2
$\begingroup$

The task is:

Population of the students has been divided into the following three groups:

  1. Students with the mean of grades below 3.5
  2. Students with the mean of grades between 3.5 and 4.5
  3. Students with the mean of grades above 4.5

Each student in the population is described by a vector of random variables $x= (x^1\ x^2\ x^3)^T$, taking one of three possible states: $(1\ 0\ 0)^T$ if the student belongs to the first group, $(0\ 1\ 0)^T$ if the student belongs to the second group,
and $(0\ 0\ 1)^T$ if the student belongs to the third group. The distribution of $x$ is categorical distribution (also known as generalized Bernoulli distribution or Multinoulli distribution) with parameters $\theta= (\theta_1\ \theta_2\ \theta_3)^T$. From the population of the students N examples were drawn. Calculate the maximum likelihood estimator of $\theta$.

I tried to do it similarly to Bernoulli case, but I'm stuck. The idea was to find $\theta^*$ by finding the maximum of probability distribution function. So my try was

$$ M(x\mid\theta)=\prod_{d=0}^D \theta_d^{x_d}=\theta_1^{x_1} \theta_2^{x_2} \theta_3^{x_3}\\ \theta^* = \operatorname*{argmax}_\theta M(x\mid\theta) = \operatorname*{argmax}_\theta \ln(M(x\mid\theta))\\ \ln(M(x\mid\theta))= \ln(\theta_1^{x_1} \theta_2^{x_2} \theta_3^{x_3}) = x_1\ln\theta_1 + x_2\ln\theta_2 + x_3\ln\theta_3 = x^T (\ln\theta_1\ \ln\theta_2\ \ln\theta_3)^T $$

Next step would be calculating derivative with respect to $\theta$ and finding it's zero, but we don't have $\theta$ in the function.

I'm not sure where is my mistake. Or perhaps there is no mistake and it is possible to convert $(\ln\theta_1\ \ln\theta_2\ \ln\theta_3)^T$ to some form with $\theta$?

$\endgroup$
  • $\begingroup$ "Multinoulli" isn't a word. I think you want "multinomial". But it's a nice not-a-word and perhaps ought to be one so I haven't edited to change it. $\endgroup$ – Ethan Bolker Apr 6 '18 at 21:04
  • $\begingroup$ Well, that's what's in our task :p I believe the most common name is categorical distribution. Edited. $\endgroup$ – Yksisarvinen Apr 6 '18 at 21:26
  • $\begingroup$ The next step would NOT be calculating the derivative with respect to $\theta$ and finding it's zero. This is constrained maximization. You might try Lagrange multipliers, although that's not the only way. It is unfortunate that many have made the technique of finding zeros of the derivative into a knee-jerk reflex rather than something to be used consciously. $\endgroup$ – Michael Hardy Apr 6 '18 at 21:28
0
$\begingroup$

Since $(\theta_1,\theta_2,\theta_3)$ must satisfy the constraint $$\theta_1+\theta_2+\theta_3 = 1,\tag 0$$ one way to do this is by Lagrange multipliers. You have $$ \operatorname{grad} (\theta_1+\theta_2+\theta_3) = (1,1,1) \tag 1 $$ and $$ \operatorname{grad} (x_1\log\theta_1 + x_2\log\theta_2 + x_3\log\theta_3) = \left( \frac{x_1}{\theta_1}, \frac{x_2}{\theta_2}, \frac{x_3}{\theta_3} \right). \tag 2 $$ So you want a value of $(\theta_1,\theta_2,\theta_3)$ for which $(2)$ is a scalar multiple of $(1).$ That happens only if the ratio $\theta_1:\theta_2:\theta_3$ is equal to the ratio $x_1:x_2:x_3.$ But the constraint $(0)$ must also hold. Consequently you get $$ \theta_1 = \frac{x_1}{x_1+x_2+x_3} $$ and similarly for the other two values of the subscript.

$\endgroup$
  • $\begingroup$ I'm trying to understand the answer and I wonder if there is some notational confusion here? I believe the question is on finding the ML estimate of $\theta$ given all the examples $X = (\vec{x}^{(1)}, \ldots, \vec{x}^{(N)})$. Any single example $\vec{x}^{(i)} \in \{(1, 0, 0), (0,1,0), (0,0,1)\}$ so in the answer above, $\theta_1$ always becomes either 0 or 1, which doesn't make sense to me. $\endgroup$ – lemonad Oct 17 '18 at 9:54

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.