Empirical Evaluation of Distributional Shifts in FDD Systems Based on Ray-Tracing

Recent research on frequency division duplex (FDD) systems discovered that machine learning (ML) methods for downlink (DL) applications can be exclusively trained on uplink (UL) data without suffering from large performance degradations. The conjecture in this context is that, although the channels in the UL and DL in an FDD system are separated by a frequency gap, their channel distributions are similar enough for the training of ML methods. The aim of this manuscript is to validate this conjecture based on ray-tracing (RT) simulations. Our results show that, also when considering channels obtained from RT, the performance of ML methods that are trained on UL channels does not degrade significantly when evaluated on DL channels. With the help of hypothesis testing, we show that, despite the raw channel data for different frequency gaps may be very dissimilar, the differences are in part caused by the discrepancies of the channel norms in the UL and DL. This effect is caused by larger path gains for lower carrier frequencies and can be alleviated by a suitable data preprocessing.