2
$\begingroup$

Let $X_1,..,X_n$ i.i.d from an exponential distribution : \begin{align*} f(x) = \lambda e^{-\lambda x} , x > 0 \end{align*} I've computed the MLE and the Fisher information number : \begin{align*} \lambda_{MLE} &= \frac{1}{\bar{X}} \\ I(\lambda) &= \frac{1}{\lambda^2} \end{align*} I therefore obtain the following asymptotic distribution for $\lambda_{MLE} $ : \begin{align} \sqrt{n} (\lambda_{MLE} - \lambda_0) \xrightarrow{D} \mathcal{N}(0,\lambda^2) \end{align} I wish to check that my result is correct by using the CLT theorem and the Delta Method. CLT : \begin{align} \sqrt{n} (\bar{X}-\mu) \xrightarrow{D} \mathcal{N}(0,1/\lambda^2) \end{align} Because $\lambda^2$ is the variance of an exponential random variable.

The delta method states that : \begin{align} \sqrt{n}(g(\hat{\theta}) -g(\theta)) \xrightarrow{D} \mathcal{N}(0,(g'(\theta))^2 \sigma^2(\theta) ) \end{align} With in this case : \begin{align} g(x) &= \frac{1}{x} \\ \frac{dg(x)}{dx} &= \frac{-1}{x^2} \end{align} Therefore : \begin{align} \sqrt{n}(\lambda_{MLE} - \lambda_0) \xrightarrow{D} \mathcal{N}(0,\Big[\frac{-1}{\lambda^2}\Big]^2 \frac{1}{\lambda^2}) \end{align} Which would yields an asymptotic variance of $1/\lambda^6$ which cannot be true. What am I doing wrong ?

$\endgroup$

1 Answer 1

1
$\begingroup$

When you use delta method, you should use $\theta = \frac{1}{\lambda}$ but not $\theta = \lambda$. In this case

\begin{align} \sqrt{n}(\lambda_{MLE} - \lambda_0) \xrightarrow{D} \mathcal{N}(0,\Big[\frac{-1}{\theta^2}\Big]^2 \frac{1}{\lambda^2}) \end{align}

$\endgroup$
1
  • 1
    $\begingroup$ It makes sense, thank you $\endgroup$ Nov 9, 2021 at 10:34

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .