I have the following distribution, defined for $0 < x < \theta$, its value is $0$ otherwise.

$$f_\theta(x)= \frac{2x}{\theta^2} $$

Find the MLE of $\theta$

I tried:

$$\prod_{i=1}^n \frac{2}{\theta^2}x_i =\left(\frac{2}{\theta^2}\right)^n \prod_{i=1}^n x_i $$

Taking the natural logarithm gives us:

$$ n \ln{\frac{2}{\theta^2}} + \sum\ln({x_i})=2n \ln\left({\frac{\sqrt2}{\theta}}\right) + \sum\ln({x_i})$$

Taking the derivative with respect to $\theta$:

$$ \frac{2n\theta}{\sqrt{2}}=0$$

After this question I get other questions about this MLE (unbiasedness, consistency, sufficiency etc), so I have the feeling that this estimator has to be a 'concrete' value...

What's going on? :-)

  • $\begingroup$ Something is fishy about your MLE for $\theta$. See this recent question for the correct answer. $\endgroup$ – Dilip Sarwate Nov 1 '12 at 1:33
  • $\begingroup$ You forgot the indicator in the likelihood. $\endgroup$ – Xi'an Nov 1 '12 at 9:24
  • $\begingroup$ I see that the MLE is Max{X1,X2...Xn} tnx $\endgroup$ – Applied mathematician Nov 1 '12 at 12:03
  • 4
    $\begingroup$ Sorry, but a good example on why one should always start with a plot and not confuse likelihood or its maximum as involving derivatives $\endgroup$ – phaneron Nov 1 '12 at 15:11

So that the question leaves the unanswered queue: As a comment pointed out, we should define the density in one expression and not with branches, since the support depends on the parameter. So

$$f_\theta(x)= \frac{2x}{\theta^2}\cdot \mathbf 1_{\{x\leq\theta \}}$$

for a sample of size $n$, this will give the likelihood

$$L(x) = \prod_{i=1}^n\frac{2x_i}{\theta^2}\cdot \mathbf 1_{\{x_i\leq\theta \}} = \left(\frac{2}{\theta^2}\right)^n\cdot \min_i{\mathbf 1_{\{x_i\leq\theta \}}}\cdot\prod_{i=1}^n2x_i $$

No need to consider the log-likelihood in order to deduce that $L(x)$ is decreasing in $\theta$ related to the term $\left(\frac{2}{\theta^2}\right)^n$, and so we would want $\theta$ as small as possible. But if even one $x_i$ is bigger than the chosen $\theta$ then the likelihood value will be exactly zero, while in all other cases it will be greater than zero, because the indicator function will be zero. But since by increasing $\theta$ we are lowering the value of the likelihood, we increase $\theta$ no more than necessary to avoid making the likelihood equal to zero. Hence $$\hat \theta_{MLE} = \max_i\{x_i\}$$

| cite | improve this answer | |

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.