Neural posterior estimation of Galactic Binary signals for the LISA mission

Tanguy Delmond, Natalia Korsakova, Thomas Oberlin, Sylvain Marsat, Antoine Basset, Nicolas Dobigeon · 2026

Read Neural posterior estimation of Galactic Binary signals for the LISA mission by Tanguy Delmond, Natalia Korsakova, Thomas Oberlin, Sylvain Marsat, Antoine Basset, Nicolas Dobigeon (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-29039.txt.

Introduction

Delmond, Korsakova, Oberlin, Marsat, Basset and Dobigeon speed up source analysis for LISA, a planned space mission that will detect gravitational waves at very low frequencies. Galactic binaries, pairs of stars orbiting each other, should be LISA's most numerous sources. Their signals will overlap heavily in time and frequency, making standard Bayesian sampling slow and hard to scale.

The team use simulation-based inference instead. A conditional normalising flow, a neural network that reshapes simple distributions into complex ones, learns to estimate source parameters from simulated data. No likelihood calculation is needed in training or use. Once trained, the estimator produces thousands of posterior samples per second. The authors show results for one source in a narrow frequency band, then wider bands. As a proof of concept, they tackle two overlapping sources, the harder case motivating the whole approach.

Catalog record

CREATORTanguy Delmond, Natalia Korsakova, Thomas Oberlin, Sylvain Marsat, Antoine Basset, Nicolas Dobigeon
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE89 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.29039

