I'm currently dealing with a problem where an underlying assumption is that, in general, subsets of my data follow somehow distinguishable multivariate normal distributions. I'm estimating via MLE approach parameters $(\mu_i, \Sigma_i),\ \ i=1,2,3,...$ of these gaussians, but I started to wonder: is there a rule of thumb specifing how many points are needed for a decent estimation of parameters of multivariate normal distribution as a function of data dimension? Obviously it depends on what is a measure of being decent and how these points are distributed and thus I don't need a precise answer; more of a hint, as in hypothesis testing, where i.e. normal approximation is kinda valid for a sample size of, say, at least $25$ observations.
Asked
Active
Viewed 31 times
0
thesecond
- 360
-
1The short answer is that the data dimension is irrelevant. – whuber Jan 08 '22 at 17:23
-
1@whuber Thank you. – thesecond Jan 08 '22 at 17:37