Ferkans — Interactive Telecom Tutor

Why Free Probability?

The Marchenko-Pastur law tells us the spectrum of $\mathbf{H}\mathbf{H}^{H}$ for an i.i.d. Gaussian $\mathbf{H}$ . But modern systems rarely have one clean random matrix: iterative receivers multiply, add, and invert several independent random matrices at each iteration. What is the spectrum of $\mathbf{W}_1 + \mathbf{W}_2$ when $\mathbf{W}_1, \mathbf{W}_2$ are large, independent, and Hermitian?

In the scalar case, the distribution of a sum of independent random variables is given by convolution. The Fourier transform (characteristic function) turns convolution into multiplication — so computing the distribution of a sum is a multiply-then-invert operation.

Free probability is the non-commutative generalization. Large independent random matrices are asymptotically free, and there is an analog of the characteristic function — the R-transform — that turns the "free convolution" of spectra into addition. This is the cleanest tool available for computing asymptotic spectra of composite random matrix expressions.

Definition:
R-Transform

Let $\mu$ be a compactly supported probability measure on $\mathbb{R}$ with Stieltjes transform $m_\mu(z)$ . Define the functional inverse $K_\mu(z) = m_\mu^{-1}(z)$ on a neighborhood of $0$ . The R-transform is $R_\mu(z) = K_\mu(z) - \frac{1}{z}.$

$R_\mu$ is analytic on a neighborhood of $0$ , and its series expansion encodes the free cumulants of $\mu$ . The R-transform plays the role that the log-characteristic-function plays in classical probability.

Theorem: R-Transform Additivity under Free Convolution

Let $\mathbf{A}_N, \mathbf{B}_N$ be sequences of $N \times N$ Hermitian random matrices with limiting spectral distributions $\mu_A, \mu_B$ , and assume that $\mathbf{A}_N, \mathbf{B}_N$ are asymptotically free (e.g., one is rotated by an independent Haar-distributed unitary). Then the ESD of $\mathbf{A}_N + \mathbf{B}_N$ converges almost surely to a measure $\mu_A \boxplus \mu_B$ whose R-transform is $R_{\mu_A \boxplus \mu_B}(z) = R_{\mu_A}(z) + R_{\mu_B}(z).$

This is the direct analog of the classical fact that the log-characteristic functions of independent random variables add under convolution. The R-transform linearizes free convolution. To compute the spectrum of a sum: (1) compute each R-transform, (2) add them, (3) invert.

Proof

Asymptotic freeness (Voiculescu)

By Voiculescu's theorem, if $\mathbf{A}_N$ has bounded spectrum and $\mathbf{U}_N$ is Haar-distributed independent of $\mathbf{A}_N$ , then $\mathbf{A}_N$ and $\mathbf{U}_N \mathbf{B}_N \mathbf{U}_N^H$ are asymptotically free: their mixed moments factor according to the free probability recipe.

Moment-cumulant relations

The free cumulants $\kappa_n(\mu)$ are defined by inverting the moment-cumulant relation using non-crossing partitions rather than all partitions. Voiculescu showed $\kappa_n(\mu \boxplus \nu) = \kappa_n(\mu) + \kappa_n(\nu)$ .

R-transform as free cumulant generating function

$R_\mu(z) = \sum_{n=0}^\infty \kappa_{n+1}(\mu) z^n$ . Hence additivity of free cumulants implies additivity of R-transforms. Functionally inverting via $m_{\mu \boxplus \nu}$ then recovers the density of the free convolution.

,

Example: R-Transform of Marchenko-Pastur + Semicircle

Compute the R-transform of (a) the standard Wigner semicircle measure $\mu_W$ on $[-2, 2]$ and (b) the Marchenko-Pastur measure $\mu_{\text{MP}}$ with ratio $\beta$ . State the R-transform of their free convolution.

Solution

Wigner semicircle

The Stieltjes transform of the standard semicircle satisfies $m_W(z)^2 + z m_W(z) + 1 = 0$ , so $m_W(z) = (-z + \sqrt{z^2 - 4})/2$ (with sign chosen for analyticity in $\mathbb{C}^+$ ). Inverting: $K_W(z) = z + 1/z$ , hence $R_W(z) = z$ .

Marchenko-Pastur

From the MP fixed-point equation $\beta z m^2 + (z - (\beta-1)) m + 1 = 0$ , one can derive $R_{\text{MP}}(z) = \beta/(1 - z)$ for $\beta \in (0, 1]$ (after subtracting the atom at zero when $\beta > 1$ ).

Free convolution

$R_{\mu_W \boxplus \mu_{\text{MP}}}(z) = z + \beta/(1 - z)$ . The resulting measure is obtained by functionally inverting $K(z) = z + \beta/(1-z) + 1/z$ . This measure describes, for example, the spectrum of a signal matrix corrupted by additive Hermitian noise with semicircular limit.

Free Convolution: Sum of Independent Spectra

Compute $\mu_A \boxplus \mu_B$ numerically from the R-transforms. Slide the parameters to see how the sum of two random matrix spectra differs from the classical convolution.

