Sometimes we only need a single conditional statistic of $y$ given $x$ (such as its mean or median) rather than the full conditional distribution $p(y \mid x; \theta)$. In that case we can minimize a cost functional — a mapping from functions to real numbers — instead of an ordinary cost function. This lets us treat learning as choosing a function rather than a set of parameters: the cost functional is designed so that its minimum is attained at the function that outputs the desired statistic. Optimizing over functions in this way relies on the calculus of variations.