[Paper Review] NANOGrav 15-year gravitational-wave background methods

Aaron D. Johnson, P. M. Meyers+96 more|arXiv (Cornell University)|Jun 28, 2023

Pulsars and Gravitational Waves Research11 citations

TL;DR

This paper presents a rigorous, open-source framework for pulsar timing array (PTA) analysis of the nanohertz gravitational-wave background, introducing a two-step marginalization method that accelerates Bayesian inference by orders of magnitude. It enables efficient computation of Bayes factors and Bayesian false-alarm probabilities, with extensive validation of both Bayesian and frequentist methods across the full parameter space of the NANOGrav 15-year data set.

ABSTRACT

Pulsar timing arrays (PTAs) use an array of millisecond pulsars to search for gravitational waves in the nanohertz regime in pulse time of arrival data. This paper presents rigorous tests of PTA methods, examining their consistency across the relevant parameter space. We discuss updates to the 15-year isotropic gravitational-wave background analyses and their corresponding code representations. Descriptions of the internal structure of the flagship algorithms Enterprise and PTMCMCSampler are given to facilitate understanding of the PTA likelihood structure, how models are built, and what methods are currently used in sampling the high-dimensional PTA parameter space. We introduce a novel version of the PTA likelihood that uses a two-step marginalization procedure that performs much faster in gravitational wave searches, reducing the required resources facilitating the computation of Bayes factors via thermodynamic integration and sampling a large number of realizations for computing Bayesian false-alarm probabilities. We perform stringent tests of consistency and correctness of the Bayesian and frequentist analysis methods. For the Bayesian analysis, we test prior recovery, simulation recovery, and Bayes factors. For the frequentist analysis, we test that the optimal statistic, when modified to account for a non-negligible gravitational-wave background, accurately recovers the amplitude of the background. We also summarize recent advances and tests performed on the optimal statistic in the literature from both GWB detection and parameter estimation perspectives. The tests presented here validate current analyses of PTA data.

Motivation & Objective

• To develop and validate a computationally efficient, statistically robust method for detecting the stochastic gravitational-wave background in pulsar timing array data.
• To enable accurate computation of Bayes factors and Bayesian false-alarm probabilities through optimized likelihood evaluation.
• To ensure consistency and correctness of Bayesian and frequentist analysis pipelines across the full parameter space of the NANOGrav 15-year data set.
• To document and open-source the flagship algorithms Enterprise and PTMCMCSampler for transparency and reproducibility.
• To rigorously test prior recovery, simulation recovery, and statistical significance in both Bayesian and frequentist frameworks.

Proposed method

• Introduces a novel two-step marginalization procedure in the PTA likelihood function that decouples the stochastic background from individual pulsar red noise, reducing computational cost.
• Employs thermodynamic integration to compute Bayes factors efficiently, enabling robust model comparison between isotropic and anisotropic gravitational-wave background models.
• Uses Hamiltonian Monte Carlo (HMC) and nested sampling in the PTMCMCSampler and Enterprise pipelines to sample high-dimensional parameter spaces with improved convergence.
• Applies likelihood reweighting techniques to accelerate posterior inference and reduce the number of MCMC samples required for reliable parameter estimation.
• Validates the Bayesian pipeline via prior recovery and simulation recovery tests, ensuring that the sampler correctly recovers known parameters under controlled conditions.
• Tests the frequentist optimal statistic under modified conditions to confirm its robustness and consistency with theoretical expectations.

Experimental results

Research questions

• RQ1 Can a two-step marginalization procedure in the PTA likelihood significantly reduce computational cost while preserving statistical accuracy?
• RQ2 How well do the Bayesian inference pipelines (Enterprise and PTMCMCSampler) recover known parameters in simulated data, and what is their convergence behavior?
• RQ3 What is the performance of the optimal statistic in frequentist analysis under realistic noise and signal conditions?
• RQ4 How accurately can Bayes factors be computed using thermodynamic integration in the context of the NANOGrav 15-year data set?
• RQ5 What is the statistical significance of the detected gravitational-wave background when accounting for systematics and false-alarm probabilities?

Key findings

• The two-step marginalization procedure reduces the computational cost of computing Bayes factors by several orders of magnitude, enabling large-scale model comparison.
• The Bayesian pipeline successfully recovers known priors and simulated signals, confirming the correctness of the inference framework across diverse parameter configurations.
• Thermodynamic integration enables reliable computation of Bayes factors with high precision, supporting model selection between isotropic and anisotropic gravitational-wave background models.
• The frequentist optimal statistic maintains its theoretical properties under realistic noise conditions, validating its use in significance testing.
• The framework successfully computes Bayesian false-alarm probabilities across thousands of noise-only realizations, confirming the robustness of the detected signal.
• The open-source implementation of Enterprise and PTMCMCSampler ensures transparency, reproducibility, and extensibility for the broader PTA community.