In point-to-point communications, Claude Shannon established that Gaussian inputs are optimal for Additive White Gaussian Noise (AWGN) channels. The intuition is rooted in geometry and calculus: under a fixed power (variance) constraint, the Gaussian distribution maximizes differential entropy. Because mutual information is defined as $I(X;Y) = H(Y) - H(Y|X)$, and the channel noise $H(Y|X)$ is fixed, maximizing the total entropy of the received signal $H(Y)$ requires making the input $X$ Gaussian.
But what happens when we move from a simple point-to-point pipe to a complex network? When multiple transmitters interfere, or when an eavesdropper listens in, does this simple bell curve still reign supreme? One might expect that avoiding structured interference or hiding information from a spy would require complex, non-linear, or discrete signaling strategies. Remarkably, a broad consensus across information theory shows that Gaussian signaling remains optimal across a vast array of multi-user, multi-antenna (MIMO), and secure communication scenarios.
Taming Interference in Multi Antenna Networks
In multi-user networks, the primary challenge is managing interference without central coordination. For instance, in a Multiple-Input Multiple-Output (MIMO) interference channel, signals from different transmitters mix spatially, creating a tangled web of dependencies.
To find the capacity limits of these networks, researchers analyze the conditional entropy of Gaussian random vectors. Xiaohu Shang et al. demonstrated that the optimality of Gaussian signaling in MIMO interference channels can be rigorously verified through specific entropy lemmas. By optimizing Gaussian input distributions under strict covariance constraints, Xiaohu Shang et al. established the sum-rate capacities for vector Gaussian interference channels, proving that structured non-Gaussian signaling cannot yield higher throughput.
This optimality holds even when transmitters cooperate or possess asymmetric information. For example, Wei Wu et al. analyzed weak interference channels where one transmitter has access to both users' messages (degraded message sets). Even with this cooperative advantage, the authors proved that Gaussian inputs remain the most efficient strategy for maximizing capacity.
When interference is known only causally (as it happens), Reza Khosravi-Farsani et al. demonstrated that linear encoding strategies achieve optimal performance when variables are restricted to jointly Gaussian distributions. Similarly, in vector broadcast channels where the transmitter is blind to the channel state but the receivers are fully informed, Hyeji Kim et al. showed that Gaussian input distributions are still optimal—a surprising result given that Gaussian inputs typically fail in general fading channels without transmitter state information.
The Geometry of Secrecy
When a channel requires physical-layer security—transmitting a message to a legitimate receiver while keeping it perfectly hidden from an eavesdropper—the optimization problem becomes non-convex and mathematically intractable at first glance. The goal is to maximize the rate difference between the legitimate channel and the wiretap channel.
Why does Gaussian signaling work here? The intuition lies in the "stable" nature of Gaussian noise. Alex Dytso et al. showed that for stable noise distributions (like Gaussian or Cauchy), the input distribution that maximizes point-to-point capacity inherently aligns with the one that maximizes secrecy capacity. Because the sum of independent Gaussian variables remains Gaussian, optimizing the signal for the legitimate receiver simultaneously presents the most challenging "worst-case" noise-like distribution to the eavesdropper.
This principle simplifies complex multi-antenna security architectures:
- MIMO Wiretap Channels: Ronit Bustin et al. utilized a Minimum Mean Square Error (MMSE) approach to prove that Gaussian inputs achieve the secrecy capacity of MIMO wiretap channels.
- Matrix Generalizations of EPI: To extend this to broadcast settings with multiple confidential layers, Ruoheng Liu et al. introduced a vector generalization of Costa’s Entropy-Power Inequality (EPI), replacing scalar parameters with positive semidefinite matrices to prove that Gaussian codebooks maximize secure transmission rates.
- Secret Superposition Coding: Using these tools, Ghadamali Bagherikaram et al. proved that a combination of Gaussian code superposition and layer randomization is optimal for degraded vector Gaussian broadcast channels.
Whether managing a mix of common, private, and confidential messages, as explored by Haim H. Permuter et al. and Hung D. Ly et al., or broadcasting over independent parallel channels as shown by Ashish Khisti et al., the optimal strategy consistently reduces to optimizing the covariance matrices of simple Gaussian codebooks. Furthermore, Joffrey Villard et al. showed that in Gaussian environments with eavesdropper side information, using the source signal as known channel state at the transmitter achieves the theoretical limits of the rate-distortion-equivocation trade-off.
Gaussian Test Channels in Distributed Sensing
The optimality of Gaussian distributions extends beyond transmission (source coding/signaling) into distributed estimation and lossy compression, often formalized via Gaussian test channels. In these scenarios, multiple distributed agents observe corrupted versions of a source and must compress their observations for a central decoder (the CEO problem).
A Gaussian test channel is a mechanism where an agent simply adds independent Gaussian noise to their observation before linear encoding. Why is this simple linear strategy optimal compared to complex, non-linear quantization?
Yigit Ugur et al. proved this for the Vector Gaussian CEO Problem under logarithmic loss. By leveraging the de Bruijn identity, which connects differential entropy to Fisher information, the authors demonstrated that non-linear representational strategies yield no additional performance gains; simple Gaussian test channels perfectly hit the rate-distortion limit.
[Source X] ---> (Observation + Noise) ---> [Agent] ---> Add Gaussian Noise ---> [Decoder]
(Optimal Strategy)
This linear, noise-adding approach generalizes across various distributed tasks. For distributed hypothesis testing against conditional independence, Abdellatif Zaidi et al. showed that the optimal communication strategy requires using Gaussian test channels combined with time-sharing. Furthermore, when dealing with multi-stage compression under complex distortion constraints, Yinfei Xu et al. developed a powerful extremal inequality using a "monotone path" argument and a "doubling trick" to confirm that Gaussian distributions remain optimal for vector successive refinement with degraded side information.
Key Foundational Work
On the Capacity of Gaussian Weak Interference Channels with Degraded Message sets (2006) — contributed 1 concept elements to this direction.
Vector Gaussian Multiple Description With Individual and Central Receivers (2007) — contributed 1 concept elements to this direction.