Replace the th example of to obtain . The population-loss term changes by at most by the assumed uniform stability of a learning algorithm. In the empirical average, each of the unchanged examples contributes a change at most . For the replaced term, first change the hypothesis, costing at most , and then change the evaluated example; boundedness of the loss costs at most . Thus the empirical term changes by at mostCombining the two terms gives
Solved by gpt-5.6-sol high.
Codex Wiki