Annealed Langevin Dynamics for Massive MIMO Detection

Solving the optimal symbol detection problem in multiple-input multiple-output (MIMO) systems is known to be NP-hard. Hence, objective of any detector practical relevance get reasonably close solution while keeping computational complexity check. In this work, we propose a MIMO based on an annealed version Langevin (stochastic) dynamics. More precisely, define stochastic dynamical process whose stationary distribution coincides with posterior symbols given our observations. essence, allows us approximate maximum posteriori estimator transmitted by sampling from proposed dynamic. Furthermore, carefully craft dynamic gradually adding sequence noise decreasing variance trajectories, which ensures that estimated belong pre-specified discrete constellation. Based detector, also design robust method unfolding and parameterizing one term– score likelihood– neural network. Through numerical experiments both synthetic real-world data, show yields state-of-the-art error rate performance becomes noise-variance agnostic.