Parameter Estimation in Astronomy with Poisson-Distributed Data. I. The Chi-Square-Gamma Statistic

Abstract

Applying the standard weighted mean formula, [sum_i {n_i sigma^{-2}_i}] / [sum_i {sigma^{-2}_i}], to determine the weighted mean of data, n_i, drawn from a Poisson distribution, will, on average, underestimate the true mean by ~1 for all true mean values larger than ~3 when the common assumption is made that the error of the ith observation is sigma_i = max(sqrt{n_i},1). This small, but statistically significant offset, explains the long-known observation that chi-square minimization techniques which use the modified Neyman's chi-square statistic, chi^2_{N} equiv sum_i (n_i-y_i)^2 / max(n_i,1), to compare Poisson-distributed data with model values, y_i, will typically predict a total number of counts that underestimates the true total by about 1 count per bin. Based on my finding that the weighted mean of data drawn from a Poisson distribution can be determined using the formula [sum_i [n_i + min(n_i,1)] (n_i+1)^{-1}] / [sum_i (n_i+1)^{-1}], I propose that a new chi-square statistic, chi^2_gamma equiv sum_i [n_i + min(n_i,1) - y_i]^2 / [n_i + 1], should always be used to analyze Poisson-distributed data in preference to the modified Neyman's chi-square statistic. I demonstrate the power and usefulness of chi-square-gamma minimization by using two statistical fitting techniques and five chi-square statistics to analyze simulated X-ray power-law 15-channel spectra with large and small counts per bin. I show that chi-square-gamma minimization with the Levenberg-Marquardt or Powell's method can produce excellent results (mean slope errors <=3%) with spectra having as few as 25 total counts.