allaun/SilverSight
1
0
Fork 0
mirror of https://github.com/allaunthefox/SilverSight.git synced 2026-08-20 14:07:29 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions
SilverSight/docs/build_logs/2026-06-22_session_bmcte_hachimoji_bridge.md
allaun e02eab7181 feat(bmcte): N=20000 sweep on neon-64gb + Hachimoji bridge theorem 
- BMCTE sweep: N=20000, p=12,14 on neon-64gb (ARM64, 62 GB, 18 cores)
- Optimizations: N×p isometry (O(N·p²) QR), vectorized Ryser permanent (65× speedup)
- p=12: H=11.607, λ=0.9928; p=14: H=11.699, λ=0.9902; total 73.6s, 0 overflows
- HachimojiBridging.lean §11: lambdaBMCTE = exp(-p²/N) with monotonicity proofs

Build: 2987 jobs, 0 errors (lake build)
2026-06-22 14:17:31 -05:00

2 KiB
Raw Permalink Blame History

Build Log: 2026-06-22 — BMCTE N=20000 Sweep & Hachimoji Bridge

Session Summary

Extended the BMCTE sweep to N=20000, p=12,14 on neon-64gb (ARM64, 62 GB, 18 cores). Applied two performance optimizations to extension_v1.py: N×p tall-skinny QR instead of full N×N QR, and vectorized Ryser permanent via numpy broadcasting (65× speedup at p=14). Formalized the BMCTE→Hachimoji link theorem in HachimojiBridging.lean §11.

Files Changed

File Change
experiments/bosonic_continuous/extension_v1.py N×p isometry (O(N·p²) vs O(N³)); vectorized Ryser (65× faster); removed debug logging
formal/CoreFormalism/HachimojiBridging.lean §11 — lambdaBMCTE(p,N), monotonicity proofs, H_max/entropyRatio
docs/build_logs/2026-06-22_session_bmcte_hachimoji_bridge.md NEW — this file
AGENTS.md Updated HachimojiBridging entry, extension_v1 entry
docs/PROJECT_MAP.md Regenerated
docs/PROJECT_MAP.json Regenerated

Sweep Results (N=20000, neon-64gb)

p λ(p) H_mean H_std nonzero modes runtime
12 0.9928 11.607 0.092 8845 4.1s/seed
14 0.9902 11.699 0.146 9898 10.6s/seed

Total: 73.6s for 10 runs (5 seeds each), 0 overflows.

Build Baseline

lake build CoreFormalism.HachimojiBridging → 2978 jobs, 0 errors
lake build → 3201 jobs, 0 errors (full workspace)

Key Decisions

  • N×p isometry over full N×N QR: BMCTE only uses the first p columns of the unitary. QR on N×p Gaussian produces the correct marginal distribution of a Haar-random isometry, matching the scattering-matrix physical model, at O(N·p²) instead of O(N³).
  • Vectorized Ryser: replaced 2²ᵖ Python loop with (2ᵖ, p) boolean inclusion matrix × matrix multiply. 65× faster at p=14 (0.28s → 0.0043s per permanent).
  • neon-64gb for large-N CPU sweep: 62 GB RAM and 18 ARM64 cores handled N=20000 where RTX 4070 (12 GB VRAM) OOM'd on full unitary.