When would you use Hoeffding's inequality for a CI?

hard

Answer

$P( \mid X\; - \;{\mu} \mid \; > \;t)$
$Gives ε for target coverage: ε = (b-a) * \sqrt(log(2/α) / (2n)).$
$Uses: bandits (UCB), online learning regret, certification without normality.$
$Weaker (wider) than CLT-based CIs for well-behaved data but valid for any n and any distribution on [a, b].$

Check yourself — multiple choice

Requires normality
Distribution-free CI on bounded [a,b] variables; ε = (b-a)√(log(2/α)/(2n)); used in bandits / UCB / anytime-valid inference
Same as CLT
Only for regression

Hoeffding CI: distribution-free on bounded RVs; foundation of bandits / anytime.

#confidence-interval#theory

215 interview questions in this topic.