Some statistical issues regarding the estimation of $\fnl$ in CMB non-Gaussianity

Abstract

We consider the problem of estimating the parameter $\fnl$ in the standard local model of primordial CMB non-Gaussianity. We determine the properties of maximum likelihood (ML) estimates and show that the problem is not the typical ML estimation problem as there are subtle issues involved. In particular, The Cramer-Rao inequality is not applicable and the likelihood function is unbounded with several points of maxima. However, the particular characteristics of the likelihood function lead to ML estimates that are simple and easy to compute. We compare their performance to that of moment estimators. We find that ML is better than the latter for values $\fnl$ away from the origin. For small values of $\fnl$, the fifth order moment is better than ML and the other moment estimators. However, we show how for small $\fnl$, one can easily improve the estimators by a simple shrinkage procedure. This is clearly important when the goal is to estimate a very small $\fnl$. In the process of studying the inference problem, we address some basic issues regarding statistical estimation in general that were raised at the Workshop on Non-Gaussianity in Cosmology held in Trieste in July 2006.