- 15D Gaussian
- Update on 0.5.3
- current list of O3 issues on git
Minutes Attendance: Matt, Simon, Carl, Charlie, Colm, Greg, Isobel, Paul, Shanika, Sylvia, Virginia
Matt has signed off on prior files
Waiting on 0.5.3 to get merged in, then re-run fiducial stuff, pp and known events
Calibration — Sylvia led: https://git.ligo.org/lscsoft/bilby_pipe/wikis/O3-review/calibration Looking through corner plots of amplitude & phase of two different detectors, with LALInference and Bilby posteriors, as well as priors for Bilby.
- different end times for the arrays between LI and Bilby. LI ends at Nyquist, Bilby at Nyquist - 1/duration. Difference is very small.
- p values for ks test are small.
- difference between LI and Bilby in last spline point. Bilby seems to be reproducing the prior, so difference may be due to the way the prior is being defined. Sylvia has ruled this out as being due to the position of the last spline point.
- PL: This is not something we want to hold up the review. Bilby seems to bbe doing the right thing, for all intents and purposes.
- MP: Yes, Bilby does seem to be doing the right thing. What are other tests than can bbe done that can figure this out? It would be interesting to understand regardless.
- CH: Partially agree with Paul; in terms of general physics that we care about, this is not super important to have being equivalent. However, on the fundamental level, we’re describing a Gaussian, and two Gaussian’s shouldn’t be different. One idea is to see how the different points are constructed. In LI, the prior is constructed by reading in the envelope, and then setting it onto a cubic-spline interpolant, which gives the values of the spline nodes.
- GA: suggest this goes offline as it’ll get technical. Two easy steps: a) add LALInference prior on those plots. b) what about making spline points the same as LALInference SB: Actually, that will require a large-scale code change. GA: Let’s take that offline. CT: Can we get each code to read in the calibration envelope file, and spit out the prior. Then we can see if the codes are behaving the same or different. MP: It’s worth double checking what went into the LI prior files. ie does it use a combination of MCMC and Nest SB: It’s only nest. CH: Agree with checking the priors. PL: Okay, then what? S’ppose LI also returns the prior, are we happy to sign off on review? MP: Yes, that should not hold up on review. It should be flagged, but no reason to hold off on review.
ACTION: Sylvia to figure out what is happening with LI priors.
15D Gaussian. Moritz on leave MP: Moritz has written a test for a single 15D Gaussian. This has been run. Moritz has also run a bi-modal 15D Gaussian. Output shows there’s a slight weighting between the two modes. MP has asked Moritz to run with a higher number of live points. Hopefully this will work, and then can be signed off. SP: Checking number of live points is a good idea. But presumably it would come down to the sampler, not an intrinsic Bilby issue. MP: Yes. We can easily check with other samplers if this is a problem.
GA: A quick walk through of the open issues so far. Once they’re done, 0.5.3 can be created, and then all stuff can be re-run.
Mismatch between phi_jl for Bilby and LI is actually a difference between O2 and O3 branches of LI. This means to test this, we need to re-run LI. Seb Kahn has provided scripts to re-run this. CT: Re-running sounds like a good idea. We should be able to match up the priors (limits for masses, distances, etc.). CT can do this, but would be nice to have a LI-person look over them once they’re created so we don’t waste CPU cycles. GA: If we’re to do this, we’re just going to do it for GW150914, and then throw it open for the reviewers. But, at some point, the O3 branch of LI is likely to be re-run on all these events. Is it worth doing this sooner rather than later? CH: There is no real plan to do those re-runs for O2 and O3 events. PL: How about running on a clean O3 event and get a phi_JL match there? CH, GA, SS: Agreed
ACTION: Decide on an O3 event (highest SNR), and run Bilby on that.
GA: There’s an issue with setting the waveform generator to -inf if it fails. Everyone agrees it’s a good idea, but we need to find someone. CT: Yes. Let’s also make sure we don’t hide any warning messages. CT to add to git.
GA: In summary, we have all the issues covered. Need to decide which event to choose, but other than that, everything is in flow. If people work on these soon, we can hopefully rerun everything in one or two weeks time.
CH: Fixed a convention difference between LI and Bilby in phi_1,2. This has been merged, and the match looks reasonable: https://ldas-jobs.ligo.caltech.edu/~charlie.hoy/pe/bilby_lalinference_comparison/GW150914/html/Comparison_phi_2.html there is still a residual difference CT: this is likely due to a difference between spins.
GA: Going to kick the likelihood issue down the road. It doesn’t effect results.
AOB: GA: Worth flagging the paper as a future action item. CH: At that point, we should also reconsider the choice of O3 event. PL: Actually, would probably only show GWTC-1 events, so not really an issue. CH: Agree. It’s only a convention. No physics difference.
CT: Add Bilby review to standing PE agenda. CH: Carl will add something in and send to random person to ensure nothing is missing.