Speaker
Description
The next generation of gravitational-wave observatories, such as the Einstein Telescope and Cosmic Explorer, will detect thousands of compact binary coalescences, placing unprecedented demands on parameter estimation pipelines. Standard Bayesian inference techniques, such as Markov Chain Monte Carlo (MCMC), provide accurate posterior distributions but are computationally expensive, often requiring ~100 CPU hours per event. This cost makes large-scale population analyses and real-time inference increasingly challenging.
Approximate methods, such as the Fisher matrix, offer a computationally efficient alternative but rely on the assumption of Gaussian posteriors, which breaks down in the presence of strong parameter degeneracies, multimodality, and other non-linear effects. In this work, we explore the Derivative Approximation for Likelihoods (DALI), a systematic higher-order extension of the Fisher approach, as a fast and accurate framework for gravitational-wave inference.
We present a comprehensive comparison between Fisher matrix, DALI (in its singlet, doublet, and triplet implementations), and full MCMC methods. We show that DALI captures non-Gaussian features of the posterior while achieving up to a factor of ~55 reduction in computational cost relative to standard MCMC. In particular, the singlet-DALI hybrid approach provides a compelling balance between accuracy and efficiency, significantly outperforming the Fisher approximation and reducing computational cost by an order of magnitude compared to higher-order DALI variants.
These results demonstrate that higher-order likelihood expansions offer a scalable solution for next-generation gravitational-wave data analysis. We also introduce the public release of the GWDALI code (v1.0), which integrates automatic differentiation, modern waveform models, and optimized parameter decompositions, enabling fast and reliable inference for large gravitational-wave datasets.