LeatherSagiKnowledgebase

Technical Reference -- Load-Bearing Facts for Design Choices

3 min read

Technical Reference

These facts are load-bearing for design choices. They inform optimizer selection, R-value protocols, and feasibility gates. Read this page when making decisions about estimation strategy, pricing method, or experiment design.

Likelihood Evaluation

  • RBPF N=3000: std ~0.6 (TSM-only) up to 14.97 (full constrained MLE), ~0.35 s/eval at 8 threads. Use N=3000-5000 for PMMH.
  • Hamilton filter: 2 ms/eval, deterministic, but ignores cap rates — Hamilton-optimal != RBPF-optimal (~1572-nat gap). Use only as cheap surrogate for early-stage screening.
  • MC pricing lookup (Phase D0): 116x speedup vs direct MC, 2.77-nat accuracy. Per-eval rebuild required — fixed lookup drifts +5269 nats (catastrophic).
  • T_bar = 1400 (production): validated DeltaLL = 0.0 vs T_bar = 5000 across 46 theta vectors. Safety: monitor_tbar=true.
  • Riccati convergence: A → -infinity is correct (exp(A) → 0). T_bar requirements are state-dependent; Apartment is the slow asset (needs T_bar ~ 50K at the new MAP).
  • Crisis adjustment: crisis_range=57:70, crisis_scale=10.0 (cap rate sigma x 10 during crisis). Slated for removal — see synthesis brief D1.
  • CRN variance reduction: only 1.4-1.6x (regime-path fragility). Use CRN for tie-break tier, not inner loop.

Optimization (54-D)

  • BOBYQA dominates for noisy MC objectives. NelderMead stagnates in 54-D; L-BFGS fails.
  • Per-eval MC rebuild is mandatory. AD (158 s/eval) too slow for inner loop.
  • m_g IS-stationarity constraint: 55 → 54 free params, 70,350x gradient conditioning improvement. Settled.
  • Two-stage R protocol: R=100/500 search → R=5000 polish on top-K. R=500 gives 100% hit rate (within 20-nat tolerance to DGP); R=100 has ~25% trap rate.
  • Paired-seed high-R tie-break: R=100k across 5 paired MC seeds achieves SEM ~0.013-0.34 nats. Mandatory for sub-nat ranking.
  • Truth-start saturation: ~600 evals at R=100. Going beyond 1000 evals chases MC noise.
  • Restart-allocation late-bloom: improvement is non-monotone in elite rank. Worst-screening elites can be best polishers. Use uniform allocation, not greedy.

Feasibility

  • Absorbing spectral proxy: 99.1% accuracy for no-bubble check (vs 63.3% for old get_lim_eta). 2.87x more feasible Sobol points. This is the production gate.
  • Old is_feasible_pixg_cap is deprecated.
  • Feasibility is learnable: GBT classifier (AUC 0.997, depth 6, 500 trees) gives 19.6x wall-hour speedup. Region-specific.

Curvature Landscape

See experiment pages for full details:

Key facts:

  • Condition number ~ 10^24-10^26 (extreme sloppiness)
  • 25-dim null space = Income block (24) + Risk price (1), exactly zero on Hamilton surrogate
  • Top-2 stiff directions are Monetary means — stencil-robust
  • Stencil caveat: forward-FD diagonal stencil is gradient-contaminated at any non-critical point. Ranks 3-5 and the 30/24 sign structure are forward-stencil artifacts. Only ranks 1-2 and top-4 block labels are reliable.
  • rotation_scaling.json is DIAGNOSTIC_ONLY, not fit for restart preconditioning

Linked Experiment Pages

Graph View

Backlinks