Suppose $x_i$ are drawn i.i.d. from a $p$-variate Gaussian, $\mathcal{N}\left(\mu,\Sigma\right)$. Suppose one observes $x_1,x_2,\ldots,x_n$. One also observes $s_{n+1},s_{n+2},\ldots,s_{n+m},$ where $s_i = \mbox{sign}\left(x_i\right)$ is a $p$-vector consisting of -1 and +1's. (Well, in principle, it could contain some zeroes...)

I would like to estimate $\mu$ from this information. One should be able to do at least as well as the estimate that ignores the $s_i$ (namely $\hat{\mu} = (1/n) \sum_{1\le i \le n} x_i$). But is there a method that does better? How good is the method? Is this a well-known problem?


1 Answer 1


Last m observations you can consider as censored. You could update the censored observation likelihood by probability of region.

For Example let p=3. you got the sign vector (1,-1,-1), the corresponding representation in likelihood will be $P[X_1>0,X_2<0,X<0]$.

Adding extra information is always better. This problem cannot solve analytically. Solving this problem involve numerical optimisation.

  • $\begingroup$ I interpret this answer as: compute the MLE of $\mu$ based on the $x_i$ in the 'usual way', and use the sign information based on a sign-censored Gaussian likelihood. And the whole thing will have to be done numerically. Is that right? $\endgroup$
    – shabbychef
    Mar 27, 2012 at 17:49
  • 1
    $\begingroup$ You are right. Use the censor information in the ML estimation. $\endgroup$
    – vinux
    Mar 27, 2012 at 18:03

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.