Suppose an observation $x$ comes from a Bernoulli distribution: $$p_{\theta}(x)= \theta^{x}(1-\theta)^{1-x}, x=0,1 \ \ 0 \leq \theta \leq 1$$
Let's say that a future one-step observation $y$ also comes from the same distribution. Given the observation $x$, we want to determine the predictive distribution of $y$. An estimate of the distribution of $y$ would be: $$p_{\delta}(y) = (\delta(x))^{y}(1- \delta(x))^{1-y}$$
Why is it difficult to find the minimax decision $\delta(x)$?