arXiv reader

Cumulants, Spreadability and the Campbell-Baker-Hausdorff Series

Takahiro Hasebe and Franz Lehner

arXiv
1711.00219v3 · PDF · 7 September 2023
Categories
math.OA, math.CO, math.PR
Journal
Doc. Math. 28 (2023) 515-601
DOI
10.4171/dm/923
Licence
CC BY 4.0 The site shows the text.

Abstract

We define spreadability systems as a generalization of exchangeability systems in order to unify various notions of independence and cumulants known in noncommutative probability. In particular, our theory covers monotone independence and monotone cumulants which do not satisfy exchangeability. To this end we study generalized zeta and Möbius functions in the context of the incidence algebra of the semilattice of ordered set partitions and prove an appropriate variant of Faa di Bruno’s theorem. With the aid of this machinery we show that our cumulants cover most of the previously known cumulants. Due to noncommutativity of independence the behaviour of these cumulants with respect to independent random variables is more complicated than in the exchangeable case and the appearance of Goldberg coefficients exhibits the role of the Campbell-Baker-Hausdorff series in this context. In a final section we exhibit an interpretation of the Campbell-Baker-Hausdorff series as a sum of cumulants in a particular spreadability system, thus providing a new derivation of the Goldberg coefficients.

Items 169

Remark

Remark 2.1

1WIJ
  1. In order to be able to include Boolean and monotone products and some other examples (Section ?) we do not a priori assume the unitality. In some examples \mathcal{A} and \mathcal{B} are naturally unital; then we also assume that \varphi is also unital, i.e., \varphi(1_{\mathcal{A}})=1_{\mathcal{B}} and also that involved subalgebras of \mathcal{A} contain the unit of \mathcal{A}.

  2. Usually the involved algebras are \ast-algebras and the linear maps satisfy positivity. However positivity is not essential here and we stick to the algebraic \mathcal{B}-valued setting, which allows to include the interesting example of Lie polynomials in Section ?.

Definition

Definition 2.2

5K97

Let (\mathcal{A}, \varphi) be a \mathcal{B}-ncps.

  1. A spreadability system for (\mathcal{A}, \varphi) is a triplet \mathcal{S}=(\mathcal{U}, \tilde{\varphi}, (\iota^{(i)})_{i=1}^\infty) satisfying the following properties:

    1. (\mathcal{U}, \tilde{\varphi}) is a \mathcal{B}-valued ncps.

    2. \iota^{(i)}: \mathcal{A} \to \mathcal{U} is a homomorphism such that \varphi = \tilde{\varphi} \circ \iota^{(i)} for each i \geq 1. For simplicity, \iota^{(i)}(X) is denoted by X^{(i)}, X \in \mathcal{A}, and we denote by \mathcal{A}^{(i)} the image of \mathcal{A} under \iota^{(i)}.

    3. The identity

      \tilde{\varphi}(X_1^{(i_1)}X_2^{(i_2)}\cdots X_n^{(i_n)}) = \tilde{\varphi}(X_1^{(h(i_1))}X_2^{(h(i_2))}\cdots X_n^{(h(i_n))})

      holds for any X_1, X_2,\ldots, X_n \in \mathcal{A}, any i_1, i_2, \ldots, i_n \in \mathbb{N} and any order preserving map h:\{i_1, i_2, \ldots, i_n\} \to \mathbb{N}, that is, i_p < i_q implies h(i_p)<h(i_q).

  2. A triplet \mathcal{E}=(\mathcal{U}, \tilde{\varphi}, (\iota^{(i)})_{i=1}^\infty) is called an exchangeability system if, in addition to (i), (ii) above, eq. (?) holds for any X_1, X_2, \ldots, X_n \in \mathcal{A}, any i_1, i_2, \ldots, i_n \in \mathbb{N} and any permutation h \in \mathfrak{S}_\infty:=\bigcup_{n\geq 1}\mathfrak{S}_n.

Remark

Remark 2.3

1TQO
  1. In some examples where \mathcal{A}, \mathcal{B}, \varphi are unital, the extended algebra \mathcal{U} is also unital and \iota^{(i)} are unit-preserving. To be specific, this applies to the tensor and free spreadability systems in Section ? and the unshuffle spreadability system in Section ?.

  2. It is easy to see that condition (?) can be rephrased as follows:

    (\iota^{(1)}, \iota^{(2)},\dots) \overset{d}{=} (\iota^{(n_1)}, \iota^{(n_2)},\dots)

    for any strictly increasing sequence (n_i)_{i=1}^\infty \subseteq \mathbb{N}. This is the definition given in [Koe10].

  3. It is straightforward to extend the definition of exchangeability systems (resp., spreadability systems) from \mathbb{N} to an arbitrary set (resp., arbitrary totally ordered set).

Definition

Definition 2.4

1ZJ7
  1. Using the concept of kernel partition from Definition A.10 the condition of spreadability (?) is equivalent to the requirement that

    \tilde\varphi(X_1^{(i_1)} X_2^{(i_2)} \cdots X_n^{(i_n)}) = \tilde\varphi(X_1^{(j_1)} X_2^{(j_2)} \cdots X_n^{(j_n)})

    holds whenever the kernels coincide, i.e., \kappa(i_1,i_2, \ldots, i_n) = \kappa(j_1,j_2,\ldots,j_n), see Definition A.10. That is, the expectation (?) only depends on the ordered kernel set partition \kappa(i_1,i_2,\ldots,i_n). Thus for every ordered set partition \pi \in \mathcal{OP}_n we can define a multilinear functional \varphi_\pi\colon \mathcal{A}^n \to \mathbb{C} by choosing any representative sequence (i_1, i_2, \ldots, i_n) with \kappa(i_1, i_2, \ldots, i_n)=\pi and setting

    \varphi_{\pi}(X_1, X_2, \ldots,X_n)=\tilde{\varphi}(X_1^{(i_1)} X_2^{(i_2)} \cdots X_n^{(i_n)}).

    The invariance (?) ensures that this definition is consistent and does not depend on the choice of the representative.

  2. This generalizes the corresponding notions from exchangeability systems [L04]: given an exchangeability system \mathcal{E}=(\mathcal{U}, \tilde{\varphi}, (\iota^{(i)})_{i=1}^\infty), we can define a multilinear functional \varphi_\pi, this time for any set partition \pi \in \mathcal{P}_n,

    \varphi_{\pi}(X_1, X_2, \ldots,X_n)=\tilde{\varphi}(X_1^{(i_1)} X_2^{(i_2)} \cdots X_n^{(i_n)}),

    where (i_1,i_2,\dots,i_n) is any representative such that \overline{\kappa}(i_1,i_2,\dots,i_n)=\pi, see Definition A.10.

Remark

Remark 2.5

737M

The algebra \mathcal{U} and the homomorphisms (\iota^{(i)})_{i\in\mathbb{N}} of a spreadability system can always be chosen to be the tensor algebra of the copies of \mathcal{A} (see (?) below) with the natural embeddings, respectively, in the following sense. Given a spreadability system \mathcal{S}=(\mathcal{U}, \tilde{\varphi}, (\iota^{(i)})_{i=1}^\infty) for a \mathcal{B}-ncps (\mathcal{A},\varphi), we set \hat{\mathcal{U}} := \bigsqcup_{i=1}^\infty\mathcal{A}_i where \mathcal{A}_{i} are copies of \mathcal{A} and set \hat\iota^{(i)} \colon \mathcal{A} =\mathcal{A}_i \to \hat{\mathcal{U}} to be the natural embeddings. Then we can define \hat \varphi \colon \hat{\mathcal{U}} \to \mathcal{B} by setting

\hat \varphi (\hat\iota^{(i_1)}(X_1) \hat\iota^{(i_2)}(X_2) \cdots \hat\iota^{(i_n)}(X_n)) := \tilde{\varphi} (X_1^{(i_1)} X_2^{(i_2)} \cdots X_n^{(i_n)})

for every i_1, i_2,\dots, i_n with i_k \ne i_{k+1}~(k\in[n-1]) and X_1,X_2,\dots, X_n \in \mathcal{A}, giving rise to the spreadability system \hat \mathcal{S}=(\hat{\mathcal{U}}, \hat \varphi, (\hat \iota^{(i)})_{i\ge1}) for (\mathcal{A}, \varphi). Now (\hat\iota^{(i)})_{i\ge1}\overset{d}{=} (\iota^{(i)})_{i\ge1} and thus replacing \mathcal{S} with \hat\mathcal{S} does not change basic results; in particular, the cumulants for \mathcal{S} and for \hat \mathcal{S} are identical, see Theorem 3.11.

