2015 BFY II Abstract Detail Page

Previous Page  |  New Search  |  Browse All

Abstract Title: Maximum Likelihood is Least Squares (and not just for Gaussian random variables)
Abstract: One often hears that a weighted, least squares fit produces maximum likelihood parameters only if the measured y-values are governed by Gaussian distributions.  Here we show that iteratively reweighted least squares (IRLS) also produces maximum likelihood parameters for y-values governed by Poisson, binomial and exponential distributions, whose standard deviation (or variance) depends on the distribution mean.[1]  The mathematics for this assertion are demonstrated and show a surprising appearance of the variance of the distribution involved.  The IRLS algorithm starts with initial estimates for the fitting parameters from which an initial set of fitted y-values is calculated.  The variances are then determined based on the assumed distribution type and using the fitted (not measured) y-values for the distribution mean.  Minimizing the chi-square keeping the variances fixed generates an improved fit and new variances.  IRLS uses each prior fit to calculate variances for the next and iterations continue until self-consistent.  Keeping the standard deviations fixed during each least squares fit, ultimately using their values as calculated at the best fit, guarantees the final fit parameters will be maximum likelihood estimates.
[1] A. Charnes, E.L. Frome and P.L. Yu, "Equivalence of Generalized Least Squares and the Maximum Likelihood Estimates in the Exponential Family," J. of the Amer. Stat. Assoc., 71 (1976) 169-171
Abstract Type: Poster

Author/Organizer Information

Primary Contact: Robert DeSerio
University of Florida
Department of Physics
Gainesville, FL 32611
Phone: (352) 392-1690

Presentation Documents

Contributed Poster: Download the Contributed Poster