Worst-Case Expected-Capacity Loss of Slow-Fading Channels

Jae Won Yoo, Tie Liu, Shlomo Shamai, Chao Tian

Abstract

This paper investigates the fundamental performance limits of wireless communication under delay constraints. Specifically, it quantifies the "delay penalty"—the gap between ergodic capacity, which assumes infinite delay, and expected capacity, which assumes coding over a single coherent block. For a channel with $K$ discrete fading states, the authors establish that the worst-case additive loss is $\log K$ and the multiplicative loss is $K$. The study further extends these results to the "writing on fading paper" scenario, providing a definitive characterization of the ergodic capacity and delay gap to within one bit per channel use.

Introduction

In wireless communications, fading represents the temporal fluctuation of signal strength. While ergodic capacity provides a theoretical upper bound by averaging these fluctuations over infinite time, real-world applications such as voice and live video operate under strict delay constraints. In the slow-fading regime, where coding occurs over a single block, a significant penalty is incurred because the system cannot wait for optimal channel conditions. This research addresses the fundamental question: what is the absolute worst-case penalty for this delay across all possible statistical distributions of channel fading?

Methodology

The authors utilize Marginal Utility Functions (MUFs) to provide an explicit characterization of optimal power allocation. MUFs help determine which fading states should be "active" (assigned power) to maximize the expected rate.

Figure 1: Marginal utility functions $u_l(z)$ and $u_k(z)$ for two fading states, illustrating the intersection point $z_{k,l}$ and asymptotes at $-n_l$ and $-n_k$.

As shown in Figure 1, the MUFs $u_l(z)$ and $u_k(z)$ represent the incremental gain in rate for a given power level. The intersection points and asymptotic behavior are critical for identifying the optimal power thresholds. To handle the inherent non-convexity of the expected capacity optimization, the authors employ a change of variables to transform the problem into a convex program, making it mathematically tractable.

Figure 2: Interaction of multiple marginal utility functions $u_1(z), u_2(z), u_3(z), u_4(z)$, demonstrating the complexity of determining active states in a multi-state fading environment.

Figure 2 illustrates how this framework scales with more states ($K=4$). The relative positions of these curves determine the "staircase" power allocation strategy that characterizes the worst-case distributions. The authors also use asymptotic analysis in high-SNR and low-SNR regimes to derive universal bounds.

Key Insights and Results

The research identifies that the "badness" of the delay constraint is directly tied to the number of possible fading states $K$.

  • Point-to-Point Bounds: The supremum of the additive gap is exactly $\log K$, and the multiplicative gap is exactly $K$. This means as the variety of possible channel conditions increases, the penalty grows logarithmically or linearly.
  • Fading Paper Capacity: For the "writing on fading paper" problem—where the transmitter knows the interference but not the fading—the ergodic capacity is shown to be within one bit per channel use of the interference-free point-to-point ergodic capacity.
  • Worst-Case Distributions: The authors identify specific "staircase" power-gain distributions that achieve these bounds, providing a theoretical benchmark for the most challenging channel environments.

Limitations and Future Work

The current results assume no transmitter channel state information (CSI-T). Modern systems with finite-rate feedback might exhibit significantly smaller gaps. Furthermore, characterizing the gap for multi-block delays ($L > 1$) remains an open problem due to the complexities of parallel Gaussian broadcast channels. Finally, extending these bounds to multiple-antenna (MIMO) systems is a promising direction for future research, as the expected capacity for MIMO one-block delay is currently unknown.

Summary updated

BibTeX
@article{yoo2013worst,
  author = {Jae Won Yoo and Tie Liu and Shlomo Shamai and Chao Tian},
  title = {{Worst-Case Expected-Capacity Loss of Slow-Fading Channels}},
  journal = {IEEE Trans. Information Theory},
  year = {2013},
  doi = {10.1109/tit.2013.2248877},
}
Figure