2
$\begingroup$

$X_1,X_2,...,X_n$ and $Y_1,Y_2,...,Y_n$ are independent samples from the exponential distributions with parameters $\lambda$ and $\frac{1}{\lambda}$.

What is the MLE for $\lambda$?

I used the log-likelihood function:

$logL(\lambda)=\mathbb{}log\prod\lambda e^{-\lambda x_i}\prod\ \frac{1}{\lambda} e^{-\frac{1}{\lambda} y_j}=$

$=n\cdot log(\lambda)+n\cdot log(\frac{1}{\lambda})-\lambda\sum x_i-\frac{1}{\lambda}\sum y_j \rightarrow max\lambda$

After differentiation I got:

$\frac{n}{\lambda}+n\lambda-\sum x_i-\sum y_j =0$

And this is a quadratic equation for $\lambda$.

Can someone check my calculations? Is it correct that there are two possible ML estimations for $\lambda$?

$\endgroup$
  • $\begingroup$ The derivative of $n\log(1/\lambda)$ w.r.t $\lambda$ is $-n/\lambda$, not $n\lambda$. Even simpler is to recognize that the normalization constants $\lambda^n$ and $(1/\lambda)^n$ cancel in the likelihood. Note, too, that the derivative of $-1/\lambda$ is $1/\lambda^2$. $\endgroup$ – whuber Dec 8 '14 at 23:33
6
$\begingroup$

Let's assume that $x_i, y_i > 0,$

The formula simplifies to

$$ n\cdot \log(\lambda)+n\cdot \log(\frac{1}{\lambda})-\lambda\sum x_i-\frac{1}{\lambda}\sum y_j \\=n\cdot \log(\lambda)+n\cdot \log(1)-n\cdot \log(\lambda) -\lambda\sum x_i-\frac{1}{\lambda}\sum y_j \\=-\lambda\sum x_i-\frac{1}{\lambda}\sum y_j $$

and the derivative w.r.t. $\lambda$ is $$ -\sum x_i+\lambda^{-2}\sum y_j $$

so if we define $\bar{x} = \sum x_i$ and $\bar{y}=\sum y_j $, then

$$ -\bar{x}+\lambda^{-2}\bar{y} = 0 $$

$\lambda = \sqrt{\frac{\bar{y}}{\bar{x}}}$, since $\lambda$ needs to be positive

$\endgroup$

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.