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: around0.8-1.2grad_noise_to_langevin_med: around0.1-1.0(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
epsand learning rate conservatively.
Robust validation protocol
Single-chain ESS is insufficient for full-posterior confidence. For production inference:
Run multiple chains from overdispersed starts.
Check split-Rhat and rank plots per key parameter block.
Compare posterior summaries across chains.
Use posterior predictive checks when possible.