is then a notion of how much its output differs in loss upon being presented with two datasets, Ξ and Ξ', that differ in at most one sample

O. Bousquet and A. Elisseeff. Stability and generalization. JMLR, 2(Mar):499–526, 2002.