Sampler Health Checks

This page summarizes practical checks for pSGLD/SGHMC behavior in Phase 4.

Core diagnostics

  • t_eff_var_over_target (temperature calibration proxy)

  • grad_noise_to_langevin_* (gradient-noise vs injected-noise balance)

  • per-parameter/block ESS and split-Rhat (multi-chain)

Suggested targets

  • t_eff_var_over_target: around 0.8-1.2

  • grad_noise_to_langevin_med: around 0.1-1.0 (order of magnitude)

These are heuristics, not strict guarantees.

If exploration looks too cold

  • Reduce SGHMC friction (sghmc_alpha) if overly damped.

  • Verify inference.batching.standard.shuffle=true.

  • Revisit preconditioner freeze timing (freeze_preconditioner_sampling).

  • Adjust eps and learning rate conservatively.

Robust validation protocol

Single-chain ESS is insufficient for full-posterior confidence. For production inference:

  1. Run multiple chains from overdispersed starts.

  2. Check split-Rhat and rank plots per key parameter block.

  3. Compare posterior summaries across chains.

  4. Use posterior predictive checks when possible.