Gibbs Sampler

hydra_tod.full_Gibbs_sampler is the top-level entry point for Bayesian joint calibration and map-making. It orchestrates four conditionally conjugate sampling steps that alternate until the Markov chain converges:

  1. Gain — draws smooth polynomial gain coefficients for each TOD independently (see gain_sampler).

  2. Local temperatures — draws receiver-noise and noise-diode coefficients for each TOD independently (see tsys_sampler).

  3. Noise parameters — draws \((\log f_0, \alpha)\) for each TOD independently via MCMC or NUTS (see noise_sampler_fixed_fc).

  4. Sky temperature — draws shared celestial pixel temperatures jointly across all TODs, synchronised over MPI ranks (see tsys_sampler).

Steps 1–3 are independent per TOD and are parallelised across MPI ranks (one or more TODs per rank). Step 4 uses an MPI_Allreduce over accumulated normal equations, so all ranks must call it collectively.

TOD_Gibbs_sampler() accepts its many arguments in four conceptual groups:

  • Data — local_TOD_list, local_t_lists

  • Operators — local_gain_operator_list, local_Tsky_operator_list, local_Tloc_operator_list

  • Priors — prior_cov_inv_*, prior_mean_*

  • Config — n_samples, gain_model, sampler, Est_mode, jeffreys, bounds, solver tolerances

For a working end-to-end example, see the Quick Start page and the tutorial notebooks in the examples/ directory.

See also

Parameter Samplers — individual Gibbs step modules. Simulation — generating synthetic TOD for input. Quick Start — annotated end-to-end example.