Signal Processing Aspects of Massive MIMO

Sammanfattning: Massive MIMO (multiple-input-multiple-output) is a technology that uses an antenna array with a massive number of antennas at the wireless base station. It has shown widespread benefit and has become an inescapable solution for the future of wireless communication. The mainstream literature focuses on cases when high data rates for a handful of devices are of priority. In reality, due to the diversity of applications, no solution is one-size-fits-all. This thesis provides signal-processing solutions for three challenging situations.  The first challenging situation deals with the acquisition of channel estimates when the signal-to-noise-ratio (SNR) is low. The benefits of massive MIMO are unlocked by having good channel estimates. By the virtue of reciprocity in time-division duplex, the estimates are obtained by transmitting pilots on the uplink. However, if the uplink SNR is low, the quality of the channel estimates will suffer and consequently the spectral efficiency will also suffer. This thesis studies two cases where the channel estimates can be improved: one where the device is stationary such that the channel is constant over many coherence blocks and one where the device has access to accurate channel estimates such that it can design its pilots based on the knowledge of the channel. The thesis provides algorithms and methods that exploit the aforementioned structures which improve the spectral efficiency.  Next, the thesis considers massive machine-type communications, where a large number of simple devices, such as sensors, are communicating with the base station. This thesis provides a quantitative study on which type of benefits massive MIMO can provide for this communication scenario — many devices can be spatially multiplexed and their battery life can be increased. Further, activity detection is also studied and it is shown that the channel hardening and favorable propagation properties of massive MIMO can be exploited to design efficient detection algorithms.  The third part of the thesis studies a more specific application of massive MIMO, namely federated learning. In federated learning, the goal is for the devices to collectively train a machine learning model based on their local data by only transmitting model updates to the base station. Sum channel estimation has been advocated for blind over-the-air federated learning since fewer communication resources are required to obtain such estimates. On the contrary, this thesis shows that individually estimating each device's channel can save a huge number of resources owing to the fact that it allows for individual processing such as gradient sparsification which in turn saves a huge number of resources that compensates for the channel estimation overhead.