I am given a bag containing marbles of two colors, with an unknown total number of marbles $N$. I randomly sample $n$ marbles ($n=n_1+n_2 < N$, where $n_1$ and $n_2$ are the number or marbles of the two colors) from the bag, without replacement. The likelihood of the true number of the two colors of marbles in the bag, $N_1$ and $N_2$, is $p(N_1,N_2|n_1,n_2)=\frac{\binom{N_1}{n_1} \binom{N_2}{n_2}}{\binom{N_1+N_2}{n_1+n_2}}$. If I make many repeated samplings from the same bag (i.e. return all marbles and draw randomly again, potentially with different $n$), is it possible to accurately estimate $N_1$ and $N_2$?
I have tried maximum likelihood by optimizing a continuous form of the log-likelihood (replacing the factorials with gamma functions), but I have found that the final estimate depends on the initial estimate, instead of always converging to the correct values. Are there any better ways to approach this problem or is there a fundamental limitation?
