Rate-cost tradeoffs in control. Part I: lower bounds

Victoria Kostina, Babak Hassibi

Abstract

This paper investigates the fundamental limits of controlling linear stochastic systems over communication channels with limited capacity. By bridging control theory and information theory, the authors derive explicit lower bounds for the "rate-cost" tradeoff, extending previous results to complex vector systems with non-Gaussian noise. The work establishes that the minimum communication rate required to maintain a target control performance is primarily dictated by the system's unstable modes and the entropy power of the noise.

Introduction

In the era of the Internet of Things (IoT) and remote robotics, control systems increasingly operate over communication networks with limited bandwidth. Unlike classical control theory, which assumes perfect information exchange, these networked systems must balance control performance—typically measured by the Linear Quadratic Regulator (LQR) cost—against the available communication rate. While previous research explored this "rate-distortion" tradeoff, results were often restricted to simple scalar systems or Gaussian noise. This paper fills that gap by providing a rigorous information-theoretic framework for multidimensional, non-Gaussian systems.

System Architecture

The paper considers a standard feedback loop where a sensor observes a system's state and sends information through an encoder and a limited-capacity channel to a controller.

System architecture for control over a communication channel.

The challenge lies in determining the minimum number of bits per second required to achieve a specific target LQR cost. The architecture assumes that the controller must act based on quantized data received through a channel, introducing a fundamental tradeoff between the bit rate and the resulting control error.

Methodology

The authors utilize several sophisticated information-theoretic techniques to establish these limits:

  • Directed Mutual Information: This metric captures the causal dependence of control actions on past sensor data, accounting for the "arrow of time" inherent in feedback loops.
  • Causal Shannon Lower Bound: Developed using the Entropy Power Inequality (EPI), this technique bounds control error recursively without requiring exact knowledge of the state's probability distribution.
  • Lattice Quantization: To prove the tightness of their bounds, the authors propose a practical Differential Pulse Code Modulation (DPCM) architecture where the encoder only transmits the "innovation"—the difference between the actual state and the controller's prediction.

The Rate-Cost Tradeoff

A key result of this work is the derivation of an explicit lower bound for the rate-cost function $R(b)$. This bound reveals that the required communication rate is primarily dictated by the system's unstable modes.

The rate-cost tradeoff showing the lower bound (converse) and the performance of a uniform innovations quantizer.

As shown in the figure, the "converse" represents the theoretical minimum rate. The required rate $r$ scales logarithmically with the target cost $b$ and the determinant of the system matrix $A$ (representing instability). Notably, for non-Gaussian noise, the limit depends on the "entropy power" of the noise rather than just its variance. This implies that systems with "peakier" noise distributions may actually be easier to control at low bitrates than Gaussian systems of equivalent power.

Key Insights

  • Innovation-Based Communication: Efficiency is maximized by communicating only the "surprise" or innovation that the controller cannot predict from the system's physics.
  • Tightness at High Rates: The gap between the theoretical lower bound and practical lattice quantization schemes vanishes as the allowed communication rate increases.
  • Dither-Free Performance: Unlike many theoretical models, this paper shows that deterministic lattice quantizers can achieve near-optimal performance without the need for shared random noise (dither) between the sender and receiver.

Limitations and Future Work

While the results for fully observed systems handle non-Gaussian noise, the analysis for partially observed systems (where the sensor itself is noisy) currently relies on Gaussian assumptions to leverage Kalman filter optimality. Future research may explore whether these bounds are reachable using fixed-length quantization and investigate non-lattice quantizers for the "medium-cost" regime where a small gap remains between the lower bound and current practical schemes.

Summary updated

BibTeX
@article{kostina2016rate,
  author = {Victoria Kostina and Babak Hassibi},
  title = {{Rate-cost tradeoffs in control. Part I: lower bounds}},
  journal = {IEEE Trans. Automatic Control},
  year = {2016},
}
Figure