Dip update 49/n
[Orig: Aug 5, 2017]
Past few days have been very productive at LCO. First, some notes on the data pipeline, and the aforementioned "detrending".
Background: A second set of 40cm telescopes became available at TFN and OGG at the end of June. It wasn't too long after when we found that one of the new cameras had a hot pixel landing on/near our science target. This is bad because the way we add up all the light in a circular aperture around the star to measure the star's brightness. If you have a hot pixel within this circular aperture, this contributes to your count, but this contribution is not real (aka, from the star), so your measurements will be all goofed up.
After discussing with LCO science support, we went for the easy fix and changed our LCO scheduling requests to apply a shift in the telescope pointing void this hot pixel interfering with the data. We were initially happy with this result, but after a couple weeks we noticed that at the same time we changed the request to shift the pointing, some of the comparison stars had shifted ever so slightly dimmer or brighter compared to the pre-pointing shift observations. We confirmed that this shift in comparison star brightness was not real, i.e., astrophysical, by having independent checks with other data sets to confirm they were constant (courtesy of B. Gary @HAO and J. swift @Thacher). So this indicates that the shift in comparison star brightness was due to the change in telescope pointing: our target (and the comparison stars) were now falling on a different part of the detector. In a perfect world, the variable response across the detector should be mitigated with standard calibration frames (darks, flats, and biases). In practice, it does a good job, but at not the precision level we desire. This introduced our need to "detrend" the data taken during this time (the issue we are experiencing is somewhat analogous to a "meridian flip" (Google it for details) within an observing session.).
For now, we have decided to detrend using a very simple linear model that is dependent on the target's (X, Y) position on the chip. This choice on detrending parameters (X,Y position) was determined most effective based on the considerable improvement in the Bayesian information criterion (BIC) for the comparison stars. This solution is acceptable, but not necessarily the best. So, long story short, LSU graduate student Tyler Ellis is working on the best way to treat the data that needs detrending using Gaussian Processes, which essentially will allow for corrections using little or no assumptions from the user. When we have a solution we are happy with (and all methods converge), he will post a write up on the process.
In the meantime, as of July 24, 2017 we decided to change the pointing back to the "normal" pre-shift position, which can be handled in the regular data pipeline. This means we have meaningful, science quality data immediately. The data taken during July with the pointing shift will eventually be at this level, but we need to work out the details described above. Going back to the pre-shift pointing also means that there is a chance for the camera to land the target on a hot pixel (the original issue), but we have found the actual amount of unusable data due to this is only ~5%, a loss that is acceptable for the time being given the pros of this strategy.
To sum it up, the graph below is data taken from May to today. It shows the data using the simple linear detrending on X,Y position, and to mitigate any lingering trends, we also only use comparison stars close in proximity to the science target on the camera. NB, the pointing shift happened from ~50-80 on the x-axis on this graph. Data to the left of ~50 and to the right of ~80 are taken using the same settings.
Are we going for a dip? Whatever you do take care of your shoes! #PhishQuotesForScience.
~Tabby and team