Unsupervised User Selection for Efficient MIMO Communications Spectrum Sharing

Key to the success of multi-user multiple input/multiple output (MU-MIMO) spectrum sharing communications systems is the ability to efficiently cluster users. Clustering can help to reduce channel estimation overhead, maximize aggregate capacity, and efficiently share communications resources among groups of users. Frequently, published clustering algorithms presume that the user data on which the algorithms operate is inherently clustered, i.e. that the data is clusterable. However, in practical MU-MIMO scenarios this is not always the case. Blind clustering may result in increased computational complexity with little benefit. Information metrics are proposed as an instrument with which to predict user channel clusterability, and select data sets appropriate for clustering. Here, we model aggregated user channel data under three illustrative correlation scenarios and investigate coarse geometric properties of each. Next, we apply information-theoretic metrics that summarize those geometric properties, and relate the metrics to the clusterability of the aggregated user channel data as well as the associated effect on capacity. Simulation results suggest the utility of these metrics to discerningly select those user scenarios which will be most appropriate for clustering, and do so robustly across a significant SNR range.