A universal prediction lemma and applications to universal data compression and prediction

We consider finite-alphabet sequences which are emitted by a stationary source with unknown statistics. We treat the optimization problem by deriving performance bounds for a restricted class of empirical conditional distributions (predictors).