[Paper Review] CosmoPMC: Cosmology Population Monte Carlo

M. Kilbinger, K. Benabed+8 more|arXiv (Cornell University)|Jan 5, 2011

Cosmology and Gravitation TheoriesPhysics and Astronomy3 references17 citations

TL;DR

CosmoPMC introduces a Bayesian sampling framework using Population Monte Carlo (PMC) to efficiently explore cosmological parameter spaces, combining adaptive importance sampling with parallelized likelihood computation via MPI. It enables accurate posterior estimation and Bayesian evidence calculation, validated on cosmic shear, SNIa, CMB, and galaxy clustering data with high-dimensional parameter spaces.

ABSTRACT

We present the public release of the Bayesian sampling algorithm for cosmology, CosmoPMC (Cosmology Population Monte Carlo). CosmoPMC explores the parameter space of various cosmological probes, and also provides a robust estimate of the Bayesian evidence. CosmoPMC is based on an adaptive importance sampling method called Population Monte Carlo (PMC). Various cosmology likelihood modules are implemented, and new modules can be added easily. The importance-sampling algorithm is written in C, and fully parallelised using the Message Passing Interface (MPI). Due to very little overhead, the wall-clock time required for sampling scales approximately with the number of CPUs. The CosmoPMC package contains post-processing and plotting programs, and in addition a Monte-Carlo Markov chain (MCMC) algorithm. The sampling engine is implemented in the library pmclib, and can be used independently. The software is available for download at http://www.cosmopmc.info.

Motivation & Objective

• To develop a scalable, parallelized Bayesian inference tool for cosmological parameter estimation using adaptive importance sampling.
• To improve posterior exploration and evidence computation in high-dimensional cosmological parameter spaces compared to traditional MCMC.
• To provide a modular, extensible software framework for cosmological data analysis with support for multiple probes (e.g., CMB, weak lensing, SNIa).
• To enable accurate model comparison via Bayesian evidence estimation using PMC's adaptive proposal distributions.
• To integrate advanced cosmological likelihood modules (e.g., CAMB, WMAP, halofit) into a unified, high-performance computational pipeline.

Proposed method

• Employs Population Monte Carlo (PMC), an adaptive importance sampling method that iteratively improves proposal densities to approximate the posterior distribution.
• Uses the pmclib library as the core PMC engine, with MPI parallelization to distribute likelihood evaluations across multiple processors.
• Applies importance weights to samples drawn from a sequence of proposal densities, with weights computed as the ratio of unnormalized posterior to proposal density.
• Supports multiple cosmological probes through modular likelihood modules: cosmic shear, SNIa, CMB anisotropies, galaxy clustering, BAO, and redshift distributions.
• Integrates external tools like CAMB for CMB power spectra, WMAP likelihood code, and halofit for non-linear matter power spectra.
• Provides post-processing tools for marginalization, credible interval estimation, evidence computation, and reparameterization using R, Python, or Yorick.

Experimental results

Research questions

• RQ1 Can PMC outperform standard MCMC in sampling efficiency and evidence estimation for high-dimensional cosmological parameter spaces?
• RQ2 How accurately can PMC estimate Bayesian evidence for cosmological model comparison across diverse data sets like cosmic shear and SNIa?
• RQ3 To what extent does MPI parallelization reduce wall-clock time in likelihood evaluation for large-scale cosmological surveys?
• RQ4 How robust is PMC to model complexity and non-Gaussian posterior shapes in cosmological inference?
• RQ5 Can the modular design of CosmoPMC be effectively extended to include new cosmological probes or likelihood models?

Key findings

• CosmoPMC enables efficient exploration of high-dimensional cosmological parameter spaces using adaptive importance sampling, reducing reliance on random-walk MCMC.
• The method achieves accurate Bayesian evidence estimation, crucial for model comparison, as demonstrated in Kilbinger et al. (2010) using PMC.
• Parallelization via MPI significantly reduces wall-clock time, making large-scale cosmological analyses feasible on high-performance clusters.
• The framework successfully integrates diverse cosmological probes—including cosmic shear, SNIa, CMB, and galaxy clustering—into a single inference pipeline.
• Post-processing tools allow reliable estimation of marginalized posteriors, credible intervals, and Bayes factors, with support for 1D and 2D marginal plots.
• The codebase supports extensibility through modular design, enabling users to add new likelihood modules or modify existing ones with minimal changes.