---
title: Second Order Free Cumulants Overview
url: https://www.emergentmind.com/topics/second-order-free-cumulants
type: topic
---

# Second Order Free Cumulants Overview

Searching arXiv for recent and foundational papers on second-order free cumulants and related variants.
Second-order free cumulants are the fluctuation-level analogues of free cumulants in a second-order noncommutative probability space $(A,\varphi,\varphi_2)$. While first-order free cumulants $\kappa_n$ are indexed by noncrossing partitions, second-order cumulants $\kappa_{p,q}$ are indexed by annular noncrossing structures and partitioned permutations, reflecting the covariance of traces and other fluctuation moments. In the formulation developed for second-order even elements and second-order $R$-diagonal elements, they admit explicit product, square, and modulus formulas, second-order $R$-transform identities, and direct random-matrix realizations such as products of Ginibre and Wishart matrices [1912.09873].

## 1. Foundational framework

A second-order probability space is a triple $(A,\varphi,\varphi_2)$, where $A$ is a unital $^\ast$-algebra over $\mathbb{C}$, $\varphi:A\to\mathbb{C}$ is a tracial state with $\varphi(1)=1$, and $\varphi_2:A\times A\to\mathbb{C}$ is a bilinear, symmetric functional that is tracial in each variable and satisfies $\varphi_2(1,a)=\varphi_2(a,1)=0$. For $a\in A$, the moments are $\varphi(a^n)$, while $\varphi_2(a^m,a^n)$ are fluctuation moments, interpreted as covariances of traces in large-$N$ random-matrix models [1912.09873].

First-order free cumulants are defined through the noncrossing moment–cumulant relation
\[
\varphi(a_1\cdots a_n)=\sum_{\pi\in NC(n)}\kappa_\pi(a_1,\dots,a_n),
\]
with multiplicativity over the blocks of $\pi$. Second-order cumulants are defined on annular combinatorial objects. If $S_{NC}(m,n)$ denotes the noncrossing annular permutations and $PS_{NC}(m,n)'$ the partitioned permutations obtained by joining exactly one cycle on the outer circle to one cycle on the inner circle, then
\[
\varphi_2(a_1\cdots a_m,a_{m+1}\cdots a_{m+n})
=\sum_{\pi\in S_{NC}(m,n)}\kappa_\pi(a_1,\dots,a_{m+n})
+\sum_{(U,\pi)\in PS_{NC}(m,n)'}\kappa_{(U,\pi)}(a_1,\dots,a_{m+n}).
\]
The second sum contains one second-order cumulant $\kappa_{p,q}$ on the marked pair of cycles and first-order cumulants on the remaining cycles [1912.09873].

Second-order freeness is characterized by vanishing mixed cumulants. Subalgebras $A_1,\dots,A_s\subset A$ are second-order free if all mixed first-order and second-order cumulants vanish. In particular, for a cumulant $\kappa_n(a_1,\dots,a_n)$ with $a_i\in A_{j_i}$, the cumulant is zero whenever there exist $i\neq i'$ with $j_i\neq j_{i'}$, and the same vanishing condition holds for $\kappa_{m,n}$ [1912.09873].

## 2. Second-order even elements and second-order $R$-diagonal elements

A self-adjoint element $x=x^\ast$ is second-order even if
\[
\varphi(x^{2n+1})=0 \quad \forall n\ge 0,\qquad
\varphi_2(x^p,x^q)=0 \quad \text{unless both }p,q\text{ are even}.
\]
Equivalently, $\kappa_{2n+1}(x,\dots,x)=0$ and $\kappa_{p,q}(x,\dots,x)=0$ unless $p,q$ are even. Its first-order determining sequence is $\beta_n^{(x)}:=\kappa_{2n}(x,\dots,x)$, and its second-order determining sequence $\beta_{p,q}^{(x)}$ is defined by $\kappa_{2p,2q}^{(x)}$ together with a sum over even annular noncrossing permutations with only through cycles satisfying a parity-preserving condition [1912.09873].

