The boundary value 0 of the population proportions makes the log
likelihood as -Inf. Since degeneracy of subpopulations can affect
the maximizing processes in EM steps. This problem is usually caused
by bad initializations, and may suggest that number of cluster
\(K\) may be too large.
Two methods have been implemented when any \(\eta_k\) less than
the lower bound (\(1/N\) or \(1e-16\)).
The ADJUST
(default) will adjust the \(\eta_k\) to the
lower bound, and the IGNORE
will stop the iterations and return
errors.