[Paper Review] A fully-automated end-to-end pipeline for massive black hole binary signal extraction from LISA data

Michael L. Katz|arXiv (Cornell University)|Nov 1, 2021

Pulsars and Gravitational Waves Research74 references35 citations

TL;DR

This paper presents a fully automated, GPU-accelerated end-to-end pipeline for extracting massive black hole binary (MBHB) signals from LISA data, using brute-force search, heterodyning-based likelihood maximization, and parameter estimation. The method achieves converged posterior distributions in under an hour with no human intervention, demonstrating high-accuracy parameter recovery on both noiseless and noisy LDC-1A datasets, including a realistic higher-harmonic posterior for improved source characterization.

ABSTRACT

The LISA Data Challenges Working Group within the LISA Consortium has started publishing datasets to benchmark, compare, and build LISA data analysis infrastructure as the Consortium prepares for the launch of the mission. We present our solution to the dataset from LISA Data Challenge (LDC) 1A containing a single massive black hole binary signal. This solution is built from a fully-automated and GPU-accelerated pipeline consisting of three segments: a brute-force initial search; a refining search that uses the efficient Likelihood computation technique of Heterodyning (also called Relative Binning) to locate the maximum Likelihood point; and a parameter estimation portion that also takes advantage of the speed of the Heterodyning method. This pipeline takes tens of minutes to evolve from randomized initial parameters throughout the prior volume to a converged final posterior distribution. Final posteriors are shown for both datasets from LDC 1A: one noiseless data stream and one containing additive noise. A posterior distribution including higher harmonics is also shown for a self-injected waveform with the same source parameters as is used in the original LDC 1A dataset. This higher-mode posterior is shown in order to provide a more realistic distribution on the parameters of the source.

Motivation & Objective

• To develop a fully automated, end-to-end data analysis pipeline for massive black hole binary (MBHB) signal extraction from LISA data, requiring no human intervention or prior knowledge of the signal.
• To enable rapid and accurate parameter estimation for MBHBs in realistic LISA data conditions, including non-stationary noise and instrumental effects.
• To demonstrate the feasibility of high-accuracy Bayesian inference on LISA data using efficient likelihood computation techniques like heterodyning (relative binning).
• To provide a more realistic posterior distribution by including higher harmonic modes in the waveform model, improving parameter constraints beyond the dominant l=m=2 mode.
• To establish a scalable framework that can be extended to multi-source scenarios and future LISA data challenges involving overlapping signals.

Proposed method

• The pipeline uses a brute-force initial search across the full prior volume of MBHB parameters, accelerated via GPU computation to handle the high-dimensional parameter space efficiently.
• A subsequent maximum-likelihood refinement step employs the heterodyning (relative binning) technique for fast and accurate likelihood evaluation, significantly reducing computational cost compared to standard matched filtering.
• The parameter estimation phase leverages the same heterodyning method to compute posterior distributions efficiently, enabling full Bayesian inference on the signal parameters.
• The method is applied to two LDC-1A datasets: one with no noise and one with realistic LISA instrumental noise (SciRDv1 curve), both containing a single MBHB injection.
• A higher-harmonic waveform model is used to generate a more realistic posterior distribution for the same source parameters, incorporating l=m=2, l=3,m=2, and l=2,m=1 modes.
• The pipeline is fully automated, requiring no manual tuning or prior knowledge, and runs to convergence in under one hour on standard GPU hardware.

Experimental results

Research questions

• RQ1 Can a fully automated, GPU-accelerated pipeline achieve accurate MBHB parameter estimation in under one hour from randomized initial parameters?
• RQ2 How does the inclusion of higher harmonic modes in the waveform model affect the posterior distribution and parameter constraints?
• RQ3 Can the heterodyning (relative binning) likelihood method enable efficient and accurate likelihood maximization in the presence of realistic LISA noise?
• RQ4 How robust is the pipeline to non-stationary noise and data gaps, as modeled in the LDC-1A dataset?
• RQ5 Can the method be extended to multi-source scenarios where signals may overlap in the merger-ringdown phase?

Key findings

• The pipeline successfully extracts the MBHB signal from both noiseless and noisy LDC-1A datasets, producing converged posterior distributions in under one hour of runtime with no human intervention.
• The posterior distributions for the extrinsic parameters (right ascension, declination, polarization angle, and inclination) are well-constrained, with 1σ, 2σ, and 3σ contours closely surrounding the true injected values in the noisy dataset.
• The inclusion of higher harmonic modes in the waveform model results in a more realistic and tighter posterior distribution, reducing parameter uncertainties compared to the dominant-mode-only analysis.
• The heterodyning (relative binning) method enables efficient likelihood computation, allowing the pipeline to explore the full prior volume and converge to the maximum-likelihood point rapidly.
• The method demonstrates robustness to realistic LISA noise, as evidenced by the accurate recovery of source parameters in the noisy dataset, validating its readiness for future LISA data challenges.
• The pipeline is scalable and extensible, with the potential to handle multiple MBHBs and overlapping signals through global fitting, provided merger-ringdown overlaps are rare.