Iterative Mercury/Waterfilling for Parallel Multiple Access Channels

This paper describes a multiuser power allocation strategy for fixed constellation over parallel Gaussian channels (e.g. OFDM systems). The criterion under consideration is mutual information, given arbitrary input distributions over users and over subcarriers. The algorithm achieves with very low complexity the multi-user aggregate sum mutual information upper bound. The algorithm is based on an iterative Mercury/waterfilling procedure. Moreover, we extend the framework to a decentralized scenario using a linear approximation of the MMSE function. We show, in particular that each user can, under certain assumptions, independently determine the power allocation without knowing the channel information of other users. Simulation results validate the theoretical claims.