In the majority of the examples below we choose \mathcal{U} to be the tensor algebra, in a few cases however, it is more natural to choose another \mathcal{U}. A notable exception is the case of unital algebras, where the unital free product \ast_{i \in \mathbb{N}} \mathcal{A} and the infinite tensor product \bigotimes_{i\in \mathbb{N}} \mathcal{A} are more convenient.

Remark

Remark 2.6

XGO4

For d=1 a binary operation \operatorname*{\odot} satisfying conditions (?)–? and ? is called a quasi-universal product in [Muraki:2002:five]. (Note that (?) easily follows from ?). On the other hand, a binary operation \operatorname*{\odot} satisfying conditions (?)–? is called a natural product [M03]. The main result of [M03] is that ? follows from (?)–?, i.e., a natural product is a quasi-universal product.

Remark

Remark 2.7

P245

A more general setup was discussed in [ManzelSchurmann:2016:noncommutative], where the algebras are allowed to have an additional structure of faces. The theory of cumulants is also developed in [ManzelSchurmann:2016:noncommutative] for universal products with multistates and multifaces. It seems that some modifications are needed in order to adapt the notion of spreadability system to the structure of faces but this issue is not discussed in the present paper.

Definition

Definition 2.8

ISM2

Let d \in\mathbb{N} and \operatorname*{\odot} be a universal product for \operatorname{AlgP}_d satisfying conditions ? and ?. Let \mathcal{A} be an algebra and \varphi = (\varphi^1, \dots, \varphi^d) be a tuple of \mathbb{C}-linear functionals on \mathcal{A}.

  1. A sequence (\mathcal{A}_i)_{i\in [p]} of subalgebras of \mathcal{A} is said to be \operatorname*{\odot}-independent if the identity

    \varphi^j \circ \tilde{\bigsqcup_{i\in[p]}}\gamma_i = \left[\operatorname*{\odot}_{i\in [p]} (\varphi\circ \gamma_i) \right]^j \quad \text{on}\quad \bigsqcup_{i\in [p]} \mathcal{A}_i,

    holds for all j\in[d], where \gamma_i\colon \mathcal{A}_i \hookrightarrow \mathcal{A} is the embedding for i\in [p] and \tilde\sqcup_{i\in[p]}\gamma_i is the canonical arrow \sqcup_{i\in[p]} \mathcal{A}_i \to \mathcal{A}.

    If we further assume ? the above definition is equivalent to the requirement that for any n \in \mathbb{N}, i_1, i_2, \dots, i_n \in [p], X_k \in \mathcal{A}_{i_k}~(k \in[n]) we have

    \varphi^j(X_1 X_2 \cdots X_n) = \sum_{\substack{\sigma \in \mathcal{P}_n\\ \sigma \le \bar\pi}} \sum_{f\colon\, \sigma \to [d]} u^j (\sigma, f; \pi) \prod_{S \in \sigma} \varphi^{f(S)}(X_S), \qquad j \in [d].
  2. For an arbitrary totally ordered set I, a family of subalgebras (\mathcal{A}_i)_{i\in I} is called \operatorname*{\odot}-independent if any finite subfamily is \operatorname*{\odot}-independent in the sense above. If the universal product \operatorname*{\odot} in addition satisfies ? then one can define \operatorname*{\odot}-independence for a family of subalgebras (\mathcal{A}_i)_{i\in I} with any index set I because one can define \operatorname*{\odot}_{i\in I} (\varphi\circ \iota^{(i)}) for any index set I.

  3. A family (S_i)_{i \in I} of subsets of \mathcal{A} with totally ordered index set I is said to be \odot-independent if the family of subalgebras \mathcal{A}_i generated by S_i is \odot-independent.

Used by Proof
Remark

Remark 2.9

SYKN

More examples may be extracted from associative universal products in [BLS96, H10] for d=2 or d=3, but we omit them here.

Definition

Definition 2.10

GMYV

Let \mathcal{S}=(\mathcal{U}, \tilde{\varphi},(\iota^{(i)})_{i \geq 1}) be a spreadability system for a \mathbb{C}-ncps (\mathcal{A}, \varphi). For a set partition \pi\in\mathcal{P}_n we define a multiplicative extension \varphi_{(\pi)} of \varphi by setting

\begin{aligned} &\varphi_{(\pi)}(X_1,X_2,\ldots, X_n) :=\prod_{P\in \pi} \varphi(X_P) \quad \text{for}\quad X_1, X_2,\dots, X_n \in \mathcal{A}, \end{aligned}

where we used notation (?) for the ordered product X_P.

Then \mathcal{S} is said to have a calculation rule if there exists a family of complex numbers \mathbf{C}= \{s(\sigma; \pi)\in\mathbb{C}: n \in \mathbb{N}, \pi\in \mathcal{OP}_n, \sigma\in\mathcal{P}_n, \sigma \leq \bar{\pi}\} such that the equality

\varphi_\pi=\sum_{\substack{\sigma\in\mathcal{P}_n,\\ \sigma \leq \bar{\pi}}} s(\sigma; \pi) \, \varphi_{(\sigma)}

holds as functionals on \mathcal{A}^n for all \pi \in \mathcal{OP}_n and n\in \mathbb{N}. Note that we can always take s(\{1\}; (1))=1 because of \tilde \varphi \circ \iota^{(i)} = \varphi.

Remark

Remark 2.11

7RKI
  1. It is straightforward to see that a spreadability system \mathcal{S}_\odot^1 constructed from a universal product \odot on \operatorname{AlgP}_1 subject to axioms ?, ? and ? has a calculation rule with s(\sigma; \pi):=u^1(\sigma; \pi) (where f is omitted because there is a unique f\colon \sigma \to [1]).

  2. Actually we can construct such a spreadability system for any given family of constants \mathbf{C} =\{s(\sigma; \pi)\in\mathbb{C}: n \in \mathbb{N}, \pi\in \mathcal{OP}_n, \sigma\in\mathcal{P}_n, \sigma \leq \bar{\pi}\} with s(\{1\}; (1))=1 and any given \mathbb{C}-ncps (\mathcal{A}, \varphi). Let \mathcal{U} be the coproduct (i.e., the tensor algebra) of the countable copies of \mathcal{A} in the category of associative algebras and \iota^{(i)}\colon \mathcal{A}\to \mathcal{U} the natural embedding as the ith component. Then we can define a linear functional \tilde{\varphi} on \mathcal{U} as follows: for each tuple X_j \in \mathcal{A}~(1 \le j \le n) and indices i_1,i_2,\dots, i_n \in \mathbb{N} with i_k \ne i_{k+1} (k\in[n-1]) we set

    \tilde \varphi (X_1^{(i_1)} X_2^{(i_2)} \cdots X_n^{(i_n)}) := \sum_{\substack{\sigma\in\mathcal{P}_n,\\ \sigma \leq \bar{\kappa}(i_1,\dots, i_n) }} s(\sigma; \kappa(i_1,\dots, i_n)) \, \varphi_{(\sigma)} (X_1,X_2,\dots, X_n).

    The fact that the value of (?) depends only on the partition \kappa(i_1,i_2,\dots,i_n) induced by the sequence (i_1,i_2,\dots, i_n) yields that \mathcal{S}_{\mathbf{C}}:= (\mathcal{U}, \tilde\varphi, (\iota^{(i)})_{i\ge1}) is indeed a spreadability system.

  3. Muraki’s example [Muraki:private] yields a spreadability system with calculation rule coming from a nonassociative universal product. The V-monotone spreadability system is also such an example. We will investigate other instructive examples later in Examples 2.19 and 3.27.

  4. It is worth mentioning that there are spreadability systems without calculation rules. For instance, the algebra generated by left and right creation operators acting on q-Fock space gives rise to an exchangeability system [Lehner:2005:cumulants3], but it was shown in [LeeuwenMaassen:1996:obstruction] that in this case the individual distributions of independent elements with respect to the vacuum expectation is not sufficient to determine the joint distribution with respect to (\varphi_\pi)_{\pi}, and in particular there is no “q-convolution”. The c-monotone spreadability system \mathcal{S}_{\rm CM} also does not have a calculation rule since the evaluation of \varphi_\pi depends on both \varphi and \psi in general.

Definition

Definition 2.12 (\exch-independence)

QE4R

Let \mathcal{E}=(\mathcal{U},\tilde{\varphi},(\iota^{(i)})_{i\ge1}) be an exchangeability system for a \mathcal{B}-ncps (\mathcal{A},\varphi). Let I\subseteq \mathbb{N}.

