Let's consider a set of data points of some observable in several bins, and a theoretical model. The agreement between the two is using the $\chi^2/ndf$ formula (where we divide $\chi^2$ by the number of freedom). Let's concentrate on the $\chi^2$ part :
The $\chi^2$ formula itself is $\chi^2=\sum_i (d_i-t_i)^2/\sigma_i^2$
$i$ is the bin of a given observable
$d_i$ is data in the bin $i$
$t_i$ is the prediction from theory in the bin $i$
Is there a needed minimum number of entries from data in a given bin $i$ in order not bias completely the $\chi^2$, by making it exploding ?
Is there a connection with gaussian behaviour from Poisson distribution ?