🔢 mathematics

From Bayesian Asymptotics to General Large-Scale MIMO Capacity

This paper presents a unifying framework that bridges Bayesian asymptotics and information theory to derive an analytic formula for the asymptotic Shannon capacity of general large-scale MIMO channels—including those with nonlinearities and imperfect hardware—by showing that their behavior is governed solely by the single-output channel's Fisher information, thereby enabling practical constellation design and low-complexity receivers.

Original authors: Sheng Yang, Richard Combes

Published 2026-02-24

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Sheng Yang, Richard Combes

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Massive MIMO" Problem

Imagine you are trying to shout a secret message to a friend across a massive stadium. In the old days, you had one megaphone and one ear. Now, imagine you have thousands of megaphones (transmit antennas) and your friend has thousands of ears (receive antennas). This is called Large-Scale MIMO.

Theoretically, having thousands of ears should let you hear the message perfectly, even if the stadium is noisy. But there's a catch:

The hardware is imperfect: Some ears are "clipped" (they can't hear loud sounds), some are "quantized" (they only hear "loud" or "quiet," not the volume in between), and some are "blurred" (phase noise).
The math is impossible: Calculating the absolute maximum amount of data you can send through this messy, noisy, broken system is usually a mathematical nightmare. It's like trying to solve a puzzle where the pieces keep changing shape.

The Paper's Solution:
The authors (Yang and Combes) found a "magic key" that unlocks the answer for almost any of these messy systems. They realized that when you have so many ears, the problem stops being about "communication" and starts being about statistics.

They used a concept from statistics called Bayesian Asymptotics to say: "If we have enough ears, we don't need to know the exact shape of every single noise wave. We just need to know how 'sensitive' a single ear is to your voice."

The Core Concept: The "Fisher Information" (The Ear's Sensitivity)

To understand their solution, imagine you are trying to tune a radio.

The Input: You are the radio station (the message).
The Output: The listener is the radio (the antenna).
The Problem: The radio is broken. Sometimes it clips the sound, sometimes it only hears static.

The authors introduce a metric called Fisher Information. Think of this as a "Sensitivity Score" for a single antenna.

If a tiny change in your voice causes a huge change in what the antenna hears, the Sensitivity Score is high.
If the antenna is broken or the noise is too loud, a change in your voice barely registers. The score is low.

The Big Discovery:
In a system with thousands of antennas, the total capacity (how much data you can send) depends almost entirely on this single "Sensitivity Score" of one antenna. You don't need to simulate the whole stadium; you just need to measure one ear.

They call the final result the "Jeffreys Factor." Think of this as a "Volume Knob" that tells you exactly how much data the system can handle, based purely on that sensitivity score.

The "Magic Recipe" (How to Fix the System)

The paper provides a step-by-step recipe to solve these impossible problems:

Look at one antenna: Ignore the thousands of others for a moment. Just look at how one antenna reacts to your signal.
Calculate the Sensitivity: Figure out the "Fisher Information" (how sensitive that one antenna is).
Apply the "Jeffreys Factor": Plug that sensitivity number into their formula.
Result: You instantly get the maximum data rate and the perfect way to shape your signal.

Why is this cool?
Before this paper, if you had a system with "1-bit ADCs" (antennas that only hear "Yes/No" or "Up/Down"), calculating the capacity was incredibly hard. With this paper, you just calculate the sensitivity of one "Yes/No" ear, plug it into the formula, and boom—you have the answer.

The "Constellation Design" (How to Speak Better)

Once you know the "Sensitivity Score," the paper tells you exactly how to speak to get the best results.

Usually, we send signals in a standard grid (like a checkerboard). But the authors say: "No, don't use a checkerboard. Use a custom map!"

They propose a "Compander" technique. Imagine you are squeezing a balloon.

If the balloon is soft in the middle, you squeeze it gently.
If it's hard on the edges, you squeeze it harder.

They suggest reshaping your signal so that it fits the "Sensitivity Score" perfectly.

