Geometric distribution from exponential estimation

82 Views Asked by At

Problem:

$X \sim \exp(\lambda)$

Now $X$ is descretized to $Y$ using floor function

$Y_1, Y_2, \dots,Y_n$ is available from $Y$ distribution

I need to find method of moment estimator for $\lambda$ using $Y$ sample value

My solution:

$Y \sim$ geometric distribution with parameter $p$

Here $p=1-\exp(-\lambda)$

Now, moment estimator for $p$ is

$$ \hat{p} = \frac{1}{\overline{Y}}. $$

If we keep this $p$ value in above equation to get $\lambda$. $$\lambda= -\ln\left(1-(1/\overline{Y})\right)$$

But my answer is

$$\lambda= \ln\left(1+(1/\overline{Y})\right)$$

Where am I getting it wrong.

1

There are 1 best solutions below

0
On

The distribution of $\lfloor X \rfloor$ when $X$ has pdf $\lambda e^{-\lambda x}$ (which I assume is the convention you are using for the exponential distribution) will have PMF $P(Y=y)=\int_y^{y+1} \lambda e^{-\lambda x} \, dx = -e^{-\lambda(y+1)}+e^{-\lambda y}=e^{-\lambda y} \left ( 1-e^{-\lambda} \right )$. Thus it is geometric with success probability parameter $p=e^{-\lambda}$, under the failure-counting convention (so $0$ is a possible value).

The mean of it is $\frac{1}{p}-1=e^\lambda-1$, so $\overline{Y}$ is a consistent and unbiased estimator for $e^\lambda-1$. Therefore $\log(1+\overline{Y})$ is a consistent but probably biased estimator for $\lambda$.

If you use $\hat{p}$ you will arrive at an algebraically equivalent expression, by using $\frac{1}{\hat{p}}-1=\overline{Y}$ and $\hat{p}=e^{-\hat{\lambda}}$ and solving for $\hat{\lambda}$.