Gravitational-wave inference

Gravitational-wave specific proposals and reparameterisations are implemented in an additional package nessai-gw. This can be installed using pip or conda:

condapip

conda install -c conda-forge nessai-gw

Warning

Older versions of nessai included a .gw submodule but this is now deprecated in favour of nessai-gw.

Basic configuration

nessai includes a default configuration for gravitational-wave inference. This includes default reparameterisations for common parameters. This functionality is enabled by adding flow_proposal_class='gwflowproposal' to the keyword arguments for the sampler.

Marginalisation

Distance marginalisation

In the original paper we tested nessai with and without distance marginalisation and found that it can work with either. With the updated defaults, sampling with distance marginalisation is typically faster.

Phase marginalisation

The current defaults for nessai are not well suited to sampling without phase marginalisation and these result in inefficient sampling. We believe that revised settings for this case will help and we are actively investigating this.

Time marginalisation

Time marginalisation as implemented in bilby has been tested and results show that it can reduce sampling time without negatively impacting convergence.

The time_jitter parameter that is added by bilby is treated as a periodic parameter by default, however it is not clear is this treatment is necessary.

Calibration parameters

Basic tests including calibration parameters have been conducted and no major issues have been identified. Sampling is, as expected, slower but the overall convergence does not seem to be affected.

These parameters typically use Gaussian priors without hard prior bounds. By default, this is not an issue because the NullReparameterisation is used. However, if using a custom set of reparameterisations this should be accounted for.

Examples

Examples of running nessai for gravitational-wave inference can be found in the GW examples directory here.