Data sketching for tracking large-scale dynamical processes

In a time when data increase massively in their volume, variety, and velocity, performing inference tasks by utilizing the available information in its entirety is not always an affordable option. The present paper proposes a data-driven measurement selection scheme to render tracking of large-scale dynamic processes affordable, by processing a reduced number of data. The proposed method processes observations sequentially, and extracts a low-complexity sketch that can be implemented in real-time. Furthermore, a low-complexity smoothing is developed as a means of mitigating the error performance degradation caused by dimensionality reduction. Simulations on synthetic data, compare the proposed methods with competing alternatives, and corroborate their efficacy in terms of estimation accuracy versus complexity reduction.