  1. Subalgebras (\mathcal{A}_i)_{i\in I} of \mathcal{A} are said to be \mathcal{E}-independent if for any tuple of indices (i_1,i_2,\dots,i_n)\in I^n, any tuple of random variables (X_1,X_2,\dots,X_n) with X_j\in \mathcal{A}_{i_j} and any set partition \pi\in\mathcal{P}_n, we have

    \varphi_\pi(X_1,X_2,\dots,X_n) = \varphi_{\pi\wedge \bar{\kappa}(i_1,i_2,\dots,i_n)}(X_1,X_2,\dots,X_n).
  2. Subsets (S_i)_{i \in I} of \mathcal{A} are said to be \mathcal{E}-independent if the algebras \mathcal{A}_i generated by S_i are \mathcal{E}-independent. In particular, a family \bigl( (X_1(i),X_2(i),\dots,X_n(i)) \bigr)_{i\in I} of random vectors is said to be \mathcal{E}-independent if the subalgebras \mathcal{A}_i generated by its respective entries \{X_1(i),X_2(i), \dots, X_n(i)\} are \mathcal{E}-independent.

Remark

Remark 2.13

U7A1

In the case where \mathcal{A}, \mathcal{B}, \mathcal{U} and \varphi, \iota^{(i)} are unital, it is more natural to assume that the subalgebras \mathcal{A}_i contain the unit of \mathcal{A}. The same applies to Definition 2.14.

Definition

Definition 2.14 (\cS-independence)

UB8M

Let \mathcal{S}=(\mathcal{U},\tilde{\varphi},(\iota^{(i)})_{i\ge1}) be a spreadability system for a given \mathcal{B}-ncps (\mathcal{A},\varphi). Let I \subseteq \mathbb{N}.

  1. A sequence of subalgebras (\mathcal{A}_i)_{i\in I} of \mathcal{A} is said to be \mathcal{S}-independent if for any tuple of indices (i_1,i_2,\dots,i_n)\in I^n, any tuple of random variables (X_1,X_2,\dots,X_n) with X_j\in \mathcal{A}_{i_j} and any ordered set partition \pi\in\mathcal{OP}_n, we have

    \varphi_\pi(X_1,X_2,\dots,X_n) = \varphi_{\pi\curlywedge \kappa(i_1,i_2,\dots,i_n)}(X_1,X_2,\dots,X_n).
  2. A sequence of subsets (S_i)_{i \in I} is said to be \mathcal{S}-independent if the sequence of subalgebras \mathcal{A}_i generated by S_i is \mathcal{S}-independent. In particular, a sequence of random vectors ((X_1(i),X_2(i),\dots,X_n(i)))_{i\in I} is said to be \mathcal{S}-independent if the sequence of subalgebras \mathcal{A}_i generated by \{X_1(i),X_2(i),\dots, X_n(i)\} is \mathcal{S}-independent.

Example

Example 2.15

UVAT

For two subalgebras \mathcal{S}-independence reads as follows. A pair of subalgebras (\mathcal{A}_1, \mathcal{A}_2) of \mathcal{A} is \mathcal{S}-independent if the following condition holds. Given elements X_1,X_2,\dots,X_n\in \mathcal{A}_1\cup \mathcal{A}_2, let \rho=(B_1,B_2) be an ordered set partition of the index set [n] such that X_i\in \mathcal{A}_1 for i\in B_1 and X_i\in \mathcal{A}_2 for i\in B_2. Then for any ordered set partition \pi\in \mathcal{OP}_n we have

\varphi_\pi(X_1,X_2,\dots, X_n) = \varphi_{\pi\rfloor_{B_1}\pi\rfloor_{B_2}}(X_1,X_2,\dots, X_n).
Remark

Remark 2.16

NVJC

Let \mathcal{E}=(\mathcal{U},\tilde{\varphi},(\iota^{(i)})_{i\ge1}) be an exchangeability system for a given \mathcal{B}-ncps (\mathcal{A},\varphi). It can be regarded as a spreadability system and we denote it by \mathcal{S}. Then, obviously, a sequence of subalgebras (\mathcal{A}_i)_{i\in I} of \mathcal{A} is \mathcal{E}-independent in the sense of Definition 2.12 if and only if it is \mathcal{S}-independent in the sense of Definition 2.14.

Remark

Remark 2.17

YY08
  1. With the notation introduced in Definition 3.3 below, equation (?) may be rewritten as

    \varphi_\pi(X_1,X_2,\dots,X_n) = \varphi_{\pi}(X_1^{(i_1)},X_2^{(i_2)},\dots,X_n^{(i_n)}).

    This condition means that the random vectors (X_k)_{k\in [n]} and (X_k^{(i_k)})_{k\in [n]} have the same “distribution” with respect to (\varphi_\pi)_\pi. This is compatible with the concept of spreadability, which is to regard the sequence (\iota^{(j)}(\mathcal{A}))_{j \in I} as independent copies \mathcal{A} constructed in (\mathcal{U}, \tilde{\varphi}).

  2. An alternative natural definition of \mathcal{S}-independence would be to require the simplified condition that for any (i_1,i_2,\dots,i_n)\in I^n and any X_j\in \mathcal{A}_{i_j}, j \in [n],

    \varphi (X_1 X_2 \cdots X_n) = \tilde\varphi(X_1^{(i_1)} X_2^{(i_2)} \cdots X_n^{(i_n)}),

    i.e., requiring (?) only for \pi =\hat1_n. In fact, for the spreadability systems associated with a large class of universal products, (?) implies the \mathcal{S}-independence, see Proposition 2.18. In general however, condition (?) does not imply \mathcal{S}-independence, see Example 2.19.

  3. In the case of exchangeability, \mathcal{E}-independence is equivalent to the vanishing of mixed cumulants (see Proposition 5.1).

Proposition

Proposition 2.18

MYAQ

Let d \in \mathbb{N} and \operatorname*{\odot} be a universal product in the category \operatorname{AlgP}_d satisfying ?, ? and ?. Let (\mathcal{A}, \varphi^1,\varphi^2, \dots, \varphi^d) be an object in \operatorname{AlgP}_d and (\mathcal{A}_i)_{i \in I} be a family of subalgebras of \mathcal{A} with I\subseteq \mathbb{N}. Then (\mathcal{A}_i)_{i \in I} is \odot-independent if and only if it is \mathcal{S}_{\operatorname*{\odot}}^j-independent for all j\in[d], where \mathcal{S}_{\operatorname*{\odot}}^j is the spreadability system for the \mathbb{C}-ncps (\mathcal{A}, \varphi^j) constructed in Section ?.

Uses Definition 2.2 · Used by Proof, Proof
Proof of Proposition 2.18

Suppose that (\mathcal{A}_i)_{i \in I} is \odot-independent. This means that the identity

\varphi_\pi^j(X_1,X_2,\dots,X_n) = \varphi_{\pi\curlywedge \kappa(i_1,i_2,\dots,i_n)}^j(X_1,X_2,\dots,X_n)

holds for \pi = \hat 1_n, any j \in[d], any tuple (i_1,i_2,\dots,i_n)\in I^n and any tuple of random variables (X_1,X_2,\dots,X_n) with X_k\in \mathcal{A}_{i_k}, cf. (?). The goal is to verify (Proof) for any ordered set partition \pi =(P_1,P_2,\dots, P_p)\in\mathcal{OP}_n.

Let us denote \kappa(i_1,i_2,\dots,i_n) = (Q_1,Q_2,\dots, Q_q). Then \pi\curlywedge \kappa(i_1,i_2,\dots,i_n) = (P_1 \cap Q_1, P_1 \cap Q_2, \dots, P_1 \cap Q_q, P_2\cap Q_1, \dots, P_p \cap Q_q). The RHS of (Proof) is exactly the value \varphi^j(X_1X_2 \cdots X_n) when

(\{X_i\}_{i\in P_1\cap Q_1}, \{X_i\}_{i\in P_1 \cap Q_2}, \dots, \{X_i\}_{i\in P_1 \cap Q_q}, \{X_i\}_{i\in P_2 \cap Q_1}, \{X_i\}_{i\in P_2 \cap Q_2}, \dots)

is assumed to be \odot-independent. The associativity of \odot allows us to compute the RHS of (Proof) in the following two steps:

  1. First compute \varphi^j(X_1X_2 \cdots X_n) assuming that (\{X_i\}_{i\in P_1}, \{X_i\}_{i\in P_2}, \dots, \{X_i\}_{i\in P_p}) is \odot-independent. The result is exactly the RHS of (?).

  2. Then compute each factor \varphi^{f(S)}(X_{S}) by additionally assuming that

    (\{X_i\}_{i\in S\cap Q_1}, \{X_i\}_{i\in S\cap Q_2}, \dots, \{X_i\}_{i\in S\cap Q_q})

    is \odot-independent.

This is exactly how the LHS of (Proof) is computed, so that (Proof) holds as desired.

