Compression Techniques for MIMO Channels in FDD Systems

In this work, we present an innovative application of transformers and vector quantized variational autoencoders (VQ-VAE) to compress multiple-input-multiple-output (MIMO) channels in frequency-division-duplex (FDD) systems. Existing works consider multiple-input-single-output (MISO) channels across all frequencies (subcarriers) of a certain bandwidth, where high compression ratios can be achieved due to the structure of the channels across the frequency domain, or due to their sparsity in the time domain. With this work, we take into account that in reality, the channels cannot be observed for all the subcarriers inside the bandwidth, therefore, it is crucial to compress the channels considering a single subcarrier observation. Simulation results demonstrate that transformers can be used to construct efficient autoencoders with a reduced amount of parameters. Furthermore, we show that embedding the quantization during the training, using the VQ-VAE framework, helps to achieve better performances compared to a post-training quantization based on standard techniques.