[Paper Review] Cut-Set Bound Is Loose for Gaussian Relay Networks

Xiugang Wu, Ayfer Özgür|arXiv (Cornell University)|Jun 4, 2016

Cooperative Communication and Network Coding26 references3 citations

TL;DR

This paper presents a tighter upper bound on the capacity of the Gaussian primitive relay channel using a novel geometric approach based on typicality and Gaussian measure concentration, demonstrating that the classical cut-set bound is strictly loose. The key result shows a fundamental gap of at least 0.0535 bits per channel use between the cut-set bound and the true capacity, implying that current capacity approximations with linear gaps to the cut-set bound are order-optimal but cannot be improved beyond a constant pre-factor.

ABSTRACT

The cut-set bound developed by Cover and El Gamal in 1979 has since remained the best known upper bound on the capacity of the Gaussian relay channel. We develop a new upper bound on the capacity of the Gaussian primitive relay channel which is tighter than the cut-set bound. Our proof is based on typicality arguments and concentration of Gaussian measure. Combined with a simple tensorization argument proposed by Courtade and Ozgur in 2015, our result also implies that the current capacity approximations for Gaussian relay networks, which have linear gap to the cut-set bound in the number of nodes, are order-optimal and leads to a lower bound on the pre-constant.

Motivation & Objective

• To address the long-standing open problem of whether the cut-set bound is tight for Gaussian relay networks.
• To develop a new upper bound on the capacity of the Gaussian primitive relay channel that is strictly tighter than the cut-set bound.
• To quantify the fundamental gap between the cut-set bound and the true capacity, establishing a lower bound on the pre-constant in capacity approximations.
• To demonstrate that existing linear-gap capacity approximations for Gaussian relay networks are order-optimal, with the gap being fundamental and not improvable by better coding schemes.

Proposed method

• A geometric approach based on typicality and concentration of measure in high-dimensional Gaussian spaces is developed to analyze the probabilistic structure of n-letter random variables in the relay channel.
• The method translates geometric relations between typical sets into novel entropy inequalities that capture the noise-injection effect of non-decoding relays.
• A tensorization argument from Courtade and Özgür (2015) is applied to extend the single-relay bound to multi-relay networks.
• The proof leverages Gaussian measure concentration inequalities to bound the probability that a random vector lies within a neighborhood of a set, enabling precise control over typical sequences.
• Key equations involve solving for the optimal parameter $ a^* $ in a nonlinear equation derived from the trade-off between relay and source rates under power constraints.
• The analysis identifies the worst-case channel parameters (infinite SNR at both links) that maximize the gap between the new bound and the cut-set bound.

Experimental results

Research questions

• RQ1 Is the cut-set bound tight for the Gaussian primitive relay channel, or is it loose due to the relay's inability to decode and thus forward noisy signals?
• RQ2 Can a tighter upper bound be derived that captures the fundamental rate loss caused by noise amplification in non-decoding relays?
• RQ3 What is the minimum possible gap between the true capacity and the cut-set bound in Gaussian relay networks?
• RQ4 Does the linear gap in current capacity approximations to the cut-set bound represent a fundamental limit or can it be improved?
• RQ5 Can the geometric structure of typical sets in high-dimensional Gaussian spaces be used to derive new converse bounds in network information theory?

Key findings

• The proposed upper bound is strictly tighter than the cut-set bound for all non-trivial channel parameters in the Gaussian primitive relay channel.
• The maximum gap between the new bound and the cut-set bound is 0.0535 bits per channel use, achieved in the high-SNR limit when both $ P/N_1 $ and $ P/N_2 $ tend to infinity.
• This 0.0535 gap establishes a lower bound on the pre-constant in linear-gap capacity approximations for Gaussian relay networks, proving such approximations are order-optimal.
• The result implies that no coding scheme can close this gap in the worst case, making the linear gap fundamental and not due to suboptimal achievability schemes.
• The geometric method based on typicality and measure concentration successfully captures the noise-injection penalty of non-decoding relays, a phenomenon previously unquantified by standard converse techniques.
• The bound is further sharpened in certain parameter regimes, suggesting potential for even tighter bounds through refined analysis.