A Single-Letter Capacity Formula for MIMO Gauss-Markov Rayleigh Fading Channels

Rami Ezzine, Moritz Wiese, Christian Deppe, Holger BochearXiv 2022 · DOI · arXiv

This paper, recently published by researchers at the Technical University of Munich and Ruhr-Universität Bochum, provides a definitive mathematical answer to a long-standing question in wireless communications: What is the fundamental data-carrying limit of a MIMO channel that changes over time according to a Gauss-Markov process?

TL;DR

The authors derive a rigorous, "single-letter" capacity formula for MIMO Gauss-Markov Rayleigh fading channels where the receiver has perfect channel state information (CSIR). Their core finding is a bit of a "shocker" in its simplicity: the temporal memory of the channel does not actually change its capacity compared to a standard memoryless (i.i.d.) channel.

What Problem Does This Paper Address?

In the real world, wireless channels aren't static. As you move with your phone, the physical environment changes, causing the signal to "fade" over time. Engineers often model this using a Gauss-Markov process, where the channel state at any given moment depends partly on its state in the previous moment (controlled by a "memory factor" $\alpha$).

While we have had capacity formulas for simple, independent fading (where every millisecond is a total surprise), the math for channels with memory is significantly more complex. Previous research mostly focused on "finite-state" Markov channels. This paper tackles infinite-state (continuous-valued) Markov channels in a MIMO (Multiple-Input Multiple-Output) setting—the technology behind 5G and 6G. Specifically, they wanted to find a "single-letter" formula: a simple expression that characterizes the capacity without needing to calculate limits over infinite sequences of time.

Methods

The authors model the channel gain $G_i$ at time $i$ as: $$G_i = \sqrt{\alpha}G_{i-1} + \sqrt{1-\alpha}W_i$$ where $\alpha$ is the memory factor and $W_i$ is fresh Gaussian noise.

To prove the capacity formula, they employed two classic information-theoretic strategies:

  1. The Direct Proof (Achievability): They used Feinstein’s Lemma, which helps prove that a certain transmission rate is possible. A key part of this was proving that the "information density" (a measure of how much information is packed into a signal) doesn't fluctuate too wildly over time (i.e., its variance is bounded).
  2. The Converse Proof: They used Fano’s Inequality and Jensen’s Inequality to show that it is mathematically impossible to exceed the derived rate, regardless of how clever your coding scheme is.

Datasets

As a theoretical information theory paper, no empirical datasets were used. The results are derived through rigorous mathematical proofs based on established statistical models (Rayleigh fading and Gauss-Markov processes).

Key Insights and Ideas

  • The Memory Paradox: The most significant insight is that the memory factor $\alpha$ completely drops out of the final capacity equation. Even though the channel has "memory," the capacity is identical to a channel where every time step is independent.
  • The Power of CSIR: Why does memory disappear? Because the receiver has perfect Channel State Information (CSIR). If the receiver knows exactly what the channel is doing at every moment, it can perfectly "track" the variations. In this specific mathematical setup, the temporal correlation becomes redundant for the purpose of calculating the maximum data rate.
  • Single-Letter Elegance: In information theory, "multi-letter" formulas are often messy and hard to use for engineering design. By proving a "single-letter" characterization, the authors have provided a tool that is easy for engineers to plug into optimization problems for 6G system design.

Key Results

The capacity $C$ for a MIMO Gauss-Markov Rayleigh fading channel with average power constraint $P$ and perfect CSIR is: $$C(P, N_R \times N_T) = \max_{Q} \mathbb{E} \left[ \log \det \left( \mathbf{I}_{N_R} + \frac{1}{\sigma^2} \mathbf{G Q G}^H \right) \right]$$ * Quantitative Result: This formula is exactly the same as the capacity for i.i.d. MIMO Rayleigh fading channels. * Qualitative Result: The "memory" of the fading process (the $\alpha$ factor) has zero impact on the Shannon capacity when the receiver knows the channel state.

Limitations and Future Work

  • Perfect CSIR Assumption: The biggest limitation is the assumption of perfect knowledge at the receiver. In real systems, we estimate the channel using pilots, which introduces error. The authors note that without perfect CSIR, the memory factor would likely play a much larger role.
  • No CSIT: The model assumes the transmitter knows nothing about the current channel state. Adding partial or delayed feedback to the sender (CSIT) would change the optimization of the input covariance matrix $Q$.
  • Higher-Order Models: This paper uses a first-order Gauss-Markov model (the current state only depends on the immediate past). The authors suggest that future work should investigate if these results hold for higher-order Markov models, which might more accurately reflect complex real-world mobility.

Summary updated

BibTeX
@misc{ezzine2022single,
  author = {Rami Ezzine and Moritz Wiese and Christian Deppe and Holger Boche},
  title = {{A Single-Letter Capacity Formula for MIMO Gauss-Markov Rayleigh Fading Channels}},
  year = {2022},
  eprint = {2207.13357},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2207.13357},
}
Figure