LQG Control with Minimal Information: Three-Stage Separation Principle and SDP-based Solution Synthesis

In the interest of evaluating an information-theoretic requirement for feedback control, this paper proposes a framework to synthesize a control policy that minimizes Massey's directed information from the state sequence to the control sequence while attaining required Linear-Quadratic-Gaussian (LQG) control performance. Interpretation and significance of this framework is discussed in the context of networked control theory. As the main result, we show that an optimal control policy can be realized by an attractively simple three-stage decision architecture comprising (1) a linear sensor with additive Gaussian noise, (2) a Kalman filter, and (3) a certainty equivalence controller. This result suggests an integration of two separation principles previously known in the literature: the filter-controller separation principle in the LQG control theory, and the sensor-filter separation principle in zero-delay rate-distortion theory for Gauss-Markov sources. It is also shown that an optimal policy can be synthesized by semidefinite programming (SDP). Both time-varying finite-horizon problems and time-invariant infinite-horizon problems are considered. Our results can be viewed as a generalization of the data-rate theorem for mean-square stability by Nair and Evans, extended for a control performance analysis.