Conversely, suppose that (\mathcal{A}_i)_{i \in I} is \mathcal{S}_{\operatorname*{\odot}}^j-independent for all j\in[d]. Formula (Proof) for \pi = \hat1_n is exactly the desired independence relation (?).

Example

Example 2.19

NNP9

Let (\mathcal{A}, \varphi) be a \mathbb{C}-ncps and let \mathcal{U}:= \sqcup_{i=1}^\infty \mathcal{A} be the coproduct in the category of algebras (see (?)) and \iota^{(i)} \colon \mathcal{A} \to \mathcal{U} be the natural embedding as i^{\rm th} component. We define a linear functional \tilde{\varphi} on \mathcal{U} by, for each X_1, X_2 , \dots, X_n \in \mathcal{A} and i_1,i_2,\dots, i_n \in \mathbb{N} with i_j \ne i_{j+1} for all j =1,2,\dots, n-1, setting

\tilde{\varphi}(X_1^{(i_1)} X_2^{(i_2)} \cdots X_n^{(i_n)}) = \begin{cases} \varphi(X_1) & \text{~if $n=1$}, \\ \varphi(X_1) \varphi(X_2) & \text{~if $n =2$}, \\ 0& \text{~if $n\ge3$}. \end{cases}

Then we get an exchangeability system (\mathcal{U}, \tilde{\varphi}, (\iota^{(i)})_{i \in \mathbb{N}}) for (\mathcal{A}, \varphi) with calculation rule.

For \pi \in \mathcal{OP}_n and X_1,X_2,\dots, X_n \in \mathcal{A} it is easy to see that

\varphi_\pi(X_1,X_2,\dots,X_n) = \begin{cases} \varphi_{(\pi)}(X_1,X_2,\dots,X_n) & \text{~if~$\pi \in \mathcal{I}_n$ with $|\pi|\le2$,} \\ 0 & \text{~otherwise}. \end{cases}

Suppose that subalgebras \mathcal{B}_1, \mathcal{B}_2 \subseteq \mathcal{A} are \mathcal{E}-independent. Then for X \in \mathcal{B}_1 and Y \in \mathcal{B}_2 we have

\varphi_{\pi}(X, Y) = \varphi_{\pi \wedge \bar\kappa(1,2)}(X,Y)

and the case \pi =\hat1_2 yields \varphi(X Y) = \varphi(X) \varphi(Y). Moreover, independence implies

\varphi_{\pi}(X, Y, X) = \varphi_{\pi \wedge \bar\kappa(1,2,1)}(X,Y,X).

On the other hand, we infer from (?) that for \pi = \{\{1,2\}, \{3\}\} the left hand side equals \varphi(X Y) \varphi(X) =\varphi(X)^2\varphi(Y), while the RHS equals zero. This is only possible if either \varphi \rfloor_{\mathcal{B}_1}=0 or \varphi\rfloor_{\mathcal{B}_2}=0 holds, so that only the trivial examples satisfy \mathcal{E}-independence. However, nontrivial subalgebras \{\mathcal{C}_1, \mathcal{C}_2\} satisfying condition (?) exist because (?) simply reads: if X_j \in \mathcal{C}_{i_j} then for \pi := \bar \kappa (i_1,i_2,\dots, i_n)

\varphi(X_1X_2\cdots X_n) = \begin{cases} \varphi_{(\pi)}(X_1,X_2,\dots,X_n) & \text{~if~$\pi \in \mathcal{I}_n$ with $|\pi|\le2$,} \\ 0 & \text{~otherwise}. \end{cases}

For example we can start from any \mathbb{C}-ncps (\mathcal{C}_i,\varphi_i), i=1,2 and then construct \varphi on \mathcal{A} := \mathcal{C}_1 \sqcup \mathcal{C}_2 so that (?) holds.

Definition

Definition 3.1

OV4D

Let \mathcal{E}=(\mathcal{U},\tilde{\varphi},(\iota^{(i)})_{i\ge1}) be an exchangeability system for a \mathcal{B}-ncps (\mathcal{A},\varphi). Cumulants are functionals satisfying the the following requirements.

  1. Multilinearity: Cumulants are multilinear functionals K_\pi\colon \mathcal{A}^n\to\mathcal{B}, indexed by set partitions \pi\in\mathcal{P}_n, n\in\mathbb{N}.

  2. Universality: There are universal coefficients c(\sigma,\pi) \in \mathbb{C} such that

    \varphi_\pi = K_\pi + \sum_{\substack{\sigma\in\mathcal{P}_n\\ \sigma<\pi}} c(\sigma,\pi) K_\sigma.
  3. Vanishing of mixed cumulants: given a family (X_1,X_2,\dots,X_n)\in\mathcal{A}^n and a partition \pi\in\mathcal{P}_n, such that some block of P\in\pi can be partitioned P=P' \dot\cup P'' nontrivially in such a way that \{X_i\mid i\in P'\} and \{X_i\mid i\in P''\} are \mathcal{E}-independent, then

    K_\pi(X_1,X_2,\dots,X_n)=0 .

    Equivalently, if \rho\in\mathcal{P}_n partitions the family (X_1,X_2,\dots,X_n) into mutually \mathcal{E}-independent subfamilies, then K_\pi(X_1,X_2,\dots,X_n)=0 unless \pi\leq\rho.

Remark

Remark 3.2

W0RH
  1. Additivity. As a consequence of ? and ? cumulants of sums of independent random variables are additive: Denote by K_n the cumulant functional K_{\hat{1}_n}, then

    K_n(X_1+Y_1,X_2+Y_2,\dots,X_n+Y_n) = K_n(X_1,X_2,\dots,X_n) +K_n(Y_1,Y_2,\dots,Y_n) .
  2. The system of equations (?) is in triangular form and can be solved recursively and transformed into the equivalent system

    K_\pi = \varphi_\pi + \sum_{\substack{\sigma\in\mathcal{P}_n\\ \sigma<\pi}} \tilde{c}(\sigma,\pi) \varphi_\sigma

    where the matrix \tilde{c}(\sigma,\pi) is inverse to c(\sigma,\pi).

Definition

Definition 3.3

5TZV

Let (\mathcal{U}, \tilde{\varphi},(\iota^{(i)})_{i \geq 1}) be a spreadability system for a \mathcal{B}-ncps (\mathcal{A}, \varphi).

  1. Given a noncommutative random variable X \in \mathcal{A} and a finite subset A \subseteq \mathbb{N} we define

    \delta_A(X)=\sum_{i\in A} X^{(i)}

    i.e., the sum of i.i.d. copies of X. In the case A=[N] we will also write \delta_N(X) and frequently abbreviate it using Rota’s dot operation

    N.X:=X^{(1)} + X^{(2)} + \cdots + X^{(N)}

    whenever it is convenient.

  2. We extend the partitioned functionals \varphi_\pi to \bigcup \iota^{(i)}(\mathcal{A}) by setting

    \varphi_{\pi}(X_1^{(i_1)}, X_2^{(i_2)}, \ldots,X_n^{(i_n)}):=\varphi_{\pi \curlywedge \kappa(i_1,i_2,\ldots,i_n)}(X_1,X_2,\ldots,X_n),

    and

    \varphi_\pi(N_1.X_1,N_2.X_2,\ldots,N_n.X_n):= \sum_{i_1=1}^{N_1} \sum_{i_2=1}^{N_2} \dots \sum_{i_n=1}^{N_n} \varphi_{\pi}(X_1^{(i_1)}, X_2^{(i_2)}, \ldots,X_n^{(i_n)}).
Remark

Remark 3.4

TGNQ

Note that (?) is actually an abuse of notation, because \varphi_\pi is defined for elements of \mathcal{A} only (see (?)); here we pretend that \mathcal{S}=(\mathcal{U}, (\iota^{(i)})_{i=1}^\infty, \tilde{\varphi}) can be interpreted as a spreadability system for the algebra \mathcal{A}^{(1,2,\dots,N)} generated by the images \mathcal{A}^{(1)},\mathcal{A}^{(2)},\dots,\mathcal{A}^{(N)}. This is true in the case of universal product construction in (?) but needs justification otherwise; yet (?) is well-defined and convenient to keep notation manageable.

Definition

Definition 3.5

Q3AY

Let (\mathcal{A},\varphi) be a \mathcal{B}-ncps and \mathcal{S}=(\mathcal{U},\tilde{\varphi},(\iota^{(i)})_{i\ge1}) a spreadability system for (\mathcal{A},\varphi). Cumulants are multilinear functionals K_\pi indexed by ordered set partitions \pi\in\mathcal{OP}_n, n\in\mathbb{N}, which satisfy the following axioms.

  1. Multilinearity: Cumulants are multilinear functionals K_\pi\colon \mathcal{A}^n\to\mathcal{B}, indexed by ordered set partitions \pi\in\mathcal{OP}_n, n\in\mathbb{N}.

