Ferkans — Interactive Telecom Tutor

Beyond i.i.d.: Correlated Channels

The Marchenko-Pastur law assumes i.i.d. entries in $\mathbf{H}$ . In practice, MIMO channels exhibit spatial correlation: nearby antennas see similar scattering environments. A correlated channel is modeled as $\mathbf{H} = \boldsymbol{\Sigma}_{r}^{1/2} \mathbf{X} \boldsymbol{\Sigma}_{t}^{1/2}$ , where $\mathbf{X}$ has i.i.d. entries and $\boldsymbol{\Sigma}_{r}$ , $\boldsymbol{\Sigma}_{t}$ are the receive and transmit correlation matrices. The eigenvalue distribution of $\frac{1}{n_t}\mathbf{H}\mathbf{H}^H$ is no longer Marchenko-Pastur — but deterministic equivalents provide computable approximations.

Definition:
Deterministic Equivalent

Let $\{f_n\}$ be a sequence of random variables indexed by the matrix dimension $n$ . A deterministic sequence $\{\bar{f}_n\}$ is called a deterministic equivalent of $\{f_n\}$ if $f_n - \bar{f}_n \xrightarrow{a.s.} 0 \quad \text{as } n \to \infty.$ The value of $\bar{f}_n$ is that it depends only on the parameters of the model (e.g., the correlation matrices), not on the random matrix realization — it can be computed without Monte Carlo.

Deterministic equivalents are the applied output of random matrix theory. Rather than characterizing the full spectral distribution, they provide approximations for specific functionals (capacity, SINR, MSE) that are the quantities of engineering interest.

Theorem: Deterministic Equivalent for $\log\det$ (Correlated MIMO Capacity)

Consider $\mathbf{H} = \boldsymbol{\Sigma}_{r}^{1/2} \mathbf{X} \boldsymbol{\Sigma}_{t}^{1/2}$ where $\mathbf{X} \in \mathbb{C}^{n \times m}$ has i.i.d. $\mathcal{CN}(0, 1/m)$ entries, and $\boldsymbol{\Sigma}_{r} \in \mathbb{C}^{n \times n}$ , $\boldsymbol{\Sigma}_{t} \in \mathbb{C}^{m \times m}$ are Hermitian positive definite. As $n, m \to \infty$ with $n/m \to \beta$ : $\frac{1}{n}\log\det\!\left(\mathbf{I}_n + \rho\, \mathbf{H}\mathbf{H}^H\right) - \bar{V}_n \xrightarrow{a.s.} 0,$ where $\bar{V}_n$ is defined by the system of equations: $\bar{V}_n = \frac{1}{n}\log\det(\mathbf{I}_n + \rho\, \boldsymbol{\Sigma}_{r} \mathbf{T}_n) + \frac{1}{n}\log\det(\mathbf{I}_m + \rho\, \boldsymbol{\Sigma}_{t} \tilde{\mathbf{T}}_n) - \frac{\rho^2}{n}\text{tr}(\mathbf{T}_n \boldsymbol{\Sigma}_{r} \tilde{\mathbf{T}}_n \boldsymbol{\Sigma}_{t}),$ with $\mathbf{T}_n$ and $\tilde{\mathbf{T}}_n$ the unique positive definite solutions of $\mathbf{T}_n = \left(\mathbf{I}_m + \rho\, \boldsymbol{\Sigma}_{t} \tilde{\mathbf{T}}_n\right)^{-1}, \quad \tilde{\mathbf{T}}_n = \frac{1}{m}\left(\mathbf{I}_n + \rho\, \boldsymbol{\Sigma}_{r} \mathbf{T}_n\right)^{-1}.$

The deterministic equivalent replaces the random $\log\det$ (which involves all eigenvalues of the random channel) with a quantity that depends only on the deterministic correlation matrices $\boldsymbol{\Sigma}_{r}$ and $\boldsymbol{\Sigma}_{t}$ . The matrices $\mathbf{T}_n$ and $\tilde{\mathbf{T}}_n$ are found by iterating two coupled matrix equations until convergence.

Proof

Resolvent identity

The proof starts from the identity $\frac{d}{d\rho}\log\det(\mathbf{I} + \rho\mathbf{H}\mathbf{H}^H) = \text{tr}(\mathbf{I} + \rho\mathbf{H}\mathbf{H}^H)^{-1}\mathbf{H}\mathbf{H}^H$ . The right side involves the resolvent of $\mathbf{H}\mathbf{H}^H$ , whose normalized trace has a deterministic equivalent by the Marchenko-Pastur-type analysis for correlated matrices.

