Ferkans — Interactive Telecom Tutor

From One Matrix to a Million

In a MIMO system with $n_r$ receive and $n_t$ transmit antennas, the channel is an $n_r \times n_t$ matrix $\mathbf{H}$ whose entries are random. The capacity depends on the singular values of $\mathbf{H}$ — and when $n_r$ and $n_t$ are large, something remarkable happens: the empirical distribution of eigenvalues of $\frac{1}{n_t}\mathbf{H}^H\mathbf{H}$ converges to a deterministic limit. This means we can predict the capacity of a massive MIMO system without simulating millions of channel realizations. Random matrix theory provides the tools to make this precise.

Definition:
Empirical Spectral Distribution

Let $\mathbf{A} \in \mathbb{C}^{n \times n}$ be Hermitian with eigenvalues $\lambda_1, \ldots, \lambda_n$ (counted with multiplicity). The empirical spectral distribution (ESD) of $\mathbf{A}$ is the discrete probability measure $F^{\mathbf{A}}(\lambda) = \frac{1}{n}\sum_{i=1}^n \mathbf{1}_{\{\lambda_i \leq \lambda\}}.$ Equivalently, $F^{\mathbf{A}}$ places mass $1/n$ at each eigenvalue. Its density (in the distributional sense) is $f^{\mathbf{A}}(\lambda) = \frac{1}{n}\sum_{i=1}^n \delta(\lambda - \lambda_i).$

The ESD is a random measure when $\mathbf{A}$ is a random matrix. The central question of random matrix theory is: does $F^{\mathbf{A}}$ converge to a deterministic limit as the matrix dimension grows?

Definition:
Wishart-Type Matrix

Let $\mathbf{H} \in \mathbb{C}^{n \times m}$ have i.i.d. $\mathcal{CN}(0,1)$ entries. The matrix $\mathbf{W} = \frac{1}{m}\mathbf{H}^H\mathbf{H} \in \mathbb{C}^{m \times m}$ is called a (normalized) Wishart-type matrix. It is Hermitian positive semidefinite, with rank $\min(n, m)$ . Its nonzero eigenvalues are the squared singular values of $\frac{1}{\sqrt{m}}\mathbf{H}$ .

In MIMO communications, $\mathbf{H}$ is the channel matrix. The eigenvalues of $\frac{1}{m}\mathbf{H}^H\mathbf{H}$ determine the signal-to-noise ratios on the parallel sub-channels created by SVD-based precoding.

Example: ESD of a $3 \times 3$ Matrix

Compute the empirical spectral distribution of $\mathbf{A} = \begin{pmatrix} 4 & 1 & 0 \\ 1 & 3 & 1 \\ 0 & 1 & 2 \end{pmatrix}$ .

Solution

Compute eigenvalues

The characteristic polynomial gives eigenvalues (approximately) $\lambda_1 \approx 4.73$ , $\lambda_2 \approx 2.88$ , $\lambda_3 \approx 1.39$ .

Form the ESD

$F^{\mathbf{A}}(\lambda) = \frac{1}{3}\left[\mathbf{1}_{\{\lambda \geq 1.39\}} + \mathbf{1}_{\{\lambda \geq 2.88\}} + \mathbf{1}_{\{\lambda \geq 4.73\}}\right]$ . This is a staircase function with jumps of $1/3$ at each eigenvalue.

Definition:
Limiting Spectral Distribution

Consider a sequence of $n \times n$ Hermitian random matrices $\{\mathbf{A}_n\}_{n \geq 1}$ . If there exists a deterministic distribution function $F$ such that $F^{\mathbf{A}_n} \xrightarrow{a.s.} F \quad \text{as } n \to \infty,$ where convergence is in the weak (distributional) sense almost surely, then $F$ is called the limiting spectral distribution (LSD) of the sequence.

The existence of a deterministic LSD is the foundational miracle of random matrix theory: even though each eigenvalue is random, their collective histogram stabilizes as the dimension grows. This is a law-of-large-numbers phenomenon for eigenvalues.

Theorem: MIMO Capacity via Eigenvalues

Consider a MIMO channel $\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{w}$ with $\mathbf{H} \in \mathbb{C}^{n_r \times n_t}$ , equal power allocation, and $\mathbf{w} \sim \mathcal{CN}(\mathbf{0}, \sigma^2\mathbf{I})$ . The ergodic capacity is $C = \mathbb{E}\left[\log_2\det\left(\mathbf{I}_{n_r} + \frac{\rho}{n_t}\mathbf{H}\mathbf{H}^H\right)\right] = \mathbb{E}\left[\sum_{i=1}^{\min(n_r,n_t)} \log_2(1 + \rho\, \lambda_i)\right],$ where $\rho = P/\sigma^2$ is the SNR, and $\lambda_1, \ldots, \lambda_{\min(n_r,n_t)}$ are the nonzero eigenvalues of $\frac{1}{n_t}\mathbf{H}^H\mathbf{H}$ .

The SVD decomposes the MIMO channel into parallel pipes. Each pipe contributes $\log_2(1 + \rho\lambda_i)$ bits per channel use. The total capacity is the sum over all pipes — and this sum can be written as an integral against the empirical spectral distribution.

