Reparametrization in regards to Fisher information

235 Views Asked by At

I'm trying to understand equation 5.11 from Lehmann and Casella's Theory of Point Estimation (2nd edition) which is presented without proof.

It states that if $I(\theta)$ represents the Fisher information that $X$ contains about the parameter $\theta$ and $\theta = h(\xi)$ where $h$ is a differentiable function, then $I^*(\xi)=I(h(\xi))*(h'(\xi))^2$ where $I^*(\xi)$ represents the Fisher information that $X$ contains about the parameter $\xi$. In this case, $\theta$ and $\xi$ are both scalars (I know this can be generalized to the case where both parameters are vectors, but I'm not worried about that part yet).

I don't see exactly why this is true. I'm guessing it has something to do with the chain rule, but I tend to get lost in the notation. I suppose we can write $I(\theta) = I(h(\xi)) = \int ( \frac{p'[f(x,h(\xi))]\frac{df}{dh(\xi)}}{p[f(x,h(\xi)]} )^2 p[f(x,h(\xi)] dx$

Note that $\frac{p'[f(x,h(\xi))]\frac{df}{dh(\xi)}}{p[f(x,h(\xi)]}$ is the derivative of the score function for density $p_{\theta}(x)$

Then $I(\theta)*(h'(\xi))^2= \int ( \frac{p'[f(x,h(\xi))]\frac{df}{dh(\xi)} \frac{dh(\xi)}{d\xi}}{p[f(x,h(\xi)]} )^2 p[f(x,h(\xi)] dx = \int ( \frac{p'[f(x,h(\xi))]\frac{df}{d\xi}}{p[f(x,h(\xi)]} )^2 p[f(x,h(\xi)] dx$

But this isn't $I^*(\xi)$. Or is it? Or is my understanding/notation just totally backward? I'd really appreciate any insight that you all may have in regards to this problem.