Non-coherent wireless networks

In highly mobile scenarios, the wireless channel changes from packet-to-packet. In point-to-point wireless communication, one sends known training symbols for the receiver to estimate the channel. This has been information-theoretically shown to be (approximately) optimal. Current wireless networks estimate network state including channel strengths and topology through training. In point-to-point channels the cost of training is small for large coherence times and is ignored typically.

But in a network the channels states may be needed at a different location in a network. In relay networks the destination needs to know the channel states at the relays for achieving close to capacity rates. We study the problem of how to operate wireless networks accounting for the network learning costs. In particular, we ask whether training based schemes are (approximately) optimal in networks
as well.

The weak and strong links in the network are quantified using generalized degrees of freedom (gDoF) charecterization, and we study which links to use and how to estimate the channel states for best gDoF performance for the network. We showed that one can do (unboundedly) better in reliable information rates than using training by using a combination of subspace signaling and modification of a relaying strategy we had developed a few years ago (called quantize-map-forward). We also showed approximate optimality of our scheme, through new information-theoretic impossibility results. Though this is a theoretical result, we believe this poses an intriguing question on how to deploy wireless networks when there is rapid changes in channel conditions.