|
|
|
# Likelihood reweighting
|
|
|
|
|
|
|
|
Likelihood reweighting was used in O3 to include calibration marginalization for RIFT.
|
|
|
|
Since then we implemented a more generic version within `bilby_pipe`.
|
|
|
|
|
|
|
|
The other primary use case being considered for O4 is importance sampling from an accelerated (e.g., ROQ) likelihood to the regular likelihood to undo the effect of the approximation.
|
|
|
|
|
|
|
|
## `bilby_pipe` implementation (https://git.ligo.org/lscsoft/bilby_pipe/-/merge_requests/493, https://git.ligo.org/lscsoft/bilby_pipe/-/merge_requests/508, https://git.ligo.org/lscsoft/bilby_pipe/-/merge_requests/522, https://git.ligo.org/lscsoft/bilby_pipe/-/merge_requests/509)
|
|
|
|
|
|
|
|
Importance sampling (reweighting) can be specified to be performed after sampling is complete as part of the main pipeline.
|
|
|
|
The setup is specified either inline or in a separate `json` file.
|
|
|
|
Any argument for the likelihood/waveform generation or the prior can be changed as part of this.
|
|
|
|
|
|
|
|
Prior changes must retain the same parameters with the exception of adding marginalization over calibration uncertainty.
|
|
|
|
|
|
|
|
For some nested samplers, the nested samples can be reweighted directly.
|
|
|
|
This can lead to better efficiency when the new likelihood differs significantly from the initial model.
|
|
|
|
|
|
|
|
This is documented [online](https://docs.ligo.org/lscsoft/bilby_pipe/master/reweighting.html) and is run as part of the continuous integration.
|
|
|
|
|
|
|
|
## `Bilby` implementation (https://git.ligo.org/lscsoft/bilby/-/merge_requests/776, https://git.ligo.org/lscsoft/bilby/-/merge_requests/1234)
|
|
|
|
|
|
|
|
The original implementation was done a long time ago and signed off by Matt Pitkin.
|
|
|
|
The only change since then is adding parallelization to improve efficiency. |
