Uploaded July 2022 | Updated September 2026, 1 hour ago
Over the course of the LISA mission, we expect to compile a large catalogue of extreme mass ratio inspiral (EMRI) observations. Encoded in this catalogue is information regarding the underlying EMRI population, which may be used to constrain the astrophysical processes that govern their formation on cosmological timescales. Using Bayesian hierarchical inference techniques, we can extract this information, estimating the parameters for population models that best fit the observed data. However, the results of this inference are susceptible to selection biases that depend on EMRI detectability (parameterised by the selection function) which lead to incorrect recovery of population parameters. While the selection function is typically simplified to lower-dimensional analogues due to its high computational cost, this itself can introduce further biases in the recovered parameters that are challenging to quantify.
In this talk, we introduce a new approach to EMRI detectability estimation in which the selection function is learned directly using neural network interpolation, producing selection function estimates as a function of population parameters. This is achieved by training a neural network on the EMRI SNR function, which achieves sub-percent-level accuracy across the EMRI parameter space. By evaluating the SNR function across a candidate population, the selection function may be estimated in milliseconds due to the vectorized nature of the interpolator network. As an additional pre-processing step, we draw samples from the population prior volume, compute selection function estimates with our primary interpolator network and train a secondary neural network on these estimates; this process further improves the speed of selection function estimation during sampling without compromising on accuracy.
We validate our method for a representative EMRI population and successfully correct for selection biases in the recovered population parameters without significantly increasing the inference runtime. Our approach is agnostic to the chosen waveform/population model and is therefore easily adaptible to a variety of population inference problems with different waveform models.
Authors: Christian Chapman-Bird, Christopher Berry, Graham Woan
Presenter: Christian Chapman-Bird
Over the course of the LISA mission, we expect to compile a large catalogue of extreme mass ratio inspiral (EMRI) observations. Encoded in this catalogue is information regarding the underlying EMRI population, which may be used to constrain the astrophysical processes that govern their formation on cosmological timescales. Using Bayesian hierarchical inference techniques, we can extract this information, estimating the parameters for population models that best fit the observed data. However, the results of this inference are susceptible to selection biases that depend on EMRI detectability (parameterised by the selection function) which lead to incorrect recovery of population parameters. While the selection function is typically simplified to lower-dimensional analogues due to its high computational cost, this itself can introduce further biases in the recovered parameters that are challenging to quantify.
In this talk, we introduce a new approach to EMRI detectability estimation in which the selection function is learned directly using neural network interpolation, producing selection function estimates as a function of population parameters. This is achieved by training a neural network on the EMRI SNR function, which achieves sub-percent-level accuracy across the EMRI parameter space. By evaluating the SNR function across a candidate population, the selection function may be estimated in milliseconds due to the vectorized nature of the interpolator network. As an additional pre-processing step, we draw samples from the population prior volume, compute selection function estimates with our primary interpolator network and train a secondary neural network on these estimates; this process further improves the speed of selection function estimation during sampling without compromising on accuracy.
We validate our method for a representative EMRI population and successfully correct for selection biases in the recovered population parameters without significantly increasing the inference runtime. Our approach is agnostic to the chosen waveform/population model and is therefore easily adaptible to a variety of population inference problems with different waveform models.
Authors: Christian Chapman-Bird, Christopher Berry, Graham Woan
Presenter: Christian Chapman-Bird
