Abstract
In network information theory, determining the exact capacity of relay networks is a long-standing challenge. For a general Gaussian relay network with $N$ nodes, the capacity was previously known to be approximable within an additive gap of $1.3N$ bits. However, this gap scales linearly with the number of relays, making it impractical for large-scale networks. This paper presents a significant advancement for the Gaussian Diamond Network, demonstrating that the capacity can be achieved within a much tighter $O(\log N)$ gap. By employing modified Noisy Network Coding (NNC) with coarse quantization and a Partial Decode-and-Forward (PDF) strategy based on superposition coding, the authors provide a more scalable approximation that remains robust as the network size increases.
Introduction
The Gaussian Diamond Network is a fundamental two-stage relay topology where a single source ($S$) communicates with a destination ($D$) through $N$ parallel relays. The source broadcasts its signal to the relays, which then process and forward their observations to the destination.
Historically, the best-known approximation for the capacity of such networks was within $1.3N$ bits of the cutset upper bound. While theoretically significant, this linear growth in the gap means that for a network with 100 relays, the approximation error could exceed 130 bits—potentially larger than the capacity itself. This research addresses this limitation by seeking strategies where the gap grows only logarithmically with $N$.
Relaying Strategies and Methodology
The authors investigate two primary relaying schemes to achieve the $O(\log N)$ gap:
1. Modified Noisy Network Coding (NNC)
Standard NNC typically suggests that relays should quantize their received signals at the noise level. This paper proves that such a heuristic is suboptimal for large diamond networks. Instead, the authors propose coarse quantization, where the quantization noise power is allowed to scale linearly with the number of relays $N$. This approach reduces the rate required to describe relay observations to the destination, effectively trading off signal precision for a significant reduction in the "rate penalty" of the description.
2. Partial Decode-and-Forward (PDF)
In the PDF scheme, the source utilizes superposition coding to transmit independent messages to each relay. Each relay decodes its intended message and forwards it to the destination. To analyze this, the network is decomposed into two distinct phases: the broadcast phase and the multiple-access phase.
The authors employ Edmond’s polymatroid intersection theorem to prove the performance of PDF. By modeling the two phases as polymatroids, they demonstrate that their intersection contains a rate point that stays within $2 \log N$ bits of the global cutset bound.
Key Insights
- Suboptimality of Noise-Level Quantization: The common practice of quantizing at the noise level is inefficient in large networks. Coarser quantization is superior because the savings in description bits outweigh the loss in signal quality.
- Efficiency of Superposition Coding: While superposition coding often introduces interference from undecoded messages, the authors show that careful rate selection ensures the total loss across $N$ relays only accumulates logarithmically.
- Strategic Rate Loss Concentration: In the PDF strategy, it is more effective to concentrate the rate loss due to superposition coding on the weakest relay rather than distributing it across the entire network.
Results and Performance
The theoretical analysis yields the following bounds:
- NNC: Achieves a rate within $\log(N+1) + \log N + 1$ bits of the cutset bound.
- PDF: Achieves a rate within $2 \log N$ bits of the cutset bound.
Numerical simulations under Rayleigh and shadow fading models support these findings. In Rayleigh fading environments, the gap for PDF typically remains near $\log N$. In shadow fading scenarios, where channel gains can vary drastically, PDF significantly outperforms simpler methods like Amplify-and-Forward (AF), which often fails when some relays experience very weak channels.
Conclusion
This work provides a tighter capacity approximation for the $N$-relay Gaussian Diamond Network, reducing the additive gap from $O(N)$ to $O(\log N)$. While these results are specific to the diamond topology, the underlying principles of coarse quantization and strategic superposition coding offer valuable insights for more complex network structures. Future research aims to generalize these $O(\log N)$ approximations to arbitrary multi-layered topologies.
Summary updated