I am quite stumped by the following problem. The usual log-likelihood route with differentiation doesn't work. The problem is as follows:

Let $X_1,...,X_n$ be i.i.d. random variables from a distribution with p.d.f.:

$$f(x; \theta_1, \theta_2) = \frac{1}{\theta_2} \text{exp}(-\frac{(x - \theta_1)}{\theta_2}) \: \text{for } x \geq \theta_1$$

with parameters $\theta_1 \in R$ and $\theta_2 > 0$.

Find maximum likelihood estimators of $\theta_1$ and $\theta_2$

Any hints or insights?

What I tried:

enter image description here

  • $\begingroup$ If you set the partial derivatives to 0 can't solve the 2 equations in 2 unknowns. If the solution is not unique verify which one is a global maximum. If the solution is unique toy should still be able to verify that it is a maximum. $\endgroup$ Commented Apr 12, 2018 at 3:46
  • 2
    $\begingroup$ In the middle of your notebook paper you have $\log{L}=-\frac{\sum _{i=1}^n x(i)}{\theta _2}+\frac{\theta _1 n}{\theta _2}-n \log \left(\theta _2\right)$. Note that this function can increases with increasing $\theta_1$ and the largest value $\theta_1$ can possibly take is the minimum $x_i$. That should tell you what the maximum likelihood estimator of $\theta_1$ is. $\endgroup$
    – JimB
    Commented Apr 12, 2018 at 4:29
  • 1
    $\begingroup$ There seems to be an error in your last calculation. Shouldn't the last term be $-n\theta_1/\theta_2^2$ $\endgroup$
    – RDK
    Commented Apr 12, 2018 at 4:39
  • 2
    $\begingroup$ You should review your basic material on finding global and local maxima. You should have been taught that setting derivatives to 0 won't always find the maximum (e.g. try it on "find the maximum of $f(x)=1/x$ in the interval $1\leq x\leq 2$"). If you try to set the derivative to $0$ you get nowhere. You've forgotten some important facts! As a first step, consider $\theta_2=1$ and try drawing a picture of the log-likelihood as a function of $\theta_1$ (keeping in mind the relationship between $\theta_1$ and any observation). $\endgroup$
    – Glen_b
    Commented Apr 12, 2018 at 6:20

1 Answer 1


To close this one:

This is the pdf of a shifted Exponential random variable, $Y \sim \text{Exp}(\theta_2)$ (scale parameter), and $X = Y + \theta_1$.

From the condition $X\geq \theta_1$, the MLE is immediately obtained as

$$\hat \theta_{MLE} = X_{(1)}$$

i.e. the minimum order statistic and so the actual estimate will be the minimum value available in the sample. This can be verified by directly inspecting the log-likelihood as suggested in a comment.

Then apply "concentrated ML" estimation, i.e. plug the minimum $x$-value in place of $\theta_1$ and optimize as usual with respect to $\theta_2$ only.


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.