#approximation-estimation

With $n$ parameters and $\ell$ training examples, a learning method's generalization error is bounded by $E(n, \ell) \le \frac{1}{n} + \sqrt{\frac{n}{\ell}}$: an approximation term that falls with $n$ and an estimation term that grows with it. If $n$ is chosen optimally for each $\ell$, the bound decreases like $\ell^{-r}$. What is $r$?