Free Probability: Foundations¶

Job: free independence in noncommutative probability spaces, free additive/multiplicative convolution, $R$- and $S$-transforms, and asymptotic freeness of random matrices.

Core¶

Noncommutative probability spaces¶

A noncommutative probability space is a pair $(A,\varphi)$ with $A$ a unital algebra over $\mathbb{C}$ and $\varphi:A\to\mathbb{C}$ linear, $\varphi(1)=1$ (a state if $A$ is $*$-algebra, $\varphi(a^*a)\ge 0$; often a faithful tracial state).

Distribution of self-adjoint $x=x^*\in A$: the moment sequence $\varphi(x^n)$, or the compactly supported measure $\mu_x$ when moments match $\int t^n\,d\mu_x$ (Riesz). Joint distributions of tuples: all mixed moments $\varphi(x_{i_1}\cdots x_{i_k})$.

Classical	Free / NC
$(\Omega,\mathbb{P})$, r.v. $L^\infty$	$(A,\varphi)$
Independence	freeness
Characteristic function	Cauchy / moment generating via $R,S$
Convolution of laws	free convolution $\boxplus$

Free independence¶

Unital subalgebras $A_1,\ldots,A_n\subset A$ are free (w.r.t. $\varphi$) if for all $k\ge 1$, $a_j\in A_{i_j}$ with $i_1\neq i_2\neq\cdots\neq i_k$ (adjacent indices different) and $\varphi(a_j)=0$, $$ \varphi(a_1a_2\cdots a_k)=0. $$ Elements are free when the unital algebras they generate are free. Free family of random variables: freeness of generated algebras.

Free cumulants $\kappa_n:A^n\to\mathbb{C}$ (Speicher): multilinear forms defined via moment-cumulant formula over noncrossing partitions $NC(n)$: $$ \varphi(a_1\cdots a_n)=\sum_{\pi\in NC(n)}\kappa_\pi(a_1,\ldots,a_n). $$ Theorem. Subalgebras are free iff all mixed free cumulants vanish (cumulants with arguments from more than one distinct subalgebra).

Free convolution¶

For compactly supported probability measures $\mu,\nu$ on $\mathbb{R}$, there exist free self-adjoints $x,y$ in some $(A,\varphi)$ with laws $\mu,\nu$; the law of $x+y$ depends only on $\mu,\nu$ and is the free additive convolution $\mu\boxplus\nu$.

Free multiplicative convolution $\mu\boxtimes\nu$: law of $xy$ for free positive (or unitary) variables — defined carefully on $\mathbb{R}_+$ or on the circle.

Operation	Classical	Free
Sum of independent	$\mu*\nu$	$\mu\boxplus\nu$
CLT limit	Gaussian	semicircle
Poisson limit	Poisson	Marchenko–Pastur (free Poisson)

Semicircular family. Centered $s$ with $\kappa_2(s,s)=\sigma^2$ and $\kappa_n=0$ for $n\neq 2$: law $\sigma_{\mathrm{sc}}$ scaled. Free CLT: free i.i.d. centered variance $\sigma^2$ partial sums $\to$ semicircle.

$R$-transform¶

The Cauchy–Stieltjes transform $G_\mu(z)=\int\frac{1}{z-t}\,d\mu(t)$ ($z\notin\operatorname{supp}\mu$). Invert near infinity: $K_\mu(G_\mu(z))=z$ for large $z$. $R$-transform: $$ R_\mu(z)=K_\mu(z)-\frac{1}{z}=\sum_{n\ge 0}\kappa_{n+1}(\mu)\,z^n. $$

Theorem (Voiculescu). $R_{\mu\boxplus\nu}=R_\mu+R_\nu$.

Computational recipe. Compute $G_\mu$, invert to $K_\mu$, form $R_\mu$, add, invert back to get $G_{\mu\boxplus\nu}$ and recover $\mu\boxplus\nu$ by Stieltjes inversion.

Examples.

Law	$R(z)$
Semicircle radius $2\sigma$ (var $\sigma^2$)	$\sigma^2 z$
Free Poisson (rate $\lambda$, jump $1$)	$\lambda/(1-z)$ (Marchenko–Pastur family)
Point mass $\delta_a$	$a$

$S$-transform¶

For measures on $\mathbb{R}_+$ with nonzero mean (or suitable unitaries), the $S$-transform satisfies $$ S_{\mu\boxtimes\nu}=S_\mu S_\nu. $$ Defined from the moment series $\psi(z)=\int\frac{zt}{1-zt}\,d\mu(t)$ via $\chi$ inverse of $\psi$ and $S(z)=\frac{1+z}{z}\chi(z)$. Tool for products of free positive matrices (Wishart products, etc.).

