Definition

F(\theta) = g_{ij}(\theta) &:= E_{\theta}\left[ \frac{\partial\ln L(\theta|x)}{\partial \theta_{i}}\frac{\partial\ln L(\theta|x)}{\partial \theta_{j}} \right] = \int_{\mathbb{R}} \frac{\partial\ln L(\theta|x)}{\partial \theta_{i}}\frac{\partial\ln L(\theta|x)}{\partial \theta_{j}} p(x, \theta)dx\\ &= -E_{\theta}\left[ \frac{\partial^{2}\ln L(\theta|x)}{\partial \theta_{i}\theta_{j}} \right] = -\int_{\mathbb{R}} \frac{\partial^{2} \ln L(\theta|x)}{\partial \theta_{i}\theta_{j}} p(x, \theta)dx \end{aligned}$$ by the [[Bartlett Identities#second-bartlett-identity|second Bartlett identity]] # Facts ## Monotonicity of Fisher Information Metrics > Suppose a [[Probability Space]] $(\Omega, {\cal F}, P)$, a [[Random Variable]] $X: \Omega \rightarrow \mathbb{R}$, and Borel measurable function $T: \mathbb{R} \to \mathbb{R}$ > Then, the fisher information metrics of two random variables $X, T\circ X$ hold relation: > $$I_{T \circ X}(\theta) \leq I_{X}(\theta)$$ > In other words, $\forall v \in \mathbb{R}_{n}, v^{\intercal}I_{T \circ X}(\theta)v \leq v^{\intercal}I_{X}(\theta)v$ > where $I_{X}(\theta) := E_{\theta}\left[ \left( \frac{\partial\ln f_{\theta}}{\partial \theta} \right)^{2} \right]$, $I_{T\circ X}(\theta) := E_{\theta}\left[ \left( \frac{\partial\ln \phi_{\theta}}{\partial \theta} \right)^{2} \right]$ and $f_{\theta}, \phi_{\theta}$ are the [[Density Function#induced-dessity-function|induced density functions]] by the $X$ and $T\circ X$ respectively. > Equality holds if and only if $T$ is [[Sufficient Statistic]] > $$E\left[ \frac{d \ln f_{\theta}}{d\theta}|T=t \right] = \frac{d\ln\phi_\theta(t)}{d\theta}$$