  2. Universality: There are universal coefficients c(\sigma,\pi) \in \mathbb{C} such that for every \pi\in\mathcal{OP}_n

    \varphi_\pi = K_\pi + \sum_{\substack{\sigma\in\mathcal{OP}_n\\ \sigma<\pi}} c(\sigma,\pi) K_\sigma

    or, equivalently, there are universal coefficients \tilde{c}(\sigma,\pi) \in \mathbb{C} such that

    K_\pi = \varphi_\pi + \sum_{\substack{\sigma\in\mathcal{OP}_n\\ \sigma<\pi}} \tilde{c}(\sigma,\pi) \varphi_\sigma.
  3. Extensivity: given a family (X_1,X_2,\dots,X_n)\in\mathcal{A}^n and an ordered partition \pi\in\mathcal{OP}_n,

    K_{\pi}(N.X_1,N.X_2,\ldots,N.X_n) =N^{\left\lvert \pi \right\rvert}K_{\pi}(X_1,X_2,\ldots,X_n),

    where the extension of K_\pi to arguments from \bigcup \iota^{(i)}(\mathcal{A}) is defined via formula (?) and (?) as

    \begin{aligned} K_\pi(X_1^{(i_1)},X_2^{(i_2)},\ldots,X_n^{(i_n)}) &:= \sum_{\substack{\sigma\in\mathcal{OP}_n\\\sigma \leq \pi}} \tilde{c}(\sigma,\pi)\,\varphi_{\sigma}(X_1^{(i_1)},X_2^{(i_2)},\ldots,X_n^{(i_n)}) \\ &= \sum_{\substack{\sigma\in\mathcal{OP}_n\\\sigma \leq \pi}} \tilde{c}(\sigma,\pi)\,\varphi_{\sigma\curlywedge \kappa(i_1,i_2,\ldots,i_n)}(X_1,X_2,\ldots,X_n) \intertext{and thus } K_\pi(N_1.X_1,N_2.X_2,\ldots,N_n.X_n)&:= \sum_{\substack{\sigma\in\mathcal{OP}_n\\\sigma \leq \pi}} \tilde{c}(\sigma,\pi)\,\varphi_{\sigma}(N_1.X_1,N_2.X_2,\ldots,N_n.X_n), \end{aligned}

    where the diagonal coefficients are \tilde{c}(\pi,\pi)=1.

Theorem

Theorem 3.6

68ZE

Cumulants are uniquely determined by axioms ?–? from Definition 3.5.

Proof of Theorem 3.6

Assume that there are two sets of cumulants K_\pi and K'_\pi satisfying axioms ?–? with universal coefficients c(\sigma,\pi) and c'(\sigma,\pi), respectively. Then the axioms imply that

\begin{gathered} \varphi_{\pi}(N.X_1,N.X_2,\ldots,N.X_n) \\ \begin{aligned}[t] &= N^{\left\lvert \pi \right\rvert}K_{\pi}(X_1,X_2,\ldots,X_n) + \sum_{\sigma < \pi} N^{\left\lvert \sigma \right\rvert}K_{\sigma}(X_1,X_2,\ldots,X_n)\,c(\sigma,\pi) \\ &= N^{\left\lvert \pi \right\rvert}K'_{\pi}(X_1,X_2,\ldots,X_n) + \sum_{\sigma < \pi} N^{\left\lvert \sigma \right\rvert}K'_{\sigma}(X_1,X_2,\ldots,X_n)\,c'(\sigma,\pi) \end{aligned} \end{gathered}

for any N\in\mathbb{N} and the coefficients of the leading term N^{\left\lvert \pi \right\rvert} must coincide.

Lemma

Lemma 3.7

SORR

Let \pi\in \mathcal{OP}_n be an ordered set partition. Fix a block P of \pi and a number m\in\mathbb{N} and let (i_1,i_2,\dots,i_n), (i'_1,i'_2,\dots,i'_n) be n-tuples such that

i'_k= \begin{cases} i_k + m & \text{if $k\in P$}\\ i_k & \text{if $k\not\in P$} \end{cases}

Then \pi\curlywedge \kappa(i_1,i_2,\dots,i_n) = \pi\curlywedge \kappa(i'_1,i'_2,\dots,i'_n).

Used by Proof
Remark

Remark 3.8

KA4F

Note that without performing the quasi-meet operation in the lemma above the kernel partitions \kappa(i_1,i_2,\dots,i_n) and \kappa(i'_1,i'_2,\dots,i'_n) may well be nontrivial permutations of each other or even \bar\kappa(i_1',i_2',\dots,i_n')\ne\bar\kappa(i_1,i_2,\dots,i_n).

Theorem

Theorem 3.9

K6GY

For any ordered set partitions \rho \le \pi \in \mathcal{OP}_n and numbers N_1,N_2,\ldots\in\mathbb{N} we have

\varphi_{\rho}(N_{\pi(1)}.X_1,N_{\pi(2)}.X_2,\ldots,N_{\pi(n)}.X_n) = \sum_{\sigma \leq \rho}\varphi_{\sigma}(X_1,X_2,\ldots,X_n)\,\gamma_{\underline{N}}(\sigma,\rho, \pi),

with \gamma_{\underline{N}}(\sigma, \rho, \pi) as in Definition A.28 and \pi(i) as in Definition A.8?. Specializing to the case \rho=\pi yields

\begin{aligned} \varphi_{\pi}(N_{\pi(1)}.X_1,N_{\pi(2)}.X_2,\ldots,N_{\pi(n)}.X_n) = \sum_{\sigma \leq \pi}\varphi_{\sigma}(X_1,X_2,\ldots,X_n)\,\beta_{\underline{N}}(\sigma,\pi) \end{aligned}

and this is a polynomial in N_1,\dots, N_{\left\lvert \pi \right\rvert} without constant term. Furthermore, specializing to the case N:=N_1 = N_2 = \cdots entails

\varphi_{\pi}(N.X_1,N.X_2,\ldots,N.X_n) = \sum_{\sigma \leq \pi}\varphi_{\sigma}(X_1,X_2,\ldots,X_n)\,\beta_N(\sigma,\pi),

which is a polynomial in N consisting of monomials of degree at least \left\lvert \pi \right\rvert.

Used by Proof
Proof of Theorem 3.9

We begin with (3.9) and then verify the general case (3.9). The proof boils down to the enumeration of the set

S_{\underline{N}}(\sigma,\pi):= \{(i_1,i_2,\ldots,i_n)\in [N_{\pi(1)}] \times [N_{\pi(2)}] \times \cdots \times [N_{\pi(n)}]: \pi \curlywedge \kappa(i_1,i_2,\ldots,i_n)=\sigma \}.

Pick an arbitrary block P_i=\{p_1,p_2,\ldots,p_k\} of \pi = (P_1,P_2,\dots). There are \binom{N_i}{\#(\sigma\rfloor_{P_i})} possible ways to choose a tuple (i_{p_1},i_{p_2}, \ldots, i_{p_k}) such that \kappa(i_{p_1},i_{p_2}, \ldots, i_{p_k}) defines the partition \sigma\rfloor_{P_i}: for each block of \sigma\rfloor_{P_i} we have to choose a distinct label from [N_i] respecting the order prescribed by the labels of the blocks of \sigma\rfloor_{P_i}. That is, we have to choose a subset of [N_i] of cardinality \#(\sigma\rfloor_{P_i}). This can be done for every block of \pi independently and thus

\# S_{\underline{N}}(\sigma,\pi)=\prod_{i =1}^{\left\lvert \pi \right\rvert} \binom{N_i}{\#(\sigma\rfloor_{P_i})}=\beta_{\underline{N}}(\sigma,\pi).

If N_i <\#(\sigma\rfloor_{P_i}) for some i, then \pi \curlywedge \kappa(i_1,i_2,\ldots,i_n) can never be equal to \sigma, and so the cardinality of S_{\underline{N}}(\sigma,\pi) is 0, in accordance with the generally adopted convention that the generalized binomial coefficient \binom{N}{k} is zero when N<k.

The general formula (3.9) requires the enumeration of the set

S_{\underline{N}}(\sigma,\rho,\pi) = \{(i_1,i_2,\ldots,i_n)\in [N_{\pi(1)}] \times [N_{\pi(2)}] \times \cdots \times [N_{\pi(n)}]: \rho \curlywedge \kappa(i_1,i_2,\ldots,i_n)=\sigma \}.

