notes.org

To Do

REVIEW Inference interface

There seem to be problems with the inference interface. There seems to be around 3 orders of magnitude difference between the likelihoods with and without the waveform uncertainty, with the waveform uncertainty bringing log-likelihoods into the region of -4000 c.f. -4. It seems highly likely that this is a problem with the normalisation of the likelihood.

For the likelihood with waveform uncertainty the (additive) log-normalisation is implemented as

normalisation = torch.sum(torch.log(psd)) - torch.log(torch.prod(psd / psd.max()) + torch.prod(variance / variance.max())) + torch.log(psd.max())*len(psd) + torch.log(variance.max())*len(variance)

but for the case without the uncertainty it is implemented as

normalisation = 0

I suspect this is the root of the problem!

In the case of no waveform uncertainty we can just treat this as if this quantity is zero, and so

normalisation = torch.sum(torch.log(psd)) - torch.log(torch.prod(psd / psd.max())) +  torch.log(psd.max())*len(psd)

This seems to have fixed the problem, and the two likelihoods now look reasonably sensible. This needs to be reviewed though.

REVIEW Open questions

• [ ] Is FFTing the time-domain covariance matrix the equivalent of a 1- or 2-sided PSD?

Documentation

Add class diagram to the documentation

DOING Antenna functions

• [ ] Need two polarisations
• [ ] How does the mixing of uncertainties work

Bilby interface

• Probably just a wrapper around our own likelihood function.

Considerations for Tests of GR

• Tests of GR generally assume that an approximant is correct how will this affect things.

Considerations for population inference

• What effect will the systematics have on this?

Check the lowest possible mass for model

• Defined by NR training waveforms

Inference against GW150914

Inference with analytical models as training data

• Using two models to train a GPR in order to estimate the overall uncertainty
• e.g. comparison in the (SEOBNR?) paper

GW190412

• High mass ratio

GW190521

• Very high total mass

GW190727

Software injections

Integrate additional waveform catalogues