Let $X = (x_1,\ldots,x_n)$ be an i.i.d sample from distribution $F%$ and let $y = \prod_{i=1}^n x_i$

Can we derive a randomized, unbiased. estimator $\hat{y}$ of $y$ that on average considers only a subsample of $X$?

Weak conditions may be imposed on $F$, for instance we may assume it has finite mean and variance, we may also assume that $\forall i, x_i >0$, but otherwise nothing is known about $F$.

  • $\begingroup$ What precisely do you mean? E.g., since $y$ is a random variable, it can't have an estimator (in standard statistical terminology), only a predictor. Should this be some function of the x_i or ("randomized") a Markov kernel? And what does "on average" mean here? $\endgroup$ – Lutz Mattner Sep 16 '14 at 14:28
  • $\begingroup$ I mean that $\hat{y}$ is a random variable such that $E(\hat{y}) = E(y)$, that there exists a sampling procedure for $\hat{y}$ that examines a random subsample of $X$, and that this subsample has on expectation less than $n$ elements. $\endgroup$ – Arthur B Sep 16 '14 at 14:39
  • $\begingroup$ And the answer to my problem seems to be the generalized poisson estimator. $\endgroup$ – Arthur B Sep 16 '14 at 14:40

The solution is the Poisson estimator. Let $a_i$ = $\log x_i$. We would like to approximate $\exp (\sum_{i=1}^n a_i) = \exp (n \hat{a})$

Draw $\kappa \sim \textrm{Poisson}(\lambda)$, then draw with replacement $a_{i_1}, \ldots, a_{i_\kappa}$ and compute $y = e^{\lambda}\prod_{j=1}^{\kappa} n a_{i_j}$

$$E(y) = \sum_{k=0}^{\infty} \frac{1}{k!} E\left(\prod_{j=1}^{k} n a_{i_j}\right) = e^{n\hat{a}}$$


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.