A One-Sided McDiarmid-Type Inequality for Heavy Tails via Reverse KL Tensorization
This work extends McDiarmid's inequality to provide exponential concentration bounds for functions of independent random variables, even when the function has infinite variance. The key achievement is a one-sided inequality that controls large upper deviations using a novel 'reverse KL tensorization' technique, which is crucial for analyzing heavy-tailed data and unbounded loss functions.