{"id":"arxiv/2606.06855","url":"https://arxiv.org/abs/2606.06855","tags":["stat.ML","cs.LG","math.ST"],"score":0,"title":"Stability beyond Bounded Differences: Sharp Generalization Bounds under Finite $L_p$ Moments","domain":"stat.ML","source":"arxiv","authors":["Qianqian Lei","Soham Bonnerjee","Yuefeng Han","Wei Biao Wu"],"reviews":0,"subject":"stat.ML","summary":"While algorithmic stability is a central tool for understanding generalization of learning algorithms, existing high-probability guarantees typically rely on uniform boundedness or sub-Gaussian/sub-Weibull tail assumptions, which can be overly restrictive for modern settings with heavy-tailed or unbounded losses. We develop a stability-based framework that requires only a finite $L_p$ moment condition. Our first contribution is sharp concentration inequalities for functions of independent random variables under $L_p$ constraints, extending McDiarmid's bounded-differences techniques beyond…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-06-05","reactions":{}}