Bounds which apply when testing on the Test set (aka hold out data) of some learning problem.

Under the usual assumptions of independence, they can be derived using Hoeffding's inequality, or tighter versions of it based on analyzing the tails of the Binomial distribution

See here