Since the clusters can be degenerated or highly flat, these cause very
large positive or negative exponents in densities.
The log likelihood will tend to be inaccurate (not finite).
Since the mixture structures can be over fit, this also cause very
tiny mixing proportions.
The poster probabilities can also unstable (NaN).
These can be solved by rescaling the range of exponents carefully
and adjust the scaling factor on the log values.
See CONTROL
for details about constrains on E- and M-steps.