The principal structural result for second-order even elements is the formula
\[
\kappa_{p,q}(x^2,\dots,x^2)=\sum_{(U,\pi)\in PS_{NC}(p,q)}\beta^{(x)}_{(U,\pi)}.
\]
Equivalently,
\[
\kappa^{(x^2)}=\beta^{(x)}*\zeta,\qquad \beta^{(x)}=\kappa^{(x^2)}*\mu,
\]
where $\zeta$ and $\mu$ are the second-order zeta and Möbius functions. The proof proceeds by expanding cumulants with products as entries, specializing to $x^2$, and using parity constraints to identify precisely the contributing annular permutations [1912.09873].

An element $a$ is $R$-diagonal if the only possibly nonzero first-order $^\ast$-cumulants are the alternating even ones,
\[
\kappa_{2n}(a,a^\ast,\dots,a,a^\ast),\qquad
\kappa_{2n}(a^\ast,a,\dots,a^\ast,a).
\]
It is second-order $R$-diagonal if, in addition, the only non-vanishing second-order cumulants are alternating on each circle:
\[
\kappa_{2p,2q}(a,a^\ast,\dots,a,a^\ast)
=
\kappa_{2p,2q}(a^\ast,a,\dots,a^\ast,a).
\]
The determining sequences are
\[
\beta_n^{(a)}:=\kappa_{2n}(a,a^\ast,\dots,a,a^\ast),\qquad
\beta_{p,q}^{(a)}:=\kappa_{2p,2q}(a,a^\ast,\dots,a,a^\ast).
\]
For such an element,
\[
\kappa_{p,q}(aa^\ast,\dots,aa^\ast)=\sum_{(U,\pi)\in PS_{NC}(p,q)}\beta^{(a)}_{(U,\pi)},
\]
or equivalently $\kappa^{(aa^\ast)}=\beta^{(a)}*\zeta$ and $\beta^{(a)}=\kappa^{(aa^\ast)}*\mu$ [1912.09873].

A further stability property is that if $r$ is second-order $R$-diagonal and $b$ is second-order free from $r$, then $rb$ is also second-order $R$-diagonal. The combinatorial mechanism is that, after expanding cumulants of products, freeness forces blocks to lie entirely in the $r$-family or the $b$-family, while alternation in the $r$-blocks propagates to the full cumulant and annihilates all non-alternating patterns [1912.09873].

## 3. Generating series, Möbius inversion, and computation

For a fixed element $a$, the cumulant series are
\[
R(z)=\frac{1}{z}\sum_{n\ge1}\kappa_n^a z^n,\qquad
R(z,w)=\frac{1}{zw}\sum_{p,q\ge1}\kappa_{p,q}^a z^p w^q.
\]
The corresponding moment series are
\[
G(z)=\frac{1}{z}\sum_{n\ge0}\varphi(a^n)z^{-n},\qquad
G(z,w)=\frac{1}{zw}\sum_{p,q\ge1}\varphi_2(a^p,a^q)\,z^{-(p+1)}w^{-(q+1)}.
\]
They satisfy the first-order identity
\[
\frac{1}{G(z)}+R(G(z))=z
\]
and the second-order $R$-transform identity
\[
G(z,w)=G'(z)G'(w)\,R(G(z),G(w))
+\frac{\partial^2}{\partial z\,\partial w}\log\!\left(\frac{G(z)-G(w)}{z-w}\right).
\]
The logarithmic correction can also be written as
\[
\frac{G'(z)G'(w)}{(G(z)-G(w))^2}-\frac{1}{(z-w)^2}.
\]
These identities package the passage from first-order laws to fluctuation laws into a closed analytic form [1912.09873].