Analogy: If your ears are most sensitive to soft whispers, you should whisper more often and shout less. If they are sensitive to loud noises, shout more.
The Result: By reshaping your signal to match the "Jeffreys Prior" (the perfect distribution), you can send significantly more data than with standard methods, especially when the hardware is broken or low-quality.

The "Low-Complexity Receiver" (The Smart Summarizer)

Finally, the paper addresses a practical problem: Processing power.
If you have 10,000 ears, a computer trying to process every single sound wave individually would overheat.

The authors suggest a Summarizer:
Instead of recording every single sound wave, the receiver just bins the sounds into a few categories (e.g., "Very Quiet," "Quiet," "Loud," "Very Loud").

The Magic: They prove that even if you throw away the fine details and just keep the "bins," you lose almost zero capacity.
The Benefit: The computer doesn't need to be a supercomputer. It can be a simple, cheap processor because it only has to count how many sounds fell into each bin, not analyze the exact frequency of every wave.

Summary: The Takeaway

The Problem: Modern wireless systems have too many antennas and too many hardware flaws (noise, clipping, low precision) for traditional math to handle.
The Insight: When you have massive numbers of antennas, the whole system behaves like a single, super-sensitive statistical estimator.
The Tool: The Fisher Information (sensitivity of one antenna) is the only thing that matters. It acts as a universal "currency" for capacity.
The Benefit:
- Simplicity: You can calculate the speed limit of a broken, complex system with a simple formula.
- Optimization: You can design better signal shapes (constellations) that adapt to the broken hardware.
- Efficiency: You can build cheaper receivers that don't need to process every single bit of raw data.

In short, the authors found a way to turn a chaotic, broken, massive system into a simple, predictable, and highly efficient machine by realizing that in the crowd, the average sensitivity of one person tells you everything you need to know.

1. Problem Statement

The paper addresses the challenge of analyzing and computing the Shannon capacity of large-scale (massive) MIMO systems where the number of receive antennas ( $n_r$ ) grows large. While standard linear MIMO capacity is well-understood, real-world scenarios often involve:

Nonlinearities: Hardware impairments like clipping, saturation, and low-precision Analog-to-Digital Converters (ADCs), including 1-bit quantization.
Imperfect Hardware/CSI: Phase noise, fading with imperfect Channel State Information (CSI), and optical Poisson channels.
Complex Constraints: Simultaneous peak-power and average-power constraints.

Traditional methods for calculating capacity in these nonlinear or high-dimensional settings are often intractable or rely on specific asymptotic Signal-to-Noise Ratio (SNR) regimes. The authors aim to unify these disparate problems under a single analytical framework that yields explicit, closed-form capacity expressions as $n_r \to \infty$ .

2. Methodology

The core methodology bridges Information Theory and Bayesian Asymptotics, specifically extending the seminal work of Clarke and Barron on Bayesian redundancy.

Bayesian Redundancy Link: The authors leverage the result that in the large-sample regime, the minimum expected redundancy (KL divergence) between the true distribution and an estimated one coincides with the mutual information between the parameter and the observations.
Parameterization: The channel input $x$ is treated as a parameter $\theta$ of the output distribution $p(y|\theta)$ . The per-antenna output distributions are assumed to be independent and identically distributed (i.i.d.) given the input.
Fisher Information & Jeffreys Prior:
- The framework relies on the Fisher Information Matrix (FIM), $J(\theta)$ , of the single-antenna (single-letter) output distribution.
- The Jeffreys prior, proportional to $\sqrt{\det J(\theta)}$ , is identified as the asymptotically optimal input distribution (the capacity-achieving distribution) for large $n_r$ .
Tilted Jeffreys Factor: To handle power constraints (peak and average), the authors introduce a "tilted" Jeffreys prior using a Lagrange multiplier $\lambda$ . This leads to the definition of the (tilted) Jeffreys factor, $JF(\lambda^*)$ , which encapsulates the channel distribution and power constraints into a single integral term.

