Let $X_1,\dots, X_n\sim$iid Poisson with mean $\lambda$, and consider estimating $$ g(\lambda)=P_{\lambda}(X_i=1)=\lambda e^{-\lambda} $$ One natural estimator might be the proportion of ones in the sample: $$ \hat{p}_n=\frac{1}{n}\#\{i\le n: X_i=1\}=\frac{1}{n}\sum_{i=1}^nI[X_i=1] $$ Another choice would be the maximum likelihood estimator $g(\bar{X}_n)$ with $\bar{X}_n$ the sample average.

(1) Find the asymptotic relative efficiency of $\hat{p}_n$ w.r.t. $g(\bar{X}_n)$.

(2) Determine the limiting distribution of $n[g(\bar{X}_n)-1/e]$ under $\lambda=1$.

(1) The log-likelihood function is $$ l_n(\lambda)=-n\lambda+n\bar{X}_n\log\lambda-\log \prod x_i! $$ with $l''_n=-n\bar{X}_n/\lambda^2$.

So $$I(\lambda)=-E[-n\bar{X}_n/\lambda^2]=n/\lambda$$

But how about the $\hat{p}_n$?

Note that from CLT $$ \sqrt{n}(\hat{p}_n-E[I(X_i)=1])=\sqrt{n}(\hat{p}_n-\lambda e^{-\lambda})\to^d N(0,v) $$ where $v=Var(I(X_i)=1)=\lambda e^{-\lambda}-(\lambda e^{-\lambda})^2$.

(2) I have no idea about $1/e$...


Take $f(x)=xe^{-1}-1/e$ with $f'(x)=e^{-x}-xe^{-x}$ and $f''(x)=(x-2)e^{-x}$. Note that $f(1)=f'(1)=0$ and $f''(1)=-1/e$. Thus, $$ f(x)\approx \frac{1}{2}f''(1)(x-1)^2 $$ This means $$ n[g(\bar{X}_n)-1/e]=nf(\bar{X}_n)\approx n\frac{-1}{2e}(\bar{X}_n-1)^2\to \frac{-1}{2e}\chi_1^2? $$ I am confused here because $E X_1=\lambda$ and I cannot use CLT.


1 Answer 1


Suppose $\hat g_1(\lambda)=g(\overline X_n)=\overline X_ne^{-\overline X_n}$ and $\hat g_2(\lambda)=\frac1n\sum\limits_{i=1}^n I(X_i=1)$.

Provided $\lambda\ne 1$ (so that $g'(\lambda)\ne 0$ ), by delta method,

$$\operatorname{Var}(\hat g_1) \approx \frac{\lambda (g'(\lambda))^2}{n}=\frac{\lambda e^{-2\lambda}(1-\lambda)^2}{n} \quad , \text{ for large }n$$

And the exact variance of $\hat g_2$ is

$$\operatorname{Var}(\hat g_2)=\frac{\lambda e^{-\lambda}(1-\lambda e^{-\lambda})}{n}$$

Note that $\hat g_1$ is asymptotically unbiased for $g(\lambda)$ (by delta method) and $\hat g_2$ is exactly unbiased.

Asymptotic relative efficiency of $\hat g_2$ with respect to $\hat g_1$ is the limit of the ratio of the variances of $\hat g_1$ and $\hat g_2$ as $n\to \infty$.

Your answer for the second part is correct.

When $\lambda=1$, by CLT,

$$\sqrt n(\overline X_n-1) \stackrel{d}\longrightarrow Z \,,\quad\text{ where }Z\sim N(0,1)$$

Delta method in this case says that (provided $g''(1)\ne 0$, which holds here)

$$n\left(g(\overline X_n)-g(1)\right)\stackrel{d}\longrightarrow \frac{Z^2}{2} g''(1)$$

[ The proof is similar to the proof for the usual delta method, except here you need a second order approximation. Note that $$n\left(g(\overline X_n)-g(1)\right)=\frac{n(\overline X_n-1)^2}{2!}g''(\overline X_n^*)=\frac{(\sqrt n(\overline X_n-1))^2}{2}g''(\overline X_n^*)\,,$$ where $\overline X_n^*$ lies between $\overline X_n$ and $1$. ]

Therefore, $$n\left(g(\overline X_n)-\frac1e\right)\stackrel{d}\longrightarrow -\frac{1}{2e}\chi^2_1 $$

In other words,

$$2ne\left(\frac1e-g(\overline X_n)\right) \stackrel{d}\longrightarrow \chi^2_1$$


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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