Learned Diffusion Channel Priors for Coded MIMO Turbo Receivers

Abstract

Iterative receivers for coded multiple-input multiple-output (MIMO) links estimate the channel from pilots, decoded data, and a prior. In short packets with few pilots and at low signal-to-noise ratio (SNR), the prior matters. An expectation propagation (EP) turbo receiver needs the channel's mean and covariance given an anisotropic Gaussian message from the pilots and decoder. A diffusion prior's denoiser, however, returns only a mean, and only for a channel in isotropic noise. We replace only the receiver's channel-prior module with a learned diffusion prior. At an isotropic query formed from the message, one network gives both moments: its output is the mean, its Jacobian the covariance. They form a matrix-valued EP site, the Gaussian factor representing the prior. We show why an indefinite Jacobian can make the receiver diverge, and we floor its eigenvalues. In coded block error rate (BLER), the diffusion prior fails less than a Gaussian mixture model (GMM) prior learned from the same channels. It does so on a controlled sparse specular model and a 3GPP model, at every tested receive-array and training-set size. On the former, scalar-site controls, with one variance instead of the covariance, keep this advantage, so the gain comes from the diffusion prior.