For second-order even elements and second-order $R$-diagonal elements, the determining sequences are computationally central. In the even case, if
\[
B(z)=\frac{1}{z}\sum_{n\ge1}\beta_n z^n,\qquad
B(z,w)=\frac{1}{zw}\sum_{p,q\ge1}\beta_{p,q}z^p w^q,
\]
and
\[
C(z)=\frac{1}{z}+\frac{1}{z}\sum_{n\ge1}\kappa_n z^{-n},\qquad
C(z,w)=\frac{1}{zw}\sum_{p,q\ge1}\kappa_{p,q}z^{-p}w^{-q},
\]
then
\[
\frac{1}{C(z)}+B(C(z))=z,\qquad
C(z,w)=C'(z)C'(w)\,B(C(z),C(w))
+\frac{\partial^2}{\partial z\,\partial w}\log\!\left(\frac{C(z)-C(w)}{z-w}\right).
\]
The same series relations hold in the $R$-diagonal case, with $\beta$ denoting the alternating determining sequences [1912.09873].

The core reduction step for composite variables is the products-as-entries expansion. If
\[
A_1=a_1\cdots a_{n_1},\ \dots,\ A_{r+s}=a_{n_1+\cdots+n_{r+s-1}+1}\cdots a_{n_1+\cdots+n_{r+s}},
\]
then
\[
\kappa_{r,s}(A_1,\dots,A_{r+s})
=
\sum_{(V,\pi)\in PS_{NC}(p,q)}\kappa_{(V,\pi)}(a_1,\dots,a_{p+q}),
\]
with an annular separation condition at the product boundaries. This is the principal computational engine for deriving cumulants of $x^2$, $aa^\ast$, products, commutators, and anti-commutators from cumulants of the underlying entries [1912.09873].

## 4. Random-matrix realizations and exact fluctuation formulas

A second-order semicircular element $s$ has $\kappa_2(s,s)=1$, $\kappa_n=0$ for $n\neq2$, and all second-order cumulants vanish. Its determining sequences satisfy $\beta_1=1$, $\beta_n=0$ for $n>1$, while for the second-order sequence only spoke diagrams contribute:
\[
\beta_{p,q}=
\begin{cases}
p,& p=q,\\
0,& p\neq q.
\end{cases}
\]
Consequently,
\[
\kappa_{p,q}(s^2,\dots,s^2)=\sum_{k\ge1}k\binom{p}{k}\binom{q}{k}
=
p\binom{p+q-1}{p}.
\]
This identifies the fluctuation cumulants of the square of a second-order semicircular element in completely explicit binomial form [1912.09873].

For a circular element $c$, the only nonvanishing first-order cumulants are $\kappa_2(c,c^\ast)=\kappa_2(c^\ast,c)=1$, and all second-order cumulants vanish. Hence $\beta_1=1$, $\beta_n=0$ for $n>1$, and $\beta_{p,q}=0$ for all $p,q$. The main formula for $aa^\ast$ then gives
\[
\kappa_{p,q}(cc^\ast,\dots,cc^\ast)=0,
\]
so the fluctuation moments are purely combinatorial:
\[
\phi_2((cc^\ast)^p,(cc^\ast)^q)=|S_{NC}(p,q)|
=
\frac{pq}{2(p+q)}\binom{2p}{p}\binom{2q}{q}.
\]
This is the basic large-$N$ covariance formula for the complex Ginibre singular-value model [1912.09873].

For products of second-order free circular elements $c_1,\dots,c_k$, let
\[
h=c_1c_2\cdots c_k c_k^\ast\cdots c_2^\ast c_1^\ast.
\]
Then $\kappa^{(h)}=\zeta^{*(k+1)}$, and the fluctuation moments are
\[
\phi_2(h^p,h^q)
=
\frac{k}{k+1}\frac{pq}{p+q}
\binom{(k+1)p}{p}\binom{(k+1)q}{q}.
\]
For $m$ independent complex Wishart factors,
\[
s_m=x_1x_1^\ast\cdots x_mx_m^\ast,
\]
the same convolution method yields
\[
\phi_2(s_m^p,s_m^q)
=
\frac{m}{m+1}\frac{pq}{p+q}
\binom{(m+1)p}{p}\binom{(m+1)q}{q}.
\]
The case $m=2$ recovers the Dartois–Forrester fluctuation formula
\[
\phi_2(s_2^p,s_2^q)
=
\frac{2}{4}\frac{pq}{p+q}\binom{3p}{p}\binom{3q}{q}.
\]
Throughout these models, the standard large-$N$ normalization is that $\phi$ is the $1/N$-trace and $\phi_2$ is the $1/N^2$-scaled covariance of linear statistics [1912.09873].

