Suppose p$\backslash$theta $($x$)$ and pdata$($x$)$ both place probability mass in only a very small part of$$the domain; that is$,$ consider the limit $->0.$ What happens to KL$($p$\backslash$theta $($x$)||$ pdata$($x$))$ and its derivative with$$respect to $\backslash$theta $,$ assuming that $\backslash$theta $=\backslash$theta $0?$ Why is this problematic for a GAN trained with the loss function LG$$defined in problem $3$c$?$