Mean Age of Information Not Enough: New Study Reveals Distribution Matters for Control

A new paper from researchers at The Ohio State University challenges a core assumption in networked control: that minimizing the average Age of Information (AoI) directly improves closed-loop performance. Authors Abdullah Y. Etcibasi, C. Emre Koksal, and Eylem Ekici mathematically prove that for scalar linear time-invariant (LTI) systems with delayed intermittent updates, the infinite-horizon LQR tracking cost depends on the entire distribution of inter-scheduling intervals—not just the mean. Under state-independent scheduling policies, the optimization reduces to a function of higher-order statistical moments, and in unstable or correlated regimes, exponential moments become critical. This means two scheduling schemes with identical mean AoI can yield dramatically different tracking costs, exposing the inadequacy of mean AoI as a surrogate for control performance.

The work extends to disturbances with exponentially decaying autocorrelation, deriving equivalent cost formulations that isolate the role of the full interval distribution. Empirical validation using real vehicle trajectories from the NGSIM US-101 dataset confirms the theoretical predictions: policies with identical mean AoI but different interval distributions produce distinct LQR costs. The findings have direct implications for real-time control over wireless networks—from autonomous vehicle platooning to drone swarms and industrial robotics. Designers must now consider distribution-aware metrics (e.g., variance or tail behavior) rather than relying solely on average AoI, marking a shift toward control-theoretic scheduling in cyber-physical systems.