Proof

SVD decomposition

Write $\mathbf{H} = \mathbf{U}\boldsymbol{\Sigma}\mathbf{V}^H$ . Then $\mathbf{H}\mathbf{H}^H = \mathbf{U}\boldsymbol{\Sigma}\boldsymbol{\Sigma}^H\mathbf{U}^H$ and $\det(\mathbf{I} + \frac{\rho}{n_t}\mathbf{H}\mathbf{H}^H) = \prod_i (1 + \frac{\rho}{n_t}\sigma_i^2)$ .

Eigenvalue connection

The eigenvalues of $\frac{1}{n_t}\mathbf{H}^H\mathbf{H}$ are $\lambda_i = \sigma_i^2/n_t$ . Hence $\log\det(\cdot) = \sum_i \log(1 + \rho\lambda_i)$ .

Integral representation

Using the ESD $F^{\mathbf{W}}$ of $\mathbf{W} = \frac{1}{n_t}\mathbf{H}^H\mathbf{H}$ : $\frac{1}{n_t}\log\det(\cdot) = \int \log(1 + \rho\lambda)\, dF^{\mathbf{W}}(\lambda).$ As $n_t, n_r \to \infty$ with ratio $\beta$ , the ESD converges to the LSD, and the per-antenna capacity converges to a deterministic integral.

,

Eigenvalue Histogram of $\frac{1}{m}\mathbf{H}^H\mathbf{H}$

Generate a random $n \times m$ matrix with i.i.d. $\mathcal{CN}(0,1)$ entries and plot the histogram of eigenvalues of $\frac{1}{m}\mathbf{H}^H\mathbf{H}$ . As the dimensions grow, the histogram converges to the Marchenko-Pastur density.

Parameters

n

(rows)100

m

(columns)200

Capacity Scales Linearly with Antennas

The integral representation reveals that the per-antenna ergodic capacity $C/\min(n_t, n_r)$ converges to a constant as the array grows. This means total capacity scales as $\min(n_t, n_r) \log(1 + \text{SNR})$ — a linear scaling with the number of antennas. This is the fundamental promise of massive MIMO, and random matrix theory is what makes this prediction precise.

Quick Check

The matrix $\mathbf{A} = \text{diag}(1, 1, 4, 4, 4)$ has empirical spectral distribution $F^{\mathbf{A}}$ . What is $F^{\mathbf{A}}(3)$ ?

$2/5$

$3/5$

$1/5$

$1$

Correction:

2/5

Two eigenvalues (both equal to 1) satisfy $\lambda_i \leq 3$ , so $F^{\mathbf{A}}(3) = 2/5$ .

Why This Matters: Random Matrix Theory Enables Massive MIMO Design

Without random matrix theory, predicting the performance of a 64-antenna base station serving 16 users would require generating thousands of channel realizations and averaging — expensive and uninformative. With RMT, we can compute the ergodic capacity in closed form (or via a simple fixed-point equation) as a function of the antenna ratio $\beta = n_t/K$ , the SNR, and the spatial correlation structure. This makes system-level design tractable: we can optimize antenna counts, pilot overhead, and power allocation using analytical expressions rather than brute-force simulation.

Common Mistake: Finite Dimensions vs. Asymptotic Limits

Mistake:

Assuming that the Marchenko-Pastur law holds exactly for small matrices (e.g., $4 \times 4$ MIMO) and using asymptotic capacity formulas without checking the approximation quality.

Correction:

The Marchenko-Pastur law is an asymptotic result — it describes the limit as $n, m \to \infty$ with $n/m \to \beta$ . For small dimensions, the ESD fluctuates significantly around the limiting density. In practice, the approximation becomes useful for $\min(n,m) \gtrsim 8$ -- $16$ , but always validate with finite-dimensional simulations for the specific system parameters of interest.

Historical Note: Wigner's Vision: Nuclear Physics to Wireless

1950s--2000s

Random matrix theory began not in communications but in nuclear physics. In the 1950s, Eugene Wigner proposed modeling the Hamiltonian of a heavy nucleus as a large symmetric matrix with random entries. He discovered that the eigenvalue distribution of such matrices converges to a semicircular law. Decades later, when MIMO systems pushed wireless engineers to analyze large random channel matrices, Wigner's mathematical framework found an entirely new application. The path from nuclear energy levels to wireless channel capacities is one of the most unexpected connections in applied mathematics.

Empirical Spectral Distribution (ESD)

The probability measure that places mass $1/n$ at each eigenvalue of an $n \times n$ Hermitian matrix. It is the eigenvalue histogram normalized to be a probability distribution.

Limiting Spectral Distribution (LSD)

The deterministic probability distribution to which the ESD of a sequence of growing random matrices converges almost surely. Its existence is the central concern of random matrix theory.

Related: Limiting Spectral Distribution

Key Takeaway

The eigenvalues of large random matrices exhibit a law-of-large-numbers phenomenon: their empirical distribution converges to a deterministic limit. For MIMO communications, this means the per-antenna capacity converges to a computable constant, making performance prediction tractable without Monte Carlo simulation.

Why Random Matrices Matter for Communications