A related conjugation identity is that for any $a$ second-order free from a circular $c$,
\[
\kappa^{(cac^\ast)}=\kappa^{(a)}*\zeta.
\]
This implies that the fluctuation transform of $cac^\ast$ is obtained from that of $a$ by one $\zeta$-convolution. When $a$ is deterministic, the second-order generating function reduces to the logarithmic derivative term, matching the fluctuation structure of $WAW^\ast$ ensembles [1912.09873].

## 5. Real and quaternionic second-order cumulants

The complex annular theory does not directly describe real matrix ensembles. Real second-order freeness adds an involution $x\mapsto x^t$ and an orientation-reversing annulus. For centered cyclically alternating families in a real second-order probability space, the covariance formula is
\[
\varphi_2\big(\mathaccent"7017{a}_1\cdots\mathaccent"7017{a}_p,\,
\mathaccent"7017{b}_1\cdots\mathaccent"7017{b}_q\big)=0\quad\text{if }p\neq q,
\]
and for $p=q\ge2$,
\[
\varphi_{2}\big(\mathaccent"7017{a}_{1}\cdots\mathaccent"7017{a}_{p},\,
\mathaccent"7017{b}_{1}\cdots\mathaccent"7017{b}_{p}\big)
=
\sum_{k=0}^{p-1}\prod_{i=1}^{p}\varphi_{1}\big(\mathaccent"7017{a}_{i}\mathaccent"7017{b}_{k-i}\big)
+
\sum_{k=0}^{p-1}\prod_{i=1}^{p}\varphi_{1}\big(\mathaccent"7017{a}_{i}\mathaccent"7017{b}_{k+i}^{t}\big).
\]
The second sum is the distinctive real contribution: it corresponds to annular spoke diagrams with opposite orientation and introduces the transpose. Real Ginibre matrices, Gaussian orthogonal matrices, and real Wishart matrices are asymptotically real second-order free, and they do not satisfy the complex definition of second-order freeness satisfied by their complex analogues [1101.0422].

In the combinatorial formulation of real second-order cumulants, the moment–cumulant expansion contains three types of terms: annular noncrossing permutations in the forward orientation, annular noncrossing permutations in reversed orientation with $(y_q^t,\dots,y_1^t)$, and disc-with-one-through-block contributions indexed by $PS'(p,q)$. The corresponding cumulant inversion is governed by an extended poset $Ssd(p,q)$ and its Möbius function $\mu_{sd}$, rather than only the standard annular noncrossing poset [1808.10589].

The real theory also makes explicit the connection with matrix cumulants and topological expansion. Two-vertex matrix cumulants have a finite $N\to\infty$ limit after multiplication by $N^2$, and that limit equals the real second-order free cumulant. In this sense, second-order free cumulants arise as vertex-cumulant limits of orthogonally invariant random matrices. The quaternionic version is parallel in structure but uses $\operatorname{Re}\circ\operatorname{tr}$, the normalization $2N$, and modified Weingarten asymptotics [1808.10589].

A common misconception is that the complex annular formalism merely carries over to real ensembles after replacing complex conjugation by transpose. The real and quaternionic theories show that the change is structural, not cosmetic: the combinatorial domain itself is enlarged by orientation-reversing annuli and one-through-block disc terms, and the asymptotic topologies include nonorientable contributions [1101.0422].

