimization may not converge to global optimum
+- Learning rate lr=0.005 is conservative but stable
+- Objective values may fluctuate in non-convex landscape
+
+## Exclusion Rules
+- Configs with fewer than 10 edges after count generation are skipped
+- Deletions where the edge has count 0 are not meaningful for Poisson-Gamma and are excluded
+- Decay fits with fewer than 3 valid distance shells are marked as invalid
+
+## Known Limitations
+- Pilot runs use N=150, M=150 which limits graph diameter to ~4-6 hops
+- BFS neighborhood computation includes all nodes at distance ≤ R, which for high-degree graphs may cover most of the graph at R=3-4
+- One-step downdate is a very rough approximation and may sometimes increase error
+
+## Runtime Notes
+- Full model fit: ~0.1-1.0s for N=M=150, K=5-10
+- Exact deletion (warm-started): ~0.1-0.5s each
+- Local R=2: ~0.05-0.3s each
+- Total per config (20 deletions): ~20-40s
+
+## Timestamp
+Generated: 2026-04-24