This is similar to S_{\underline{N}}(\sigma,\rho); the difference is that the blocks of \rho contained in a block P_i \in \pi are endowed with the common number N_i. This results in the product formula

\# S_{\underline{N}}(\sigma,\rho,\pi) = \prod_{i=1}^p S_{N_i}(\sigma \rfloor_{P_i},\rho \rfloor_{P_i}),

which equals \gamma_{\underline{N}}(\sigma,\rho,\pi) from Remark A.29?.

Theorem

Theorem 3.11 (Moment-cumulant formulas)

UFVK

For any \pi \in \mathcal{OP}_n, we have

\begin{aligned} K_{\pi}(X_1,X_2,\ldots,X_n) &= \sum_{\sigma \leq \pi} \varphi_{\sigma}(X_1,X_2,\ldots,X_n)\, \widetilde{\mu}(\sigma,\pi), \\ \varphi_{\pi}(X_1, X_2,\ldots, X_n) &= \sum_{\sigma \leq \pi} K_{\sigma}(X_1,X_2,\ldots,X_n)\,\widetilde{\zeta}(\sigma,\pi) \end{aligned}

with \widetilde{\mu}, \widetilde{\zeta} as in Definition A.28.

Proof of Theorem 3.11

For 0<k<N, there is no constant term in \binom{N}{k}= \frac{N(N-1)\cdots (N-k+1)}{k!} (regarded as a polynomial in N), and the coefficient of its linear term is \frac{(-1)(-2)\cdots(-k+1)}{k!} = \frac{(-1)^{k-1}}{k}. Therefore the coefficient of N_1 \cdots N_{\left\lvert \pi \right\rvert} in \beta_{\underline{N}}(\sigma,\pi)=\prod_{i=1}^{\left\lvert \pi \right\rvert}\binom{N_i}{\#(\sigma\rfloor_{P_i})} is equal to \frac{(-1)^{\left\lvert \pi \right\rvert-\left\lvert \sigma \right\rvert}}{[\sigma:\pi]}. Note that \sum_{P\in\pi}\#(\sigma\rfloor_P)=\left\lvert \sigma \right\rvert. Comparing with (A.28) the claimed formula (3.11) follows. The same argument holds true if we look at the monomial N^{\left\lvert \pi \right\rvert} when N_1 =N_2= \cdots =N.

The inverse formula (3.11) expressing moments in terms of cumulants is an immediate consequence of the fact that the modified Möbius function \widetilde{\mu} is the inverse of the modified zeta function \widetilde{\zeta} (Corollary A.32).

Definition

Definition 3.13

GZA3

For \pi\in\mathcal{OP}_n and M_i, N_i \in \mathbb{N}, i=1,2,\dots,n, we define

\begin{gathered} \varphi_\pi(M_1.(N_1.X_1),M_2.(N_2.X_2), \ldots,M_n.(N_n.X_n)) \\ := \sum_{(i_1,i_2,\ldots,i_n) \in[M_1] \times [M_2] \times \cdots \times [M_n]} \varphi_{\pi\curlywedge \kappa(i_1,i_2,\ldots,i_n)}(N_1.X_1, N_2.X_2,\ldots,N_n.X_n). \end{gathered}
Remark

Remark 3.14

I9T0

We would like to alert the reader that the formal definition of the expectation

\varphi_\pi(M_1.(N_1.X_1),M_2.(N_2.X_2),\ldots,M_n.(N_n.X_n))

is not necessarily related to “M.(N.X)” (which formally would represent the element (N.X)^{(1)}+(N.X)^{(2)}+\cdots +(N.X)^{(M)} , i.e., “the sum of i.i.d. copies of N.X”) as an element of an enlarged space, obtained, e.g., by iterating a product construction. Yet we will show below that the vector (M_{\pi(1)}N_{\pi(1)}.X_1,\ldots, M_{\pi(n)}N_{\pi(n)}.X_n) and the virtual vector “(M_{\pi(1)}.(N_{\pi(1)}.X_1), \ldots,M_{\pi(n)}.(N_{\pi(n)}.X_n))” have the same distribution with respect to \varphi_\pi. This property implies extensivity of cumulants, see Proposition 3.16 below. It is a consequence of the associativity of the corresponding universal product in the case of classical, free, monotone and Boolean independence [HS11b]. For general spreadability systems it holds on a formal level (Lemma 3.15), even when it comes from a nonassociative universal product, like the example of V-monotone independence from Section ?.

Lemma

Lemma 3.15

9EU2

Let \pi \in \mathcal{OP}_n, X_i \in \mathcal{A} and M_i,N_i \in \mathbb{N}, i=1,2,\dots, \left\lvert \pi \right\rvert.

  1. The value \varphi_\pi(M_{\pi(1)}.(N_{\pi(1)}.X_1),\ldots,M_{\pi(n)}.(N_{\pi(n)}.X_n)) is a polynomial in the variables M_i,N_i, i=1,2,\dots, \left\lvert \pi \right\rvert, and K_\pi(N_{\pi(1)}.X_1,\ldots,N_{\pi(n)}.X_n) is the coefficient of M_1 \cdots M_{\left\lvert \pi \right\rvert}.

  2. The dot operation gives rise to an action of the multiplicative semigroup \mathbb{N}^\infty

    \begin{gathered} \varphi_\pi(M_{\pi(1)}N_{\pi(1)}.X_1,\ldots,M_{\pi(n)}N_{\pi(n)}.X_n) \\ = \varphi_\pi(M_{\pi(1)}.(N_{\pi(1)}.X_1),\ldots,M_{\pi(n)}.(N_{\pi(n)}.X_n)). \end{gathered}
Uses Definition 3.3 · Used by Proof
Proof of Lemma 3.15

?   Let \underline{M}:=(M_1,\dots, M_{\left\lvert \pi \right\rvert},0,0,\dots)\in \mathbb{N}^\infty and similarly \underline{N} \in \mathbb{N}^\infty. Then the following expansion holds:

\begin{gathered} \varphi_\pi(M_{\pi(1)}.(N_{\pi(1)}.X_1),\ldots,M_{\pi(n)}.(N_{\pi(n)}.X_n))\\ \begin{aligned} &= \sum_{(i_1,\dots, i_n) \in [M_{\pi(1)}]\times \cdots \times [M_{\pi(n)}]} \varphi_{\pi \curlywedge \kappa(i_1,i_2,\ldots,i_n)}(N_{\pi(1)}.X_1,\ldots,N_{\pi(n)}.X_n) \\ &=\sum_{\sigma \leq \pi} \sum_{\substack{(i_1,\dots, i_n) \in [M_{\pi(1)}]\times \cdots \times [M_{\pi(n)}],\\ \pi \curlywedge \kappa(i_1,i_2,\ldots,i_n)=\sigma}} \varphi_{\sigma}(N_{\pi(1)}.X_1,\ldots,N_{\pi(n)}.X_n) \\ &= \sum_{\sigma \leq \pi} \varphi_{\sigma}(N_{\pi(1)}.X_1,\ldots,N_{\pi(n)}.X_n) \,\beta_{\underline{M}}(\sigma,\pi), \end{aligned} \end{gathered}

where we used the identity (Proof). Hence \varphi_\pi(M_{\pi(1)}.(N_{\pi(1)}.X_1),\ldots,M_{\pi(n)}.(N_{\pi(n)}.X_n)) is a polynomial in M_i,N_i, i=1,2,\dots, \left\lvert \pi \right\rvert, and from the proof of Theorem 3.11 we infer that the coefficient of M_1\cdots M_{\left\lvert \pi \right\rvert} is equal to K_\pi(N_{\pi(1)}.X_1,\ldots,N_{\pi(n)}.X_n).

?   We proceed with the computation of (Proof):

\begin{gathered} \varphi_\pi(M_{\pi(1)}.(N_{\pi(1)}.X_1),M_{\pi(2)}.(N_{\pi(2)}.X_2),\ldots,M_{\pi(n)}.(N_{\pi(n)}.X_n))\\ \begin{aligned} &= \sum_{\rho \leq \pi} \varphi_{\rho}(N_{\pi(1)}.X_1,N_{\pi(2)}.X_2,\ldots,N_{\pi(n)}.X_n) \,\beta_{\underline{M}}(\rho,\pi) \\ &= \sum_{\rho \leq \pi} \sum_{\sigma \le \rho } \varphi_{\sigma}(X_1,X_2, \ldots,X_n) \, \gamma_{\underline{N}}(\sigma,\rho,\pi) \, \beta_{\underline{M}}(\rho,\pi)\\ &= \sum_{\sigma\leq \pi} \varphi_{\sigma}(X_1, X_2,\ldots,X_n) \biggl( \sum_{\substack{\rho\in\mathcal{OP}_n \\ \sigma \leq \rho \leq \pi}} \gamma_{\underline{N}}(\sigma,\rho,\pi) \, \beta_{\underline{M}}(\rho,\pi) \biggr)\\ &= \sum_{\sigma \leq \pi} \varphi_\sigma(X_1, X_2,\ldots,X_n) (\gamma_{\underline{N}} \otriangle \beta_{\underline{M}}) (\sigma,\pi)\\ &= \sum_{ \sigma \leq \pi} \varphi_\sigma(X_1, X_2,\ldots,X_n) \,\beta_{\underline{M} \circ \underline{N}}(\sigma,\pi) \\ &= \varphi_{\pi}(M_{\pi(1)}N_{\pi(1)}.X_1,M_{\pi(2)}N_{\pi(2)}.X_2,\ldots,M_{\pi(n)}N_{\pi(n)}.X_n), \end{aligned} \end{gathered}

where (3.9) was used in the second equality and Corollary A.32 was used in the next to last line.

