This question is a continuation of this one. By guesswork, I found out that $\vec{\theta}=(5.2,5.3,1.0)=$ $(A,B,C)$ was a good guess that made my model


fit the $(x_i,y_i)$ points pretty well. This should mean that if I compute the log-likelihood function $\ell(\vec{\theta})$ at the particular values of $\vec{\theta}$ that I guessed, I should come close to the maximum value. In my other question I obtained the likelihood function

$$L(A,B,C | x_i,y_i)= \prod_{i=1}^n \phi\left(\frac{y_i - A \sin(x_i/B)}{C}\right).\tag2$$

Since the standard normal density is $\phi(x)=\frac{1}{\sqrt{2\pi}}e^{-x^2/2}$ I have

$$ \begin{align} L(A,B,C | x_i,y_i) &= \frac{1}{(\sqrt{2\pi})^{23}}\prod_{i=1}^{23}\exp{\left(-\frac{1}{2}\left(\frac{y_i-A\sin\left(\frac{x_i}{B}\right)}{C}\right)^2\right)}\tag3\\ &=\frac{1}{(\sqrt{2\pi})^{23}}\exp\left(-\frac{1}{2C^2}\sum_{i=1}^{23}\left(y_i-A\sin\left(\frac{x_i}{B}\right)\right)^2\right)\tag4 \end{align} $$

Taking the logarithm yields the log-likelihood function $\ell$ to be \begin{align} \ell(A,B,C|x_i,y_i)&=\ln\left(\frac{1}{(\sqrt{2\pi})^{23}}\right)+\ln\left(\exp\left(-\frac{1}{2C^2}\sum_{i=1}^{23}\left(y_i-A\sin\left(\frac{x_i}{B}\right)\right)^2\right)\right)\tag5\\ &= \ln\left(\frac{1}{(\sqrt{2\pi})^{23}}\right)-\frac{1}{2C^{2}}\sum_{i=1}^{23}\left(y_i-A\sin\left(\frac{x_i}{B}\right)\right)^2.\tag6 \end{align}

I used R to compute $\ell(\hat{\theta})$ at $\hat{\theta}=(5.2,5.3,1.0)$ and found that the result was $\approx-43.42094$, however, computing the same function at $\hat{\theta}=(12,12,8)$ gave me $\approx -35.60867.$

This should not be since i should get very close to the maximum using the guessed values. This leads me to believe that I've calculated my log-likelihood function incorrectly, or my code is incorrect, or I've just misunderstood the entire assignment.

Can anyone help me finding out which it is?

  • $\begingroup$ @Xi'an - Are you sure? In which step? I don't see why it should be $1/c^23$ $\endgroup$ – Parseval Dec 6 '18 at 10:50
  • $\begingroup$ @Xi'an - I've checked it for 3 hours since I saw your comment. I honestly don't see why $C$ should be raised tot the 23:d power anywhere :S I must be blind to something. $\endgroup$ – Parseval Dec 6 '18 at 15:39

If$$y_i=A\sin\left(\frac{x_i}{B}\right)+C\epsilon_i,\tag1$$then the density of $y_i$ is$$\frac{1}{C}\varphi\left(\left\{y_i-A\sin\left(\frac{x_i}{B}\right)\right\}/C\right)$$

  • $\begingroup$ So this means that the guy in the other thread got the density wrong? $\endgroup$ – Parseval Dec 6 '18 at 16:10

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.