Simple Bounds on Fidelity-Timeliness Tradeoff of Intelligent Communications for Real-Time CV

Real-time computer vision (CV) is expected to play a vital role in virtual/augmented reality, factory automation, digital twin, and metaverse. To fully unlock its potential, latency- or freshness-constrained communications with a fidelity criterion has attracted considerable recent attention. One of its fundamental limits is the fidelity-timeliness tradeoff (FTT) that still remains open. In this paper, we investigate real-time CV oriented video streaming over AWGN and fading channels with fixed/variable-length lossy compression or even burst arrivals, in which reinforcement learning (RL) inspired intelligent cross-layer scheduling is adopted to minimize the distortion while satisfying a hard-delay or age-of-information (AoI) constraint. A low-complexity algorithm based on constrained Markov decision process (CMDP) and binary search is presented to compute the optimal FTT numerically. To shed more light on FTT, simple analytical upper and lower bounds are derived by leveraging the structural Markovian analysis of conceived bounding policies and saddle-point approximations. An asymptotic analysis will demonstrate that the gap between two bounds vanishes as the signal-to-noise-ratio (SNR) increases, thereby allowing the squeeze theorem to give an optimal but yet analytical FTT in the high-SNR regime. Finally, we find a distortion floor, below which the target AoI/latency becomes infinite, no matter how we increase the power supply.