Two-Sample Tests for Validating the UL-DL Conjecture in FDD Systems

In this work, we present a two-sample tests analysis based on the maximum mean discrepancy metric to validate the recently proposed uplink-downlink conjecture for frequency division duplex systems. This novel concept shows that a neural network trained with uplink channel data can adequately generalize to downlink channel data. With this paper, we focus on a particular application of this idea, namely an autoencoder neural network, which has been introduced lately to generate channel feedback, without requiring any training effort at the mobile terminals. Simulation results with several datasets demonstrate that application-based low-dimensional representations for two-sample testing give a deeper insight into the similarities and dissimilarities between the uplink and downlink data distributions and are in accordance with the performance of the neural network that is applied to the respective datasets.