What is needed to exploit knowledge of primary transmissions?

Recently, Tarokh and others have raised the possibility that a cognitive radio might know the interference signal being transmitted by a strong primary user in a non-causal way, and use this knowledge to increase its data rates. However, there is a subtle difference between knowing the signal transmitted by the primary and the actual interference at our receiver since there is a wireless channel between these two points. We show that even an unknown phase results in a substantial decrease in the data rates that can be achieved, and thus there is a need to feedback interference channel estimates to the cognitive transmitter. We then consider the case of fading channels. We derive an upper bound on the rate for given outage error probability for faded dirt. We give a scheme that uses appropriate "training" to obtain such estimates and quantify this scheme's required overhead as a function of the relevant coherence time and interference power.