#hoeffding-inequality

d2l bounds the test error of a fixed classifier $f$ on $n$ fresh examples with Hoeffding's inequality, $P(\epsilon_\mathcal{D}(f) - \epsilon(f) \geq t) < \exp(-2nt^2)$, and gets its "roughly 15,000 examples" for $t = 0.01$ at 95% confidence by setting the right-hand side equal to $0.05$. Using the same recipe, how many test examples are needed to estimate the error to within $t = 0.0001$ with probability 99.9%?