Fixed-point equations

The deterministic equivalent of $\frac{1}{n}\text{tr}(\mathbf{I} + \rho\mathbf{H}\mathbf{H}^H)^{-1}\boldsymbol{\Sigma}_{r}$ leads to the coupled equations for $\mathbf{T}_n$ and $\tilde{\mathbf{T}}_n$ . These generalize the scalar MP fixed-point to the matrix-valued case, capturing the structure of $\boldsymbol{\Sigma}_{r}$ and $\boldsymbol{\Sigma}_{t}$ .

Integration

Integrating the deterministic equivalent of the derivative from $0$ to $\rho$ yields the log-det approximation $\bar{V}_n$ . The correction term $-\frac{\rho^2}{n}\text{tr}(\mathbf{T}_n \boldsymbol{\Sigma}_{r} \tilde{\mathbf{T}}_n \boldsymbol{\Sigma}_{t})$ accounts for the coupling between the two fixed-point equations.

,

Example: Deterministic Equivalent Reduces to MP for i.i.d. Channels

Show that when $\boldsymbol{\Sigma}_{r} = \mathbf{I}_n$ and $\boldsymbol{\Sigma}_{t} = \mathbf{I}_m$ (no correlation), the deterministic equivalent simplifies to the capacity formula from the Marchenko-Pastur law.

Solution

Simplify the fixed-point equations

With identity correlation matrices, the coupled equations become scalar: $T = (1 + \rho\tilde{T})^{-1}$ and $\tilde{T} = \frac{1}{\beta}(1 + \rho T)^{-1}$ (where $T$ and $\tilde{T}$ are now scalars times identity).

Solve the scalar fixed point

Substituting: $T = \frac{1}{1 + \rho/({\beta(1 + \rho T)})}$ . Let $\delta = \rho T$ . This gives $\delta(1 + \delta/\beta) = \rho/(1 + \delta)$ , which is equivalent to the MP Stieltjes transform fixed-point from Section 21.3.

Recover the capacity

The deterministic equivalent $\bar{V}_n$ reduces to $\log(1 + \rho T) + \frac{1}{\beta}\log(1 + \rho\tilde{T}\beta) - \rho^2 T \tilde{T}$ , which matches the closed-form MP capacity expression from Telatar.

Example: Deterministic Equivalent with Exponential Correlation

A massive MIMO base station has $n_t = 64$ antennas with exponential correlation $[\boldsymbol{\Sigma}_{t}]_{i,j} = r^{|i-j|}$ where $r = 0.7$ , serving $n_r = 16$ single-antenna users (no receive correlation, $\boldsymbol{\Sigma}_{r} = \mathbf{I}$ ). Compute the per-user ergodic capacity at $\text{SNR} = 10$ dB using the deterministic equivalent vs. Monte Carlo.

Solution

Setup

$\beta = n_r/n_t = 1/4$ , $\rho = 10$ . The transmit correlation is a Toeplitz matrix with $[\boldsymbol{\Sigma}_{t}]_{i,j} = 0.7^{|i-j|}$ .

Fixed-point iteration

Initialize $\tilde{\mathbf{T}} = \frac{1}{n_t}\mathbf{I}_{n_r}$ . Iterate: $\mathbf{T} \leftarrow (\mathbf{I}_{n_t} + \rho\, \boldsymbol{\Sigma}_{t} \tilde{\mathbf{T}})^{-1}$ , $\tilde{\mathbf{T}} \leftarrow \frac{1}{n_t}(\mathbf{I}_{n_r} + \rho\, \mathbf{T})^{-1}$ . Since $\boldsymbol{\Sigma}_{r} = \mathbf{I}$ , the $\tilde{\mathbf{T}}$ equation simplifies. Convergence in ~10 iterations.

Result

The deterministic equivalent gives approximately $3.8$ bits/s/Hz per user. Monte Carlo with 1000 realizations gives $3.82 \pm 0.04$ bits/s/Hz — the deterministic equivalent is accurate to within 1% even at these moderate dimensions.

Deterministic Equivalent vs. Monte Carlo for Correlated MIMO

Compare the deterministic equivalent capacity approximation with Monte Carlo simulation for a MIMO channel with exponential transmit correlation.