Proposition

Proposition 3.16

D5U4

The cumulants from Definition 3.10 satisfy extensivity (?) and more generally,

K_{\pi}(N_{\pi(1)}.X_1,N_{\pi(2)}.X_2,\ldots,N_{\pi(n)}.X_n) =N_1N_2 \cdots N_{\left\lvert \pi \right\rvert}K_{\pi}(X_1,X_2,\ldots,X_n).
Proof of Proposition 3.16

By definition, \varphi_\pi applied to the same arguments has the expansion

\begin{gathered} \varphi_{\pi}(M_{\pi(1)}N_{\pi(1)}.X_1, M_{\pi(2)}N_{\pi(2)}.X_2,\ldots,M_{\pi(n)}N_{\pi(n)}.X_n) \\ =\Biggl( \prod_{i=1}^{\left\lvert \pi \right\rvert}M_i N_i \Biggr) K_{\pi}(X_1,X_2,\ldots,X_n) + (\text{sum of monomials in $M_i N_i$ of higher degrees}). \end{gathered}

On the other hand, from Lemma 3.15? we infer

\begin{gathered} \varphi_{\pi}(M_{\pi(1)}.(N_{\pi(1)}.X_1), M_{\pi(2)}.(N_{\pi(2)}.X_2),\ldots,M_{\pi(n)}.(N_{\pi(n)}.X_n)) \\ = M_1 M_2 \cdots M_{\left\lvert \pi \right\rvert} K_{\pi}(N.X_1,N.X_2,\ldots,N.X_n) \\ + (\text{sum of monomials in $M_i$ of higher degree}). \end{gathered}

We have thus computed the expectation in two ways and it follows from Lemma 3.15? that the coefficients of M_1 \cdots M_{\left\lvert \pi \right\rvert} coincide.

Proposition

Proposition 3.17

A99M

If a spreadability system \mathcal{S} satisfies exchangeability, then both the partitioned expectations and cumulants are invariant under permutations of the blocks. More precisely, for any partition \pi=(P_1,P_2, \dots,P_k) \in \mathcal{OP}_n and permutation h \in \mathfrak{S}_k, \varphi_\pi = \varphi_{h(\pi)} and K_\pi = K_{h(\pi)}, where h(\pi)=(P_{h(1)}, P_{h(2)}, \dots,P_{h(k)}), Thus \varphi_\pi and K_\pi are completely determined by the underlying (unordered) set partition \bar{\pi}.

Proof of Proposition 3.17

Invariance of \varphi_\pi under permutations of the blocks of \pi is an immediate consequence of exchangeability and this invariance extends to the extension (?). Indeed let \pi=(P_1,P_2,\dots,P_k)\in\mathcal{OP}_n and i_1,i_2,\dots,i_n \in \mathbb{N}, then for any h \in \mathfrak{S}_k there exists a permutation g \in \mathfrak{S}_\infty such that h(\pi)\curlywedge \kappa(i_1,i_2,\dots,i_n) = g(\pi\curlywedge \kappa(i_1,i_2,\dots,i_n) ). Therefore,

\begin{aligned} \varphi_{h(\pi)}(X_1^{(i_1)}, X_2^{(i_2)},\dots, X_n^{(i_n)}) &= \varphi_{h(\pi)\curlywedge \kappa(i_1,i_2,\dots,i_n)}(X_1,X_2,\dots, X_n) \\ &= \varphi_{g(\pi\curlywedge \kappa(i_1,i_2,\dots,i_n))}(X_1,X_2,\dots, X_n) \\ &= \varphi_{\pi\curlywedge \kappa(i_1,i_2,\dots,i_n)}(X_1,X_2,\dots,X_n) \\ &=\varphi_{\pi}(X_1^{(i_1)}, X_2^{(i_2)},\dots, X_n^{(i_n)}). \end{aligned}

Consequently also \varphi_\pi(N.X_1,N.X_2,\dots,N.X_n) = \varphi_{h(\pi)}(N.X_1,N.X_2,\dots,N.X_n) and, by definition, K_\pi(X_1,X_2,\dots, X_n) = K_{h(\pi)}(X_1,X_2,\dots, X_n).

Corollary

Corollary 3.18

HIJI

If a spreadability system \mathcal{S} satisfies exchangeability then for \pi \in \mathcal{P}_n we simply write K_\pi for the partitioned cumulant functional uniquely determined in Proposition 3.17. Then for any \pi \in \mathcal{P}_n, we have

\begin{aligned} K_{\pi}(X_1,X_2,\ldots,X_n) &= \sum_{\substack{\sigma \in \mathcal{P}_n \\ \sigma \leq \pi}} \varphi_{\sigma}(X_1,X_2,\ldots,X_n)\, \mu_{\mathcal{P}}(\sigma,\pi), \\ \varphi_{\pi}(X_1, X_2,\ldots, X_n) &= \sum_{\substack{\sigma \in \mathcal{P}_n \\ \sigma \leq \pi}} K_{\sigma}(X_1,X_2,\ldots,X_n), \end{aligned}

where \mu_{\mathcal{P}} is the Möbius function on the posets \mathcal{P}_n, n\in \mathbb{N}, see (?).

Example

Example 3.19 (Cumulants in terms of moments)

5AQD

We will write the ordered kernel set partition \kappa(i_1,\dots, i_n) simply as the multiset permutation i_1i_2\dots i_n. For example, 211=(\{2,3\},\{1\}). Examples of Theorem 3.11 are given by

\begin{aligned} K_1(X) &= \varphi_1(X), \\ K_{11}(X, Y) &= \varphi_{11}(X,Y) - \frac{1}{2}(\varphi_{12}(X,Y) + \varphi_{21}(X,Y)), \\ K_{12}(X,Y) &= \varphi_{12}(X,Y), \\ K_{21}(X,Y) &= \varphi_{21}(X,Y), \\ K_{111} &= \varphi_{111} - \frac{1}{2}(\varphi_{112}+\varphi_{121}+\varphi_{122}+\varphi_{211}+\varphi_{212}+\varphi_{221}) \\ &\qquad+\frac{1}{3}(\varphi_{123}+\varphi_{132}+\varphi_{213}+\varphi_{231}+\varphi_{312}+\varphi_{321}), \\ K_{112} &= \varphi_{112} - \frac{1}{2}(\varphi_{123}+\varphi_{213}), \end{aligned}

where for the sake of compactness the arguments (X,Y,Z) are omitted in the last two formulas.

Example

Example 3.20 (Moments in terms of cumulants)

U8GI

Examples of Theorem 3.11 are given by

\begin{aligned} \varphi_1(X) &= K_1(X), \\ \varphi_{11}(X, Y) &= K_{11}(X,Y) + \frac{1}{2!}(K_{12}(X,Y) + K_{21}(X,Y)), \\ \varphi_{12}(X,Y) &= K_{12}(X,Y), \\ \varphi_{21}(X,Y) &= K_{21}(X,Y), \\ \varphi_{111} &= K_{111} + \frac{1}{2!}(K_{112}+K_{121}+K_{122}+K_{211}+K_{212}+K_{221}) \\ &\qquad+\frac{1}{3!}(K_{123}+K_{132}+K_{213}+K_{231}+K_{312}+K_{321}), \\ \varphi_{112} &= K_{112} + \frac{1}{2!}(K_{123}+K_{213}), \end{aligned}

where (X,Y,Z) are omitted in the last two formulas for simplicity.

Definition

Definition 3.21

1Y97

For n\in\mathbb{N} the n-linear maps on \mathcal{A}

K_n:=K_{\hat{1}_n}, \qquad \varphi_n:=\varphi_{\hat{1}_n}

are called the n^{\mathrm{th}} cumulant and expectation functional, respectively. The latter is simply

\varphi_n(X_1, X_2, \dots, X_n) =\varphi(X_1X_2 \cdots X_n)

for X_i \in \mathcal{A}. Following notation (?), we also write

\varphi_n(X_1^{(i_1)}, X_2^{(i_2)}, \dots, X_n^{(i_n)}) =\tilde{\varphi}(X_1^{(i_1)} X_2^{(i_2)} \cdots X_n^{(i_n)}) .

