We study the following 3-user secure computation problem: Alice and Bob have data in the form of random variables X and Y, respectively. Charlie wants to compute a function f(x,y) of the data. The users are connected by pairwise, secure, noiseless, bidirectional links; they also have access to private randomness. We require that at the end of computation neither Alice nor Bob learns anything about the other�s data. Furthermore, we require that Charlie does not learn anything more about (X,Y) than can be inferred from f(X,Y). The celebrated result of Ben-Or, Goldwasser, and Wigderson (BGW) implies that this is possible. However, the question of how much communication is required to achieve this remains open. We give new information theoretic lowerbounds which are tight for some cases.