Parameters

n_t

(transmit antennas)64

n_r

(receive antennas)16

Correlation

r

0.5

SNR (dB)10

Random Matrix Theory Tools: Summary

Tool	Input	Output	Use Case
Marchenko-Pastur Law	Aspect ratio $\beta$	Limiting eigenvalue density	i.i.d. Rayleigh MIMO capacity
Stieltjes Transform	Distribution $F$	Analytic function $m_F(z)$	Computing capacity, SINR, MSE integrals
Fixed-Point Equation	Model parameters ( $\beta$ , correlations)	Stieltjes transform value	Avoiding eigenvalue computation
Deterministic Equivalent	Correlation matrices $\boldsymbol{\Sigma}_{r}, \boldsymbol{\Sigma}_{t}$	Approximate capacity/SINR	Correlated massive MIMO analysis

🎓CommIT Contribution(2012)

Random Matrix Theory for Massive MIMO with Spatial Correlation

S. Wagner, R. Couillet, M. Debbah, G. Caire — IEEE Journal on Selected Areas in Communications, vol. 30, no. 3

Wagner, Couillet, Debbah, and Caire applied the deterministic equivalent framework to analyze linear precoding (zero-forcing, regularized zero-forcing) in the MISO broadcast channel with spatially correlated channels. Using RMT, they derived closed-form expressions for the per-user SINR in the large-system limit, enabling optimization of the regularization parameter without Monte Carlo simulation. This work demonstrated that deterministic equivalents are not just theoretical curiosities but practical tools for massive MIMO system design.

massive-MIMOdeterministic-equivalentprecodingRMTView Paper →

Historical Note: Telatar's MIMO Capacity Formula

1999

In 1999, Emre Telatar published one of the most influential papers in wireless communications, deriving the ergodic capacity of the i.i.d. Rayleigh MIMO channel using random matrix theory. The key step was recognizing that the per-antenna capacity could be expressed as an integral against the Marchenko-Pastur distribution. This single insight — connecting an abstract result from mathematical physics to a concrete engineering quantity — launched two decades of research on massive MIMO and random matrix methods for wireless communications.

Quick Check

The fixed-point iteration for the deterministic equivalent requires inverting matrices of size $n_t \times n_t$ and $n_r \times n_r$ at each step. For a system with $n_t = 64$ and $n_r = 16$ , approximately how many floating-point operations per iteration?

$O(n_t^3) \approx 2.6 \times 10^5$

$O(n_t n_r) \approx 10^3$

$O(n_t^2 n_r^2) \approx 10^6$

$O(2^{n_t})$

Correction:

O(n_t^3) \approx 2.6 \times 10^5

The dominant cost is the $n_t \times n_t$ matrix inversion, which costs $O(n_t^3)$ . With 10 iterations, this is still much cheaper than Monte Carlo.

Common Mistake: Deterministic Equivalent Iteration May Diverge with Wrong Initialization

Mistake:

Initializing the fixed-point iteration with arbitrary matrices (e.g., zero matrices) and expecting convergence.

Correction:

The standard initialization is $\tilde{\mathbf{T}}^{(0)} = \frac{1}{m}\mathbf{I}$ (or $\frac{1}{1+\rho}\mathbf{I}$ ). The iteration is a contraction map for this initialization and converges monotonically. Starting from $\tilde{\mathbf{T}}^{(0)} = \mathbf{0}$ can lead to numerical issues. Always use the recommended initialization.

Why This Matters: Deterministic Equivalents in 5G NR System Design

5G NR massive MIMO base stations (gNBs) with 32--256 antenna elements exhibit significant spatial correlation due to the half-wavelength antenna spacing and finite angular spread. Deterministic equivalents allow link-level performance prediction for different array geometries, correlation models, and user distributions without requiring channel measurements or extensive Monte Carlo campaigns. This capability is used in 3GPP system-level simulations to set spectral efficiency targets.

Deterministic Equivalent

A deterministic approximation $\bar{f}_n$ to a random quantity $f_n$ (typically involving large random matrices) such that $f_n - \bar{f}_n \to 0$ almost surely. It replaces Monte Carlo averaging with a fixed-point computation.

Related: Deterministic Equivalent

Key Takeaway

Deterministic equivalents extend random matrix theory from the i.i.d. case to correlated channels. By solving a system of coupled matrix equations (which converge in 10--15 iterations), we can compute the ergodic capacity of a correlated massive MIMO system to within 1% of Monte Carlo — at a tiny fraction of the computational cost. This makes analytical system-level optimization tractable.

Deterministic Equivalents