Vector Precoding for Gaussian MIMO Broadcast Channels: Impact of Replica Symmetry Breaking (1001.3790v3)

Abstract: The so-called "replica method" of statistical physics is employed for the large system analysis of vector precoding for the Gaussian multiple-input multiple-output (MIMO) broadcast channel. The transmitter is assumed to comprise a linear front-end combined with nonlinear precoding, that minimizes the front-end imposed transmit energy penalty. Focusing on discrete complex input alphabets, the energy penalty is minimized by relaxing the input alphabet to a larger alphabet set prior to precoding. For the common discrete lattice-based relaxation, the problem is found to violate the assumption of replica symmetry and a replica symmetry breaking ansatz is taken. The limiting empirical distribution of the precoder's output, as well as the limiting energy penalty, are derived for one-step replica symmetry breaking. For convex relaxations, replica symmetry is found to hold and corresponding results are obtained for comparison. Particularizing to a "zero-forcing" (ZF) linear front-end, and non-cooperative users, a decoupling result is derived according to which the channel observed by each of the individual receivers can be effectively characterized by the Markov chain u-x-y, where u, x, and y are the channel input, the equivalent precoder output, and the channel output, respectively. For discrete lattice-based alphabet relaxation, the impact of replica symmetry breaking is demonstrated for the energy penalty at the transmitter. An analysis of spectral efficiency is provided to compare discrete lattice-based relaxations against convex relaxations, as well as linear ZF and Tomlinson-Harashima precoding (THP). Focusing on quaternary phase shift-keying (QPSK), significant performance gains of both lattice and convex relaxations are revealed compared to linear ZF precoding, for medium to high signal-to-noise ratios (SNRs). THP is shown to be outperformed as well.