Change reference distance and interpolant behavior

When setting the minimum of the distance prior to extremely small values (like 1 Mpc to match the LALInference behavior) this leads to huge values of the reference SNRs stored in the lookup table for distance marginalization. The interpolation of the lookup table was also not verifying that the interpolated values are inside the bounds for which it was initialized, leading again to huge values of the SNRs and likelihood. This MR changes the reference distance to be the median of the prior instead of the minimum, and throws an error when the interpolation tries to exceed the bounds.

added 1 commit

• 977aaaba - Add catch for bounds_error

Compare with previous version

resolved all threads

added 1 commit

• 4cd5ed90 - Change to fill_value

Compare with previous version

unmarked as a Work In Progress

@sylvia.biscoveanu this indeed seems to help with the dlogz=inf problem observed in PhenomXPM runs!

• I ran 10 seeds of a previously failing config with this week's (Tuesday's?) unmodified master versions ['bilby_version': '1.0.0 3c93842c', 'bilby_pipe_version': '1.0.0 9a35107f#, 'parallel_bilby_version': '0.1.4 0703664'] and with distance prior minimum set to 100, and still 3/10 failed. So just increasing that minimum doens't help.
• But then applying a patch made from this MR as the only change, I ran another 8 seeds and none failed.

This isn't conclusive statistical proof, but before we saw on average 1/4 seeds fail, so I'm pretty optimistic about this.

Very related issue (I think) found in the distance-marginalisation with LALInference lalsuite!819 (closed)

( ping @john-veitch for reference)

approved this merge request

approved this merge request

changed milestone to %1.0.2

mentioned in commit 74a346d8

merged

mentioned in issue #525 (closed)

mentioned in merge request !875 (merged)