Injection runs cannot track backgrounds
Done statement: We can use up-to-date backgrounds throughout an injection run, especially for 2h and 1d time scales.
During an injection run our background collection gets polluted by the signals. Currently we get around this by using reference backgrounds, generated over a one week (BBH) or two week (BNS, NSBH, or BBH in noisy data) long run(s).
As the injection run itself cannot reliably track and update those backgrounds, the shorter background timescales (2h and 1d) quickly become inaccurate.
This is because when we load our "reference background" (marginalized stats from a background seed run), they remain static throughout the duration of the injection run. Therefore any far_2h or far_1d value output by the pipeline during an injection run with seeded backgrounds will have stale 2h or 1d FAR estimates as it will only consider the backgrounds calculated for the last 2h and 1d of the original 2 week background run - it will not update and be "tracked" during the newer injection run.
Possible Solutions
- The project to remove signals from backgrounds could improve results: #61
-
@manoj.kovalam suggests running a non-injection and injection run in parallel, with the non-injection run tracking backgrounds, and the injection run using its backgrounds as a reference.
- They'll fall out of sync with each other as the injection runs have to process more data. However, it will leave backgrounds out by a matter of hours instead of a week.
- It won't be deterministic, but may still be better for science results.
- On a run with live-streamed data the two runs should remain in sync.
Note this will be simpler once #86 is completed, so that we can more easily track & use reference backgrounds at the same time.