The Pocomc sampler

Preconditioned Monte Carlo for Accelerated Bayesian Computation

Name	pocomc
Version	1.2.1
Author(s)	Minas Karamanis, David Nabergoj, Florian Beutler, John Peacock, Uros Seljak
URL	https://github.com/minaskar/pocomc
Citation(s)	https://arxiv.org/abs/2207.05660, https://arxiv.org/abs/2207.05652
Parallelism	parallel

pocoMC is a Python package for fast Bayesian posterior and model evidence estimation. It leverages the Preconditioned Monte Carlo (PMC) algorithm, offering significant speed improvements over traditional methods like MCMC and Nested Sampling. Ideal for large-scale scientific problems with expensive likelihood evaluations, non-linear correlations, and multimodality, pocoMC provides efficient and scalable posterior sampling and model evidence estimation. Widely used in cosmology and astronomy, pocoMC is user-friendly, flexible, and actively maintained.

It generates both posterior samples and Bayesian evidences.

Installation

if using conda, do this first to ensure consistent dependencies:

conda install -c conda-forge corner pytorch tqdm

pip install pocomc

Parameters

These parameters can be set in the sampler’s section in the ini parameter file. If no default is specified then the parameter is required. A listing of “(empty)” means a blank string is the default.

Name	Type	Default	Description
n_effective	integer, default=512		Number of effective particles/walkers. It should be at least twice the number of active particles.
n_active	integer, default=256		Number of active particles/walkers. It should no more than half of the number of effective particles. It is recommended to be a multiple of the number of CPUs for optimal performance.
flow	string		The type of normalizing flow to use. Default is “nsf6” (Neural Spline Flow with 6 transforms). Other options include “nsf3”, “nsf12”, “maf3”, “maf6”, “maf12” (Masked Autoregressive Flow).
precondition	bool, default=True		Whether to use Normalizing Flow Preconditioning to accelerate sampling.
dynamic	bool, default=True		Whether to use dynamic particle allocation.
n_total	integer, default=4096		Total number of effective samples to generate. Higher values will increase the accuracy of the posterior estimate (recommended ~10000 for publication quality corner plots).
n_evidence	integer, default=4096		Number of samples to generate for the Bayesian evidence calculation.
n_steps	integer, default=None		Number of MCMC steps. This will be automatically adjusted based on the sampling efficiency. Default is n_steps=n_dim. Larger values will provide more robust results at higher computational cost.
n_max_steps	integer, default=None		Maximum number of MCMC steps. Default is n_max_steps=10*n_dim. This should be greater than n_steps.
save_every	int, default=None		Save the state of the sampler every save_every iterations. If not provided, the sampler will not save intermediate states.
seed	int	0	A random seed for reproducibility. Default is to generate automatically.
progress	bool, default=True		Whether to show a progress bar during sampling.