We discuss the on-line learning of probability distributions in a reparamet
rization covariant formulation. Reparametrization covariance plays an essen
tial role not only to respect an intrinsic property of "information" but al
so for pattern recognition problems. We can obtain an optimal on-line learn
ing algorithm with reparametrization invariance, where the conformal gauge
connects a covariant formulation with a noncovariant one in a natural way.