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Neural posterior estimation of Galactic Binary signals for the LISA mission Tanguy Delmond, Natalia Korsakova, Thomas Oberlin, Sylvain Marsat, Antoine Basset, Nicolas Dobigeon Abstract ESA’s LISA mission will open a new window onto the gravitational-wave sky by detecting signals from a wide variety of sources in the millihertz frequency band. Among these, galactic binaries are expected to be the most numerous sources observable by LISA. Their analysis and parameter estimation represent a significant challenge, as the signals are expected to strongly overlap in both the time and frequency domains. Conventional Bayesian inference approaches, such as Markov Chain Monte Carlo sampling, are difficult to scale to this setting due to the high dimensionality of the problem and the complicated likelihood landscape which can hinder convergence. In this work, we explore simulation-based inference as a means to perform efficient parameter estimation for single galactic binaries, with a potential extension to the analysis of multiple overlapping sources. Our approach relies on a conditional normalizing flow acting as a neural posterior estimator. The model is trained using samples generated according to a dedicated simulation framework that does not require any likelihood computation. Once trained, the neural posterior estimator enables the generation of thousands of posterior samples per second, again without explicit likelihood evaluation. We first present results for a single source in a narrow frequency band, and then extend the analysis to wider frequency ranges. As a proof of concept, we further investigate the more challenging case of two overlapping sources. These results demonstrate the potential of likelihood-free inference as a scalable alternative to conventional Markov chain Monte Carlo sampling for the analysis of LISA galactic binaries. I Introduction Since the first detection in September 2015 3 , gravitational waves (GW) from the coalescence of compact-object binaries are now routinely observed by ground-based interferometers such as Advanced LIGO 2 , Advanced Virgo 4 , and in the future by KAGRA 7 . However, these detectors are fundamentally limited at high frequencies by terrestrial noise and armlength of the detectors, which prevents the observation of GW sources with total masses much larger than a few hundred solar masses. The Laser Interferometer Space Antenna (LISA) 12 ; 6 , a space-based GW observatory led by ESA in collaboration with NASA and scheduled for launch in the 2030s, will overcome this limitation by operating in the millihertz frequency band, well below the sensitivity range of ground-based detectors. This will enable the observation and precise characterization of a wide variety of GW sources, including massive black hole binaries (MBHBs) with masses up to millions of solar masses 12 ; 34 , stellar-mass black hole binaries, extreme mass ratio inspirals (EMRIs) 12 ; 8 , stochastic GW signal from the energetic processes in the early Universe 12 ; 10 and an exceptionally large population of Galactic Binaries (GBs) mainly composed of double white dwarf binaries but also some neutron stars or black hole binaries 35 . Most of these binaries will form a stochastic GW foreground 48 ; 12 , however, up to tens of thousands will be resolved and characterized individually 36 . This unprecedented census will enable a detailed and essentially unbiased mapping of compact binary populations across the Milky Way, providing valuable information on binary evolution, star formation history, and Galactic structure 12 ; 45 . As a result, GBs will constitute a cornerstone of LISA’s astrophysical program and a key probe of the Milky Way in GWs. The huge number and diversity of signals potentially observable by LISA places the data analysis at the very heart of the mission. In broad terms, gravitational-wave data analysis consists in transforming raw interferometric measurements into astrophysical information about the observed sources. This process unfolds across several key steps: the calibration and characterization of detector data, the detection and identification of gravitational-wave signals, and finally the estimation of the physical parameters of the sources like its sky localization or component masses of the system. LISA data analysis is particularly challenging due to the large and unknown number of sources overlapping in both the frequency and time domains, and the diversity of source types. In addition to the astrophysical diversity of its GW sources, LISA data will have non-stationary noise 26 ; 5 and be affected by a complex environment effects including gaps 18 , and glitches 44 ; 11 . Furthermore, the data stream will be dominated by GW signals, leaving virtually no signal-free intervals for estimating instrumental noise alone. The complexity of LISA data, with its high dimensionality, multiple source types, and intricate noise properties, necessitates a global fit approach 39 ; 17 ; 32 in which all source and noise parameters are estimated simultaneously. Implementing such an analysis requires a sophisticated pipeline, where dedicated modules for different source classes or instrumental models run in parallel and communicate throughout the process. Among all LISA sources, GBs present one of the most severe data analysis challenges. The core difficulty lies in the unknown number of sources and the fact that they gradually transition into the noise floor, making it extremely hard to identify sources at low SNRs — a problem commonly referred to as the confusion problem. Until now, most of the techniques proposed in the literature to estimate their parameters posterior, rely on Bayesian trans-dimensional methods, such as the Reversible-Jump Markov Chain Monte Carlo (RJMCMC) method 24 ; 21 ; 38 . Even setting aside the trans-dimensional aspect, standard Markov Chain Monte Carlo (MCMC) approaches are computationally demanding in this context due to the cost of performing millions of likelihood evaluations, which can also lead to slow convergence. These limitations motivate the exploration of alternative Bayesian inference strategies capable of delivering accurate posterior distributions more efficiently or complementary Bayesian tools including methods that can provide fast, accurate proposals to improve the efficiency of classical MCMC and RJMCMC samplers. Bayesian inference does not necessarily have to be performed with MCMC: simulation-based inference (SBI) offers a compelling alternative, and in particular a direct route to posterior distributions over source parameters. Among SBI techniques, neural posterior estimation (NPE), through Normalizing Flows (NFs), has been widely and successfully applied to GW data analysis, demonstrating that accurate, full posterior distributions can be obtained orders of magnitude faster than with traditional samplers. In the context of ground-based detectors, the DINGO pipeline enables fast and accurate posterior estimation for compact binary mergers 25 ; 16 . NFs have also been explored for LISA sources, including attempts at targeting MBHBs via sequential neural likelihood 43 and NPE 50 , as well as within the DINGO framework. These developments are partly motivated by the need for rapid parameter estimation for sources with expected electromagnetic counterparts, but the computational benefits are relevant more broadly. In parallel, work on extending SBI methods to trans-dimensional settings has begun, with approaches that start from simpler source classes and progressively address the unknown-number problem 29 . Here, we follow a complementary strategy: rather than starting from simpler toy sources, we begin by fitting a representative isolated GB signal as observed by LISA, directly targeting the full posterior distribution over the parameters of a single GB, with the goal of building toward a trans-dimensional treatment in subsequent and future works. Even at this

Read the complete text