Quasi-Static MIMO Fading Channels at Finite Blocklength

Wei Yang, Giuseppe Durisi, Tobias Koch, Yury PolyanskiyarXiv 2013 · OpenReview

Abstract

In wireless communications, the quasi-static fading model is a cornerstone for designing delay-constrained systems. However, its standard performance metric, outage capacity, relies on an infinite blocklength assumption that contradicts the finite coherence time of real-world channels. This paper resolves this paradox by proving that for quasi-static MIMO fading channels, the channel dispersion is zero. This result implies that the rate penalty for finite-length packets vanishes at a rate of $\mathcal{O}(\frac{\log n}{n})$, significantly faster than the $\mathcal{O}(\frac{1}{\sqrt{n}})$ penalty observed in AWGN channels. Consequently, outage capacity serves as a highly accurate proxy for finite-blocklength performance.

Introduction

A quasi-static fading channel is characterized by channel conditions that remain constant for the duration of a single codeword but vary randomly from one codeword to the next. Traditionally, engineers use outage capacity—the maximum rate supportable with an outage probability below $\epsilon$—to evaluate these channels.

The theoretical tension lies in the fact that outage capacity is defined in the limit as blocklength $n \to \infty$. Yet, the quasi-static assumption is only physically valid when $n$ is small enough to fit within the channel's coherence time. This paper investigates whether outage capacity remains a meaningful metric for the short, finite blocklengths where the quasi-static model is actually applied.

Theoretical Framework and Methodology

The authors employ finite-blocklength information theory to derive tight upper (converse) and lower (achievability) bounds on the maximal achievable rate $R^*(n, \epsilon)$.

Converse and Achievability Bounds

The study utilizes the meta-converse theorem to establish absolute upper limits. For achievability, the authors apply the $\kappa\beta$ bound to a stochastically degraded channel. A key contribution is the development of a novel geometric threshold decoder for scenarios without receiver channel state information (CSI). Unlike standard decoders that rely on log-likelihoods, this decoder evaluates the principal angles between the subspace of the received signal and the subspace of the codewords.

Geometric representation of the threshold decoder used for achievability proofs without receiver CSI.

Asymptotic Analysis

By applying a Cramer-Esseen-type central limit theorem, the authors analyze the behavior of these bounds as $n$ increases. This allows for the extraction of the "channel dispersion," a term that typically dictates the speed of convergence to capacity.

The Zero-Dispersion Phenomenon

The most significant finding is that for quasi-static fading channels, the channel dispersion is exactly zero. In standard Additive White Gaussian Noise (AWGN) channels, the finite-blocklength rate penalty is proportional to $1/\sqrt{n}$. In contrast, the absence of dispersion in quasi-static channels leads to a much faster convergence rate of $\mathcal{O}(\frac{\log n}{n})$.

This phenomenon is universal, holding true across various CSI scenarios: no-CSI, CSIT (transmitter only), CSIR (receiver only), and CSIRT (both). The results provide a rigorous mathematical foundation for the long-standing engineering intuition that outage capacity is a practical design metric.

Comparison of conventional wisdom for outage capacity versus finite-blocklength bounds across different CSI scenarios.

Numerical Results and Practical Implications

The theoretical speedup in convergence has tangible practical consequences. In a simulated 1x2 SIMO Rician fading channel with a target error rate of $10^{-3}$:

  • Achieving 90% of the outage capacity requires only 120 to 320 symbols.
  • In comparison, a standard AWGN channel with equivalent capacity requires approximately 1420 symbols to reach the same 90% threshold.

Furthermore, the authors benchmarked their bounds against LTE-Advanced turbo codes. For blocklengths between 500 and 1000, these codes achieve roughly 85% of the theoretical maximal rate. The remaining 15% gap is attributed to the suboptimality of the coding schemes themselves, rather than inherent finite-blocklength penalties.

Limitations and Future Research

While the results are robust for continuous fading distributions (like Rayleigh or Rician), they encounter limitations in specific regimes:

  1. High Determinism: If the channel becomes highly deterministic (e.g., a Rician channel with an extremely large $K$-factor), it begins to resemble an AWGN channel. In this limit, the $\mathcal{O}(\frac{\log n}{n})$ terms grow significantly, and the zero-dispersion approximation becomes less accurate.
  2. MIMO Converse Complexity: For general MIMO systems without transmitter CSI, the converse bound requires an optimization over input covariance matrices that remains analytically and numerically challenging for complex antenna configurations.

In conclusion, this work validates the use of outage capacity for short-packet communications in slow-fading environments, showing that the "cost" of finite blocklength is much lower than previously feared.

Summary updated

BibTeX
@misc{yang2013quasi,
  author = {Wei Yang and Giuseppe Durisi and Tobias Koch and Yury Polyanskiy},
  title = {{Quasi-Static MIMO Fading Channels at Finite Blocklength}},
  year = {2013},
  url = {https://openreview.net/forum?id=XeBcRH-0EsQ},
}
Figure