Today some actual work happened, although only a tiny bit! I spoke with Adam Wheeler (Columbia) about our project to find a local, low-dimensional representation of LAMOST spectra, which can then be used to find abundance outliers. The model is a linear latent-variable model, but executed locally among the K nearest neighbors in the training set near the test object. It is unsupervised, because it doesn't rely on any labels (except the knowledge of where in the spectrum each element has lines). I'm interested in whether approaches like this could be built up into a full abundance system, with clever self-calibration.
I met with Megan Bedell (Flatiron) who has been playing around with observing strategies for Terra Hunting Experiment. We would like to demonstrate that randomized observing strategies bring more information than regular observing strategies. But if we simulate the data as having white (uncorrelated) noise of known variance, the sensitivity of the data to planets of various periods depends only (or almost only) on the total exposure time and total survey duration, and barely on how the observations are scheduled. So what gives? I think the big deal is correlated noise: If the spectrographs drift in their calibration properties in some correlated way over time during the survey, then the regular data will alias those drifts into periodic signals. Or that's my conjecture. I'm not confident in it, but we can test it. Oh and by the way: You always have low-level correlated drifts in your instrument calibration, which are below the calibration precision but large enough that they can affect your results when you combine many thousands of exposures over many years.
No comments:
Post a Comment