Parameter estimation for von Mises-Fisher distributions (Q2271704)

The von Mises-Fisher distribution supported on the unit hypersphere in \(R^d\) is defined by the PDF \(p(x)=C_p(\kappa)e^{\kappa\mu^Tx}\), where \(\mu\) is the mean vector (\(\| \mu\| =1\)), \(\kappa\) is the concentration parameter, and \(C_p\) is a normalizing constant. The authors propose a new algorithm for calculating the maximum likelihood estimator \(\hat \kappa\) of \(\kappa\). This algorithm is iterative and based on a fixed point approach. Simple straightforward approximations for \(\hat\kappa\) are considered. A special approximation of \(\log I_p(x)\) for large \(p\) is proposed, where \(I_p(x)\) is the modified Bessel function of the first kind of order \(p\). This function is used for calculation of the normalizing constant \(C_p\).