1
$\begingroup$

Let's say I have an observable $X$ that depends on a parameter $\theta$, and that I can find an expression for the Cramér-Rao lower bound for estimating $\theta$, as a function of $X$: $\sigma_{CRB}(X)$, by assuming I can exactly measure $X$. Now let's say I cannot measure $X$ exactly, and instead it is normally distributed. Now it seems to me there should be a way to work this into the "correct" Cramér-Rao bound, is this correct and how would it be done?

$\endgroup$
1
  • $\begingroup$ What about if $X$ is normally distributed? (I updated the question to signify that) $\endgroup$ Apr 17 '20 at 11:11
2
$\begingroup$

Recall the Fréchet-Darmois-Cramér-Rao lower bound conditions [reproduced from Wikipedia]:

Suppose $\theta$ is an unknown deterministic parameter which is to be estimated from $n$ independent observations (measurements) of $X$, each distributed according to some probability density function ${\displaystyle f(x;\theta )}$. ). The variance of any unbiased estimator ${\displaystyle {\hat {\theta }}_n}$ of ${\displaystyle \theta }$ is then bounded by the reciprocal of the Fisher information ${\displaystyle I(\theta )}$ associated with $n$ iid observations: $${\displaystyle \operatorname {var} ({\hat {\theta }_n})\geq {\frac {1}{n I(\theta )}}}$$

Therefore, to apply in this framework when observing $y_i=X_i+\epsilon_i$,

  1. There must exist an unbiased estimator based on the $Y_i$'s, which is not an immediate consequence of an existing unbiased estimator based on the $X_i$'s, unless the estimator is linear in the $X_i$'s, like $\hat{\theta}_n=\bar{X}_n$;
  2. The Fisher information is the Fisher information associated with the distribution of $Y$, rather than the Fisher information associated with the distribution of $X$. Deriving one from the other is rarely feasible.
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.