Leveraging Mutual Information in Stochastic CSI Analysis for Wi-Fi Sensing

The latest generations of wireless networks are focusing on the integration of Joint Communication and Sensing capabilities to improve environment awareness and thus enhance services and user experience. One of the elements commonly used for sensing tasks is the Channel State Information (CSI) extracted at the receiver. Most studies exploit Artificial Intelligence (AI) to perform CSI-based sensing tasks, from localization to activity recognition and more. These studies have shown, at least in small, controlled experiments, exceptional capabilities, but have hardly ever offered interpretative models, insight on the sensing capabilities of the methods or fundamental results and bounds. Fortunately, studies are starting to appear that explore an analytic characterization of the CSI. This work falls into this latter category, proposing a quantitative method to measure the relationship between different CSI's. The approach leverages a quantized representation of the CSI amplitude and a reduced-complexity matrix representation that allows to compute a form of Mutual Information (MI) between CSI's. The MI is then used as a compact indicator of the similarity of two sets of CSI, potentially collected in different environments. The proposed method aims at providing a complement or, possibly, an alternative to AI-based sensing, improving the understanding of the CSI features exploited by AI. Results, encouraging though preliminary, are obtained on real-life situations and experiments that overall cover more than six hours of data collection, spanning across several months and amounting to over 800000 CSI's.