Processes in distributed systems communicate with one another exclusively by sending and receiving messages. A process has access to its state but not to the states of other processes. Many distributed algorithms require that a process determine facts about the overall system computation. In anthropomorphic terms, processes "learn" about states of other processes in the evolution of system computation. This paper is concerned with how processes learn. We give a precise characterization of the minimum information flow necessary for a process to determine specific facts about the system.