2
$\begingroup$

Let $X_1....X_n$ be a random sample from the Poisson distribution with parameter $\lambda$. Find the uniformly minimum-variance unbiased estimator of

a) $\lambda^3$

b) $E(e^X)=e^{\lambda (e-1)}$

I'm having some difficulty with this. So far, I've found the function at the same time as the estimator using $a(\theta)[W(x)-\tau(\theta)]=\frac{\delta}{\delta \theta} Log(L(\theta|x)$ where W(x) is the estimator and $\tau(\theta)$ is the function.

Any assistance with a theorem or idea to begin here? Or can that be the right track?

$\endgroup$
  • $\begingroup$ I'm no expert in this, but what you did is proving that the score function can be written in efficient form. This means that $W(x)$ is an efficient estimator for $\tau(\theta)$, i.e. its variance is the smallest possible and reaches the Cramer-Rao lower bound. Now all you have to do is show that $W(x)$ is unbiased: $\mathrm{E}[W(x)|\theta]=\tau(\theta)$ and you're done,... I think. $\endgroup$ – StijnDeVuyst Mar 23 '15 at 23:49
11
$\begingroup$

There is one main strategy that I use for this type of problem. The key theorem is the Lehmann-Scheffe theorem which tells us that if we can get an unbiased estimator for $\tau(\theta)$ that is purely a function of a complete and sufficient statistic for $\theta$ [being a bit casual with definitions] then we have the unique BUE of our target.

The Poisson distribution is an exponential family so we know that $S := \sum\limits_{i=1}^n X_i$ is complete and sufficient for $\lambda$. We also know that $S/n$ is the MLE of $\lambda$.

For (1) we need to find an unbiased estimator of $\lambda^3$ that is only a function of $S$. We know by the invariance principle of the MLE that $S^3 / n^3$ is the MLE for $\lambda^3$ so we'll start there.

The $X_i$ are iid so we know that $S \sim Poisson(n\lambda)$ so its MGF is $M_S(t) = e^{n\lambda(e^t-1)}$. Computing the first three derivatives of this gives us that

$$ E(S^3/n^3) = \lambda^3 + 3\lambda^2 / n + \lambda / n^2. $$

We now need to eliminate the lower order powers of $\lambda$ and can do so by a linear combination of powers of $S$, giving us that

$$ E(\bar{X}^3 - \frac{3}{n}\bar{X}^2 + \frac{2}{n^2}\bar{X}) = \lambda^3. $$

Since the statistic $T(\vec{X}) = \bar{X}^3 - \frac{3}{n}\bar{X}^2 + \frac{2}{n^2}\bar{X}$ is purely a function of the complete and sufficient statistic $S$ it is the unique UMVUE for $\lambda^3$. Note that this requires $n$ to be non-random.

For (2) we will proceed by Rao-Blackwellization with $S$. This really is just Lehmann-Scheffe again but it looks a little different. The idea is that if $W$ is an unbiased estimator for $\tau(\theta)$ and $T$ is a sufficient statistic for $\theta$ then $\Phi(T) = \mathbb{E}(W|T)$ is often a better and never a worse estimator of $\tau(\theta)$ than $W$ alone. Note that $\Phi$ is only a function of $T$ so if $T$ is also complete then by Lehmann-Scheffe we have the unique UMVUE. So this is really just a different algebraic route to doing the same thing.

We've been handed our unbiased estimator by being told that $E(e^X) = e^{\lambda(e-1)}$ [note that this is $M_X(1)$].

This tells us that $\mathbb{E}(e^{X_1}|S)$ is our UMVUE [note that by the tower property we haven't changed its mean]. Let's work out what this is.

$$ \mathbb{E}(e^{X_1}|S = s) = \sum\limits_{x_1=0}^s e^{x_1} P(X_1 = x_1 | S = s) $$

$$ = \sum e^{x_1} \frac{P(X_1 = x_1 , S = s)}{P(S = s)} = \sum e^{x_1} \frac{P(X_1 = x_1) P(\sum\limits_{i=2}^n X_i = s - x_1) }{P(\sum\limits_{i=1}^n X_i = s)}. $$

Using the fact that the $X_i$ are iid and Poisson you should be able to work this out.

$\endgroup$

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.