Bayesian Techniques for Finding and Characterizing Variable Stars: Application to the Hungarian-made Automated Telescope Surveys

Abstract

Wide-field variability surveys like HAT (Bakos et al., 2004, 2013), OGLE (Udalski

et al., 1994), Pan-STARRs (Chambers et al., 2016a), ASAS (Pojmanski, 1997; Jayas-

inghe et al., 2018c), and high-precision space-borne instruments such as Kepler (Koch

et al., 2010), Gaia (Gaia Collaboration et al., 2016), and CoRoT (CoRot Team, 2016)

have revolutionized the field of variable star astronomy. The future is still brighter —

LSST (LSST Science Collaboration et al., 2009) will increase the number of known

variable stars by over three orders of magnitude once the decadal survey completes

in 2029.

Analyzing variability in 10 billion objects requires a balance between statistical

rigor and computational speed. This thesis presents the development and exploration

of several techniques for detecting and characterizing variable stars in time-series

photometry applied to data from the HATSouth (Bakos et al., 2013) automated

survey.

The first chapter develops a scalable algorithm for fitting periodic templates to

noisy, irregularly-sampled timeseries. Template fitting is important for finding low-

amplitude signals, as well as for finding and characterizing signals in very sparse

timeseries. Recently (Sesar et al., 2016) used template fitting to successfully charac-

terize O(10 5 ) RR Lyrae in Pan-STARRs DR1 photometry, where O(10) observations

per ugrizy filter make traditional period finding techniques less sensitive.

The second chapter presents a study of RR Lyrae observed by HATSouth (Bakos

et al., 2013). We crossmatch with several external variable star catalogs to obtain

several thousand RR Lyrae, and then use a Random Forest classifier to find hundreds

of new RR Lyrae in HATSouth. We perform a careful Bayesian analysis of all RR

Lyrae lightcurves to characterize double mode pulsations and the Blazhko effect.

The third chapter develops a semi-supervised framework for performing automated

variable star classification in the presence of sparse, noisy labels. We demonstrate

that these techniques are capable of learning accurate classifications despite only

using three labeled examples per 7 classes during training. This approach is also ca-

pable of handling non-Euclidean features such as the phase-invariant shape of periodic

lightcurves.