McDiarmid's inequality bounds how much a function of independent random variables can deviate from its mean, but it typically requires bounded differences or finite variance. This paper relaxes that for upper tails. The authors show that if the function is almost surely bounded above and satisfies a conditional moment condition of order p in (1,2], large deviations above the mean are exponentially unlikely, even when variance is infinite.

The proof rests on a new tensorization inequality for Kullback-Leibler divergence, which the authors call reverse KL tensorization. This is a technical contribution that may be of independent interest in information theory and probability.

The authors demonstrate the result on the sample mean and the median-of-means estimator, and derive one-sided uniform risk bounds for unbounded losses. This extends a standard tool in statistical learning theory to settings with heavy tails.