One-bit mmWave MIMO Channel Estimation using Deep Generative Networks

As future wireless systems trend towards higher carrier frequencies and large antenna arrays, receivers with one-bit analog-to-digital converters (ADCs) are being explored owing to their reduced power consumption. However, the combination of large antenna arrays and one-bit ADCs makes channel estimation challenging. In this paper, we formulate channel estimation from a limited number of one-bit quantized pilot measurements as an inverse problem and reconstruct the channel by optimizing the input vector of a pre-trained deep generative model with the objective of maximizing a novel correlation-based loss function. We observe that deep generative priors adapted to the underlying channel model significantly outperform Bernoulli-Gaussian Approximate Message Passing (BG-GAMP), while a single generative model that uses a conditional input to distinguish between Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) channel realizations outperforms BG-GAMP on LOS channels and achieves comparable performance on NLOS channels in terms of the normalized channel reconstruction error.