Investigate HMC for more efficient uncertainty estimation

[ergodic-ludwig]

Goal

Evaluate whether Hamiltonian Monte Carlo (HMC) can improve sampling efficiency for TSADAR uncertainty estimation relative to the current MCMC baseline.

The existing MCMC implementation is functional but computationally expensive; a high-quality run currently takes approximately 1–2 hours. HMC should be considered only after the baseline MCMC workflow is fully integrated.

Investigation

• Identify which fitted parameters and model paths expose usable gradients.
• Consult relevant experts and review suitable HMC implementations for the JAX-based stack.
• Define a representative inference case and compare HMC with the current MCMC baseline.
• Measure effective sample size, convergence diagnostics, wall time, and implementation complexity.
• Document incompatibilities, numerical stability concerns, or model changes needed for HMC.

Acceptance criteria

• Baseline MCMC integration prerequisites are identified and confirmed.
• A representative benchmark and comparison metrics are defined.
• At least one HMC approach is prototyped or ruled out with documented evidence.
• Results compare sampling quality and computational cost against baseline MCMC.
• A recommendation is recorded: adopt, defer, or reject HMC, with rationale.

Provenance

Action item from the Avi / Archis meeting on 2026-09-01. Discussion: HMC proposal at 00:26:34 and decision to defer until baseline integration at 00:29:09.