We study learning problems in which the underlying class is a bounded subset of L p and target Y belongs to . Previously, minimax sample complexity estimates were known under such boundedness assumptions only when p=∞. present sharp estimate that holds for any p > 4; it based on procedure suited...