Semidefinite Programming Approach to Gaussian Sequential Rate-Distortion Trade-offs

Authors:
  • TTTakashi Tanaka
    Takashi Tanaka
  • KKKwang-Ki K. Kim
    Kwang-Ki K. Kim
  • PPPablo A. Parrilo
    Pablo A. Parrilo
  • SMSanjoy K. Mitter
    Sanjoy K. Mitter

Abstract
Sequential rate-distortion (SRD) theory is concerned with the fundamental trade-off between data-rate and data-quality in zero-delay transmission of stochastic source signals. In this paper, we consider an SRD trade-off problem for a multi-dimensional and time-varying Gauss-Markov source with a mean-square distortion constraint. We first revisit the "sensor-estimator separation principle," which asserts that a Gaussian SRD problem is equivalent to the problem of designing the best linear sensor mechanism of the form {y_t=C_t x_t+v_t}, and computing the least mean-square error estimate of {x_t} given the history of {y_t} based on the Kalman filter. We then show that the optimal matrix-valued signal-to-noise ratio the sensor mechanism can be efficiently computed by semidefinite programming (SDP). This allows us to obtain an SDP-based solution synthesis procedure for the Gaussian SRD problem and a semidefinite representation of the corresponding SRD function, whose analytical expression in the general case is not known in the literature. We also consider applications of SRD theory in real-time communication theory and networked control theory.