Fisher's information in terms of the hazard rate (Q756322): Difference between revisions

Let \(\{g_{\theta}(t)\}\) be a regular family of probability densities on the real line R. Let T be a generic random variable with density \(g_{\theta}(t)\). Define the survival function corresponding to density \(g_{\theta}(t)\) by \[ G_{\theta}(t)=\int^{\infty}_{t}g_{\theta}(s)ds=\Pr ob_{\theta}\{T\geq t\} \] and the hazard rate for T by \(h_{\theta}(t)=g_{\theta}(t)/G_{\theta}(t).\) The authors prove that the Fisher information \(I_{\theta}\) can be expressed as \[ I_{\theta}=\int [((\partial /\partial \theta)g_{\theta}(t))/g_{\theta}(t)]^ 2g_{\theta}(t)dt,\quad \theta \in {\mathbb{R}}. \] This identity, in some sense, shows that the hazard rate transform has an ``isometric'' property which is important when we consider geometrical structures of a family of probability distributions. Among others, using this property, they compute curvature formulas of finite-dimensional statistical models and study the connection of this property with martingale theory.