Adapting a novel framework for rapid inference of massive black hole binaries for LISA

TL;DR

This paper presents an adapted version of the RIFT framework for efficient Bayesian inference of massive black hole binaries in LISA data, leveraging parallel architecture without waveform acceleration.

Contribution

It introduces a parallelized RIFT implementation for LISA, enabling rapid analysis of complex gravitational wave models without specialized hardware.

Findings

Accurate parameter inference of black hole binaries using NRHybSur3dq8.

Higher modes significantly impact LISA data analysis.

Inference of multiple signals remains robust.

Abstract

The Laser Interferometer Space Antenna (LISA) is designed to detect a variety of gravitational-wave events, including mergers of massive black hole binaries, stellar-mass black hole inspirals, and extreme mass-ratio inspirals. LISA's capability to observe signals with high signal-to-noise ratios raises concerns about waveform accuracy. Additionally, its ability to observe long-duration signals will raise the computational cost of Bayesian inference, making it challenging to use costly and novel models with standard stochastic sampling methods without incorporating likelihood and waveform acceleration techniques. In this work, we present our attempt to tackle these issues. We adapt ${\tt RIFT}$ for LISA to take advantage of its embarrassingly parallel architecture, enabling efficient analysis of large datasets with costly gravitational wave models without relying on likelihood or waveform acceleration. We demonstrate that our open-source code can accurately infer parameters of massive black hole binary signals by carrying out a zero-noise injection recovery using the numerical relativity surrogate model ${\tt NRHybSur3dq8}$. By utilizing all available $m\neq0$ modes in the inference, we study the impact of higher modes on LISA data analysis. We study the impact of multiple massive black hole binary signals in a dataset on the inference of a single signal, showing that the selected source's inference remains largely unaffected. Furthermore, we analyze the LDC-1A and blind LDC-2A datasets from the Radler and Sangria challenge of the LISA data challenges. When eschewing specialized hardware, we find ${\tt NRHybSur3dq8}$ injection-recovery takes approximately $20$ hours to complete, while the analysis of Sangria and Radler datasets takes about $10$ hours to complete.