3. Key Contributions

A. Unified Asymptotic Capacity Formula

The paper derives a compact, general formula for the asymptotic capacity $C(P)$ of large-scale MIMO channels:
$C(P) = \frac{d}{2} \log \left( \frac{n_r}{2\pi e} \right) + \log JF(\lambda^*) + o(1)$
Where:

$d$ is the dimension of the parameter space (degrees of freedom).
$n_r$ is the number of receive antennas.
$JF(\lambda^*)$ is the Jeffreys factor, an integral involving $\sqrt{\det J(\theta)}$ and the power constraints.
$\lambda^*$ is the optimal tilting parameter ensuring the average power constraint is met.

This formula reveals that the asymptotic behavior is governed solely by the single-antenna Fisher information, effectively reducing a massive MIMO capacity problem to a single-antenna parameter estimation problem.

B. Systematic "Recipe" for Capacity Calculation

The authors provide a step-by-step procedure to compute capacity for any channel model fitting the framework:

Identify a smooth, one-to-one mapping between the input and the output distribution parameter $\theta$ .
Compute the Fisher Information Matrix $J(\theta)$ for the single-antenna output.
Evaluate the Jeffreys factor integral to find the optimal $\lambda^*$ and the resulting capacity.

C. Practical Constellation Design

Motivated by the continuous Jeffreys prior, the paper proposes a compander-like transformation for practical discrete constellation design:

Instead of standard uniform grids (like PAM), the input symbols are generated by applying the inverse Cumulative Distribution Function (CDF) of the Jeffreys prior to a uniform grid.
This "Jeffreys Constellation" significantly outperforms conventional designs, especially in low-resolution ADC scenarios (e.g., 1-bit or 2-bit).
The paper also introduces a polynomial approximation method to compute these constellations efficiently when closed-form inverses are unavailable.

D. Low-Complexity Receiver Architecture

The authors propose a receiver structure that quantizes channel outputs into a finite number of bins to approximate the log-likelihood.

Complexity: The computational complexity becomes independent of the output dimension $n_r$ , depending only on the number of bins.
Performance: They prove that the capacity loss due to this approximation scales favorably (e.g., $O(L^{-2/m})$ for bounded support), enabling near-capacity performance with drastically reduced complexity.

4. Results and Applications

The framework is applied to a wide range of previously intractable scenarios, demonstrating its versatility:

Clipping and Saturation: Derives capacity for channels with output clipping, showing how the optimal input distribution concentrates near zero under strong clipping.
Low-Resolution ADCs: Analyzes 1-bit and multi-bit quantization. It shows that dithering (artificially shifting signals before quantization) can significantly boost capacity by increasing the Jeffreys factor.
Phase Noise & Energy Detection: Handles complex AWGN channels where only signal magnitude is observable (non-coherent detection), showing behavior similar to real-valued channels.
Imperfect CSI: Extends to fading channels where the receiver has an estimate of the channel but not the true realization.
Optical Poisson Channels: Applies the method to optical intensity channels with Poisson noise and background light.
Correlated Outputs: Discusses extensions to non-i.i.d. settings (spatial correlation), showing that if the Fisher Information Rate exists, the asymptotic capacity formula remains valid by replacing the single-letter FIM with the rate.

5. Significance

Unification: The paper unifies the analysis of diverse, complex MIMO channels (linear, nonlinear, quantized, optical) under a single theoretical umbrella.
Simplicity: It replaces complex multi-dimensional optimization problems with a tractable single-antenna Fisher Information calculation.
Practicality: The derived "Jeffreys Constellation" offers a systematic, near-optimal design for modern hardware-constrained systems (e.g., massive MIMO with 1-bit ADCs), while the low-complexity receiver architecture addresses the computational bottlenecks of large arrays.
Theoretical Insight: It establishes a profound link between statistical estimation (minimax redundancy) and communication theory (capacity), showing that the "least favorable" prior in statistics is the "best" input distribution in communication for large-scale systems.

In conclusion, this work provides a powerful, generalizable toolkit for designing and analyzing next-generation wireless systems where hardware imperfections and massive antenna counts are the norm rather than the exception.