Asymptotic freeness of random matrices¶

Theorem (Voiculescu et al.). Let $H_N^{(1)},\ldots,H_N^{(k)}$ be independent Wigner matrices (suitable moments) and/or independent Haar unitaries $U_N^{(j)}$ on $\mathrm{M}_N(\mathbb{C})$. Then in the noncommutative probability space $\bigl(\mathrm{M}_N,\varphi_N=\frac1N\operatorname{Tr}\bigr)$, the family is asymptotically free: mixed moments of noncommutative polynomials converge to those of a free family with the corresponding limiting laws (semicircular, Haar unitary, etc.).

Consequences.

ESD of $H_N^{(1)}+H_N^{(2)}$ $\to$ $\sigma\boxplus\sigma$ (dilation of semicircle — still semicircle after scaling).
Polynomials in several independent ensembles: limiting spectral law via free calculus.
Diagonal deterministic $D_N$ with converging ESD $\mu$, unitarily invariant noise: $U_ND_NU_N^*+H_N$ has free convolution limit $\mu\boxplus\sigma_{\mathrm{sc}}$ under hypotheses (additive deformed Wigner).

Free entropy / free Fisher information (pointer). Voiculescu’s free entropy $\chi$ quantifies “NC dimension”; used in structural results on free group factors and large deviations for matrix microstates.

See it¶

Marchenko–Pastur via freeness. Sample covariance $X_NX_N^*$ with rectangular Gaussian $X_N$ is asymptotically a free Poisson compound of free semicircular / Wishart constructions; free Poisson with rate $c$ and jump $1$ is the Marchenko–Pastur law $$ d\mathrm{MP}_c=\frac{\sqrt{(b-x)(x-a)}}{2\pi x}\,dx+\text{possible atom at }0, $$ $a=(1-\sqrt{c})^2$, $b=(1+\sqrt{c})^2$ for aspect ratio $c$ — matching classical RMT without orthogonal polynomials.

Free product of groups. On the reduced free product of $C^*$-algebras (or $L(\mathbb{F}_n)*L(\mathbb{F}_m)\cong L(\mathbb{F}_{n+m})$ in von Neumann setting), natural subalgebras are free w.r.t. the free product state — pure operator-algebraic source of freeness beyond matrices.

Check¶

Using $R$-transforms, compute the free convolution of two centered semicircular laws of variances $\sigma^2$ and $\tau^2$.

$R_1(z)=\sigma^2 z$, $R_2(z)=\tau^2 z$, so $R_{1\boxplus 2}(z)=(\sigma^2+\tau^2)z$, hence semicircular of variance $\sigma^2+\tau^2$. (Free sum of free semicirculars is semicircular — free Gaussian stability.)

State the moment-cumulant formula for $n=2,3$ and interpret freeness for two centered free elements $x,y$.

$n=2$: $\varphi(ab)=\kappa_2(a,b)+\kappa_1(a)\kappa_1(b)$. $n=3$: moments sum $\kappa_3+\kappa_1\kappa_2$ terms over $NC(3)$. Centered free $x,y$: $\varphi(xy)=0$, $\varphi(xyx)=\kappa_2(x,x)\varphi(y)?$ carefully — alternating centered products vanish: $\varphi(xy)=\varphi(x)\varphi(y)=0$, $\varphi(xyxy)=\kappa_2(x,x)\kappa_2(y,y)=\varphi(x^2)\varphi(y^2)$.

Pitfalls¶

Noncrossing partitions only — full partition lattice gives classical cumulants. Mixing the two is a common combinatorial error.
$R$-transform radius of convergence / analytic continuation needed for unbounded measures; formal series suffice for moment determinacy in compact support cases.
Multiplicative free convolution for variables that are not positive requires unitary or carefully polarized settings.
Asymptotic freeness $\neq$ freeness at finite $N$ — all freeness statements for matrices are limit statements unless built from exact free products.

Refs¶

Nica–Speicher, Lectures on the Combinatorics of Free Probability
Voiculescu–Dykema–Nica, Free Random Variables
Mingo–Speicher, Free Probability and Random Matrices
Anderson–Guionnet–Zeitouni, An Introduction to Random Matrices (free probability chapters)

Next¶

FP index · Free Entropy & Fisher · Operator-Algebraic Freeness · Free Stochastic Calculus · Subordination & Harmonic Analysis · Random matrices foundations · RMT / spiked — the deformed semicircle / free convolution, the BBP outlier via the $R$-transform / subordination · RMT / circular — the circular law via the Hermitization / the free Poisson, the single ring / the Brown measure · Probability foundations · Functional analysis · Mathematical physics · NCG foundations · Operator algebras — $\mathrm{II}_1$ factors and free group factors

Law	\(R(z)\)
Semicircle radius \(2\sigma\) (var \(\sigma^2\))	\(\sigma^2 z\)
Free Poisson (rate \(\lambda\), jump \(1\))	\(\lambda/(1-z)\) (Marchenko–Pastur family)
Point mass \(\delta_a\)	\(a\)