Finding periodicities in astronomical light curves using information theoretic learning

The analysis of time-variable astronomical phenomena is of great interest as it helps to improve our understanding of the structure and topology of our Universe, the mechanisms of galaxy and stellar evolution, etc. The basic tool to study variability in the sky is the light curve. Light curves are time series of stellar brightness and their analysis reveals key information about the physics behind the variable phenomena. Periodic variable stars are particularly interesting. Periodic variable stars are used to estimate the size and distance-scales of our Universe, and the period is a key parameter for stellar parameter estimation, stellar classification and exoplanet detection. The precise estimation of the period is critical in order to accomplish these scientific tasks. Astronomy is experiencing a paradigm change due to the extent volumes of data generated by current astronomical surveys. In less than 10 years, hundreds of Petabytes of astronomical images and time series catalogs will be produced. Conventional astronomy does not possess the tools required for this massive data mining operation. Nowadays there is a growing need for methods with solid statistical background to do automatic astronomical time series analysis. These methods need to be robust, fully-automated and computationally efficient. In this doctoral research I developed methods for periodicity detection and period estimation in light curves that are based on information theoretic concept of correntropy and advanced signal processing techniques. These methods are intended for automatic and efficient periodic light curve discrimination in large astronomical databases. Correntropy is a generalization of the conventional correlation to higher order statistics. In this thesis I propose the slotted correntropy estimator, the correntropy kernelized periodogram (CKP) and the correntropy non-negative matrix factorization spectrum (CNMFS). The slotted correntropy extends correntropy to unevenly sampled time series such as light curves. The CKP is a generalized periodogram that can be computed directly from the samples without regards on their sampling.
[Santiago: 2011]

Compartir: