LCDM Sampling on the HPC¶
submits run_single_lcdm_chain.sh once per (data map, chain). Specify num_maps data realizations and num_chains chains per map in the master slurm submission script sample_lcdm.sh. Averaging over maps removes the cosmic variance of any one map, and running several chains per map checks convergence.
What to set in sample_lcdm.sh¶
| Variables | Meaning |
|---|---|
nside, theta_pix, noise_level, pol |
basic configurations like pixel resolution, map size, etc, |
*_ground |
the ground-truth cosmology the data are simulated at |
*_min, *_max |
the sampling range for each parameter (by default ±5 sigma) |
num_maps, num_chains, map_prefix |
how many realizations and chains; map seeds are map_prefix * map |
sample_* |
true / false: whether each parameter is sampled (default omch2, theta_MC_100, logA). A sampled parameter starts each chain at a uniform random point in its [*_min, *_max] box, an unsampled one is held at *_ground |
*_proposal_sigma |
the Metropolis proposal width of each parameter (ombh2, omch2, theta_MC_100, log_a, ns) |
num_burn_in_fix_theta, iters_per_chain, num_burn_in_always_accept, hmc_num_steps, hmc_step_size |
sampler settings |
out_dir_name |
the run's folder inside lcdm_sampler_data/ (default joint_inference_<submission date>) |
Every variable has a default in the script and can be overridden on the command line as --<name>=<value>, e.g.
bash get_lcdm_samples/sample_lcdm.sh --noise_level=10 --omch2_proposal_sigma=2e-3 --out_dir_name=run_10uk
A misspelled name is ignored and the default used. Tune the proposal widths first with a pilot chain.
Output¶
Chains are written as *_history.txt files (one line per iteration, every parameter) under get_lcdm_samples/lcdm_sampler_data/<out_dir_name>/. The same folder holds submission_settings.txt: the submission date, every setting, and each chain's random starting point (rewritten on every run, so use a new out_dir_name when you change settings). Copy that folder to your machine and run the chain analysis script.