Use the structure of the network to tell us which sets of joint probabilities to use. Gives us a slightly intelligent way to sum out variables from the joint without actually constructing the explicit representation.