James Wicker (NAOC): New Methods of Univariate Mixture Modeling with Applications to Astronomy

Fri, Nov. 12, 2010, 3:00 p.m. (NAOC Seminar Room A308)

Since the early 1990s, researchers in Astronomy have used the KMM Mixture Model algorithm to analyze their univariate distributional data, for example in studies of Galactic and Globular Cluster metallicity, as advocated by Ashman, Bird and Zepf (AJ, 1994). However, this method has some serious handicaps which need to be addressed, including its basic inability to identify more than two distributions in a data set. Hence, although a data set might have more than two component distributions, KMM would almost always conclude that the data set is only "bimodal." I will show how we can improve on KMM with some new Mixture Modeling algorithms and better identify the number of distributions with a technique based on Information Scoring in Statistical Analysis. I would also like to have an open discussion at the end of the talk to see how this method can be improved in terms of computational speed and data size, and also what other applications in astronomy this method might have.