## 6. Later developments, finite-$N$ precursors, and higher-order generalizations

Recent work extends second-order free cumulants in several directions. For finite free cumulants, the genus expansion of the finite multiplicative free convolution organizes the $1/d$ correction by planar annular non-crossing permutations on two circles. If
\[
\alpha_{r,s}=\sum_{\pi\in S_{NC}(r,s)}\kappa_\pi,
\]
then the second-order generating function is
\[
G(z,w)=\frac{1}{zw}\sum_{r,s\ge1}\alpha_{r,s}z^{-r}w^{-s}
=
\frac{\partial^2}{\partial z\,\partial w}
\left[\log\!\left(\frac{F_\mu(z)-F_\mu(w)}{z-w}\right)\right].
\]
In the genus expansion, the planar single-circle sector gives the usual free multiplicative convolution, while the planar two-circle sector gives the annular second-order contribution. This annular sector also yields explicit infinitesimal transforms
\[
G_{\mathrm{inf}}(z)=\frac{G_\mu''(z)}{2G_\mu'(z)}-\frac{G_\mu'(z)}{G_\mu(z)},\qquad
R_{\mathrm{inf}}(z)=\frac{K_\mu''(z)}{2K_\mu'(z)}+\frac{1}{z},
\]
with concrete Hermite and Laguerre examples [2108.08489].

A distinct finite-$N$ approach introduces $\mathrm{U}(N)$-invariant polynomial precursors $K_n^{(N)}$ of free cumulants. They converge to the free cumulants as $N\to\infty$, are additive under averaging over sums of $\mathrm{U}(N)$ conjugacy orbits, and have a topological expansion whose genus-$1$ term is the $1/N^2$ correction. In the inverted moment–precursor relation,
\[
m_n(A)=\sum_{g\ge0}N^{-2g}\sum_{\alpha\vdash n}P_g(\alpha)\,K_\alpha(A),
\]
the $g=1$ term is the second-order free cumulant correction. Fluctuation estimates for the centered additive error $\delta K_n$ show Gaussian limits with variance of order $N^{-2}$, consistent with second-order freeness [2508.21483].

The product calculus has also been extended beyond squares and moduli. For second-order free $a$ and $b$, explicit formulas are now available for the second-order cumulants of $ab$, $ab-ba$, and $ab+ba$, all indexed by special subsets of non-crossing partitioned permutations. At the lowest level,
\[
\kappa_{1,1}^{ab}
=
\kappa_{2}^{a}\kappa_{2}^{b}
+\kappa_{1,1}^{a}(\kappa_{1}^{b})^2
+\kappa_{1,1}^{b}(\kappa_{1}^{a})^2,
\]
\[
\kappa_{1,1}^{ab-ba}=2\kappa_{2}^{a}\kappa_{2}^{b},
\]
\[
\kappa_{1,1}^{ab+ba}
=
2\kappa_{2}^{a}\kappa_{2}^{b}
+4\kappa_{1,1}^{a}(\kappa_{1}^{b})^2
+4\kappa_{1,1}^{b}(\kappa_{1}^{a})^2.
\]
The commutator and anti-commutator cases had not been studied before at second order, and the product formula generalizes the earlier case in which the individual second-order cumulants vanish. The semicircular specialization yields closed formulas for $\kappa_{n,m}^{ab}$, $\kappa_{n,m}^{ab-ba}$, and $\kappa_{n,m}^{ab+ba}$ [2507.21031].

Tensorial generalizations place second-order cumulants in the broader setting of local-unitary invariant random tensors. In the matrix-product scaling, the second-order sector is characterized by invariants with $\Omega(\sigma)+K(\sigma)=2$: either $K(\sigma)=2$ with both components melonic, or $K(\sigma)=1$ with $\Omega(\sigma)=1$. The exact moment-to-cumulant relations involve forests of permutations, Möbius weights, and planar monotone Hurwitz numbers; for $D=1$, they recover the matrix case. The field remains unsettled, however: several points of view have been proposed, and it is unclear at this point whether they lead to the same notions of tensorial free cumulants and freeness [2605.01887].

Across these developments, second-order free cumulants retain the same core role. They are the annular, fluctuation-level cumulants that interpolate between free probability, random-matrix covariance asymptotics, finite-$N$ genus expansions, and newer tensorial or infinitesimal formalisms, while preserving the defining principles of vanishing mixed cumulants, Möbius inversion, and noncrossing annular combinatorics [1912.09873].

Source: https://www.emergentmind.com/topics/second-order-free-cumulants