Parameters

MP aspect ratio

\beta

0.5

Semicircle radius

\sigma

1

Matrix size

N

(for overlay histogram)200

Definition:
S-Transform

For a compactly supported probability measure $\mu$ on $[0, \infty)$ with moment generating series $\psi_\mu(z) = \sum_{k \geq 1} m_k z^k$ (where $m_k = \int x^k\,d\mu(x)$ ), the S-transform is $S_\mu(z) = \frac{1 + z}{z} \cdot \chi_\mu(z),$ where $\chi_\mu$ is the inverse of $\psi_\mu$ under composition. The S-transform linearizes multiplicative free convolution: if $\mathbf{A}_N, \mathbf{B}_N$ are asymptotically free and positive, then $S_{\mu_A \boxtimes \mu_B}(z) = S_{\mu_A}(z) \cdot S_{\mu_B}(z)$ .

The S-transform is to products of free positive matrices what the R-transform is to sums of free Hermitian matrices. Together they cover the main operations needed to analyze iterative receiver architectures.

Theorem: Spectrum of an Isometric Precoded Channel

Let $\mathbf{H}$ be $n_r \times n_t$ i.i.d. Gaussian with entries of variance $1/n_r$ , and let $\mathbf{W}$ be $n_t \times k$ with orthonormal columns (isometric precoder). Assume $n_r, n_t, k \to \infty$ with $n_t/n_r \to \beta$ and $k/n_t \to \alpha$ . The LSD of $\mathbf{H}\mathbf{W}(\mathbf{H}\mathbf{W})^H$ is obtained from the free multiplicative convolution of the Marchenko-Pastur measure with the Bernoulli measure $\alpha\delta_1 + (1-\alpha)\delta_0$ .

Projecting onto $k$ out of $n_t$ orthonormal directions is multiplication by a projector with Bernoulli spectrum. Asymptotically, multiplication of free positive matrices is governed by the product of S-transforms, giving an explicit expression for the resulting LSD.

Proof

S-transform of the projector

The projector $\mathbf{W}\mathbf{W}^{H}$ has spectrum $\alpha\delta_1 + (1-\alpha)\delta_0$ (asymptotically). A direct computation gives $S_{\text{proj}}(z) = (1+z)/(\alpha + z)$ .

S-transform of MP

From the Marchenko-Pastur moments, $S_{\text{MP}}(z) = 1/(1 + \beta z)$ .

Multiply and invert

The S-transform of the product spectrum is $S(z) = \frac{1+z}{(\alpha + z)(1 + \beta z)}$ . Inversion yields the Stieltjes transform of the resulting LSD and, by Plemelj, its density.

,

Free Probability in Iterative Receivers

The most striking application of free probability in wireless is the asymptotic analysis of iterative (turbo) receivers. Each iteration combines the current estimate with fresh observations through matrix operations. Because the iterates become asymptotically free of the next set of random matrices, the spectral dynamics of the receiver reduce to a scalar recursion in the R- or S-transforms. The CommIT group has used this machinery extensively in its work on large-scale precoding and cell-free MIMO analysis.

The Three Transforms of Random Matrix Theory

Transform	Linearizes	Key identity	Use case
Stieltjes $m_\mu(z)$	Spectrum inversion (Plemelj)	$m(z) = \int (x-z)^{-1}d\mu$	Recover density from fixed-point
R-transform $R_\mu(z)$	Additive free convolution	$R_{\mu \boxplus \nu} = R_\mu + R_\nu$	Sum of Hermitian matrices
S-transform $S_\mu(z)$	Multiplicative free convolution	$S_{\mu \boxtimes \nu} = S_\mu \cdot S_\nu$	Product of positive matrices

Common Mistake: Assuming Freeness Without Checking

Mistake:

Applying the R-transform additivity formula to two random matrices that are independent but not asymptotically free (e.g., two random diagonal matrices in the same basis).

Correction:

Asymptotic freeness requires one matrix to be rotated by an independent Haar-distributed unitary, or to satisfy moment conditions that decouple mixed traces. Two random diagonal matrices in the standard basis are not free — their eigenbases are identical. The correct tool in that case is classical convolution, not free convolution.

Historical Note: Voiculescu's Revolution

1985-2005

Dan Voiculescu introduced free probability in the 1980s as a tool in operator algebras — specifically, to study the isomorphism problem for free group factors. The connection to random matrices came in 1991, when Voiculescu proved that independent unitarily-invariant random matrices are asymptotically free. For a decade, the theory stayed within pure mathematics. In the early 2000s, Tulino and Verdú, then Couillet and Debbah, adapted the machinery for wireless communications, launching what is now a standard analytical toolkit.

Quick Check

What is the R-transform of the standard Marchenko-Pastur measure with aspect ratio $\beta \in (0, 1]$ ?

$R_\mu(z) = \beta/(1-z)$

$R_\mu(z) = z$

$R_\mu(z) = 1/(1 + \beta z)$