Add support for waveform uncertainty #475

Open bilby-bot opened 5 years ago

In GitLab by @git.ligo:daniel-williams on Oct 3, 2019, 10:51

With the appearance of new models (like heron) which are capable of producing not only a waveform but also a standard deviation on that waveform it would be nice for bilby to have a way to incorporate this extra information in inference.

This should be possible to include as an additional PSD term in the calculation of match between the model.

As far as I can tell this would require updates to the bilby.gw.waveform_generator.WaveformGenerator object, to allow the additional information to be collected from the model, and also an overloaded class based on bilby.gw.likelihood.GravitationalWaveTransient to calculate the SNR and then likelihood with the adapted PSD.

I'm happy to take point on this, but I wanted to get some input about whether you'd rather see this implemented as an update to the existing classes, as new classes inheriting the existing ones, or indeed, if you'd rather this was functionality I added to the heron package instead.

In GitLab by @git.ligo:colm.talbot on Oct 3, 2019, 22:57

Since this is really a modification to the template, I wonder if it makes more sense to treat this closer to how calibration models are included than in the PSD, e.g., https://git.ligo.org/lscsoft/bilby/blob/master/bilby/gw/detector/interferometer.py#L321?

In GitLab by @git.ligo:colm.talbot on Oct 3, 2019, 23:04

Sorry to spam on this, but is the plan to marginalise over the uncertainty in the likelihood evaluation, or just to randomly sample from the uncertainty in the model? If it's the latter, it may be as simple as just calling the waveform uncertainty a calibration model and the likelihood could remain the same.

If you're planning on marginalising inside the likelihood, I imagine that will be a significant enough change that we should make it a new likelihood class inheriting from the GWT.

In GitLab by @git.ligo:daniel-williams on Oct 4, 2019, 10:26

Hmm, this is a good question. There are two approaches to dealing with the uncertainty, depending on the form of the model.

A time-domain model which doesn't use a reduced basis of some sort should be amenable to the approach of being included as an additional PSD term in the noise-weighted inner product.
A model which produces a non-analytical estimate of the waveform distribution (so effectively, non-Gaussian, or non-t-distributed, depending on the exact elliptical process used) would need to be sampled; the likelihood at a given waveform parameter would then need to be the sum of the individual waveform draws' likelihood.

The latter situation happens if you combine the outputs of two elliptical processes, so if you model amplitude and phase rather than hoft, for example.

While I appreciate the attraction of including this like a calibration term, I'm unhappy about the semantics of this (I think) as it's not strictly a calibration uncertainty. This said, as I type this I'm becoming less convinced of this conviction.

In GitLab by @git.ligo:gregory.ashton on Apr 25, 2020, 02:15

I'm going to remove the "doing" label as I think it is fair to say it is not being actively worked on. Please retag it if this is incorrect.

In GitLab by @git.ligo:daniel-williams on Apr 25, 2020, 12:10

I've changed this back. I am in fact actively working on this.

In GitLab by @git.ligo:michael.williams on Oct 3, 2024, 17:55

unassigned @git.ligo:daniel-williams

bilby-dev / bilby

Add support for waveform uncertainty #475