I was trying to implement the EM algorithm (on GMM) in code, however, I was getting that in high dimension, I get numerical issues when computing the E-step:

$$p(j|x^{(t)}) = \frac{ p(j) N(x^{(t)} | j; \mu_j, \sigma^2_j) }{\sum^{K}_{j=1} p(j) N(x^{(t)} | j; \mu_j, \sigma^2_j)}$$

I was going to ask about different numerical or mathematical methods or any method that could make this equation be computed without a NaN or divided by zero issue.

Where is this question a good place to ask? I am sure people have run the EM algorithm before, so this should be a known issue and it seems that it should be well known how to fix it. Or at least I am sure I am not the first one to try to fix such an issue. Is CV a good site to ask this type of question?

This is a boundary code, mathematics, ML question so I was not sure where I could get a useful response.

• I'm not sure where it is best asked. I suspect CV won't be the best place. You could add something like Theoretical Computer Science to your list. – gung - Reinstate Monica Apr 12 '15 at 4:29
• If you happen to use Mathematica, then you would probably get a good answer from Mathematica stack exchange, they have some very good experts there for this type of question. – image_doctor Apr 16 '15 at 17:52
• My guess would be scicomp.stackexchange.com which is the Computational Science SE rather than a Computer Science one! – Silverfish Apr 17 '15 at 23:41
• There is probably no one right answer. Stack Overflow has the largest user base of any of the Stack Exchange pages (see here for details). However, @Silverfish has a good suggestion. Wherever you decide to go, make sure you target your question to that site (e.g., focus on programming at SO or statistics at at CV). The meat of question will be the name, but you'll just want to frame it differently. – Richard Erickson Apr 25 '15 at 0:23