We extend these functionals multiplicatively to set partitions and ordered set partitions as follows.

For set partitions \pi\in\mathcal{P}_n we define multiplicative extensions \varphi_{(\pi)} of (\varphi_n)_{n\in\mathbb{N}} and K_{(\pi)} of (K_n)_{n\in\mathbb{N}} by setting

\begin{aligned} &\varphi_{(\pi)}(Y_1,Y_2,\ldots, Y_n) :=\prod_{P\in \pi} \varphi_{\left\lvert P \right\rvert}(Y_P) \qquad \text{and} \qquad K_{(\pi)}(Y_1,Y_2,\ldots, Y_n) :=\prod_{P\in \pi} K_{\left\lvert P \right\rvert}(Y_P)\\ &\qquad\qquad \text{for} \quad Y_1,Y_2,\dots, Y_n\in \bigcup_{i=1}^\infty \iota^{(i)}(\mathcal{A}) \quad \text{or} \quad Y_1, Y_2,\dots, Y_n \in \mathcal{A}, \end{aligned}

where we use notation (?) for multilinear functionals. We also define \varphi_{(\pi)}:=\varphi_{(\bar{\pi})} and K_{(\pi)}:=K_{(\bar{\pi})} for ordered set partitions \pi\in\mathcal{OP}_n.

Lemma

Lemma 3.22

OVO3

Let \pi\in\mathcal{OP}_n be an ordered set partition.

  1. Let \mathcal{S}=(\mathcal{U}, \tilde \varphi, (\iota^{(i)})_{i\ge1}) be a spreadability system for some \mathcal{B}-ncps (\mathcal{A},\varphi). Assume that there exists a family of complex numbers \{s(\sigma; \pi): \sigma \in \mathcal{P}_n, \sigma \le \bar\pi\} such that

    \varphi_\pi(Y_1, Y_2, \dots, Y_n) = \sum_{\substack{\sigma \in \mathcal{P}_n\\ \sigma \le \bar\pi }} s(\sigma; \pi) \varphi_{(\sigma)}(Y_1,Y_2,\dots, Y_n)

    holds for any tuple of random variables Y_1, Y_2, \dots, Y_n \in \bigcup_{i=1}^\infty \iota^{(i)}(\mathcal{A}). Then

    K_\pi(X_1,X_2,\dots,X_n)= s(\bar\pi;\pi) K_{(\pi)}(X_1,X_2,\dots,X_n)

    for any tuple X_1, X_2, \dots, X_n \in \mathcal{A}.

  2. Let (\mathcal{U}, \tilde \varphi, (\iota^{(i)})_{i\ge1}) be a spreadability system for some \mathcal{B}-ncps (\mathcal{A},\varphi). Assume that there exist coefficients \{t(\rho; \pi): \rho \in \mathcal{OP}_n, \bar \sigma < \bar \pi\} such that

    \varphi_\pi(Y_1, Y_2, \dots, Y_n) = \sum_{\substack{\rho \in \mathcal{OP}_n\\ \bar \rho < \bar \pi }} t(\rho; \pi) \varphi_{\rho}(Y_1,Y_2,\dots, Y_n)

    holds for any tuple of random variables Y_1, Y_2, \dots, Y_n \in \bigcup_{i=1}^\infty \iota^{(i)}(\mathcal{A}). Then the cumulant K_\pi(X_1,X_2,\dots, X_n)=0 for any tuple X_1, X_2, \dots, X_n \in \mathcal{A}.

Uses Definition 2.2 · Used by Proof, Proof
Proof of Lemma 3.22

? Choosing Y_k = \iota^{(i_k)}(X_k) and taking the sum of (?) over all indices i_1, i_2,\dots, i_n \in [N] we get formula (?) for Y_k = N. X_k as well. The desired conclusion is an immediate consequence by comparing the coefficients of N^{|\pi|} together with the definition of cumulants (Definition 3.10). The point is that, since the number of blocks of any \sigma<\bar \pi is strictly larger than the number of blocks of \pi, the coefficient of N^{|\pi|} vanishes in \varphi_{(\sigma)}(N.X_1,N.X_2,\dots,N.X_n). ? is proven in a similar way.

Remark

Remark 3.23

DVYW

Suppose the expansion (?) holds for all ordered set partitions \pi \in \mathcal{OP}_n, then \mathcal{S} is a spreadability system with calculation rule (?) (e.g. take Y_1,Y_2,\dots, Y_n \in \iota^{(1)}(\mathcal{A})). Now the expansion (?) extends the calculation rule to the variables Y_1,Y_2,\dots in \bigcup_{i=1}^\infty \iota^{(i)}(\mathcal{A}). Such an extension is not automatic; we give a counterexample in Example 3.27.

Proposition

Proposition 3.24

70W3
  1. Let K^{\mathrm{T}}_\pi be the cumulants associated to the tensor spreadability system \mathcal{S}_{\mathrm{T}}. Then

    K^{\mathrm{T}}_\pi=K^{\mathrm{T}}_{(\pi)},\qquad \pi \in \mathcal{OP}_n.
  2. Let K^{\mathrm{F}}_\pi be the cumulants associated to the free spreadability system \mathcal{S}_{\mathrm{F}}. Then

    K^{\mathrm{F}}_\pi= \begin{cases} 0, &\pi \notin \mathcal{ONC}_n, \\ K^{\mathrm{F}}_{(\pi)},& \pi \in \mathcal{ONC}_n \end{cases}

    where by \mathcal{ONC}_n we denote the set of ordered noncrossing partitions, see Definition A.14 ?.

  3. Let K^{\mathrm{T}}_\pi be the cumulants associated to the Boolean spreadability system \mathcal{S}_{\mathrm{B}}. Then

    K^{\mathrm{B}}_\pi= \begin{cases} 0, &\pi \notin \mathcal{OI}_n, \\ K^{\mathrm{B}}_{(\pi)},& \pi \in \mathcal{OI}_n \end{cases}

    where by \mathcal{OI}_n we denote the set of ordered interval partitions, see Definition A.14 ?.

  4. Let K^{\mathrm{M}}_\pi be the cumulants associated to the monotone spreadability system \mathcal{S}_{\mathrm{M}}. Then

    K^{\mathrm{M}}_\pi= \begin{cases} 0, &\pi \notin \mathcal{M}_n, \\ K^{\mathrm{M}}_{(\pi)},& \pi \in \mathcal{M}_n \end{cases}

    where by \mathcal{M}_n we denote the set of monotone partitions, see Definition A.14 ?.

These results combined with the general moment-cumulant formula (Theorem 3.11) reproduce the known formulas.

Proof of Proposition 3.24

Items ?, ? and ? satisfy exchangeability and are covered in [L04] (note that our cumulants are invariant under the permutation of the blocks of \pi, see Proposition 3.17); alternatively these cases follow from Lemma 3.22 and the arguments below.

We are left with case ?. Let u_{\rm M}(\sigma; \pi) be the universal coefficients from Axiom ? for monotone independence (where f \colon \sigma \to [1] is unique and hence omitted). We can prove (?) for s(\sigma; \pi) := u_{\mathrm{M}}(\sigma; \pi) using the technique from the proof of Proposition 2.18. Indeed, associativity allows us to deduce the identity

\varphi_{\pi}(X_1^{(i_1)},X_2^{(i_2)},\dots, X_n^{(i_n)})= \varphi_{\pi \curlywedge \kappa(i_1,\dots, i_n)}(X_1,X_2,\dots, X_n)

following the two steps ? and ? there and the resulting value is exactly the RHS of (?) with Y_k = X_k^{(i_k)}. It is known that u_{\rm M}(\bar{\pi}; \pi)=1 for \pi \in \mathcal{M}_n and u_{\rm M}(\bar{\pi}; \pi)=0 for \pi \in \mathcal{OP}_n \setminus \mathcal{M}_n, see [Muraki:2002:five, Prop. 3.2] and the claimed formula follows.

Remark

Remark 3.25

P4UZ

Other examples of noncrossing cumulants come from the c-free exchangeability systems of [BLS96], see [L04, Section 4.7]. Lemma 3.22 ? provides a new proof that crossing cumulants vanish.

Proposition

Proposition 3.26

O1VK

Let K^{\mathrm{CM}}_\pi be the \pi-cumulant associated to the c-monotone spreadability system \mathcal{S}_{\mathrm{CM}} defined in Section ? and K^{{\mathrm{M}},\psi}_n the n^{\mathrm{th}} monotone cumulant functional with respect to the linear map \psi:\mathcal{A}\to \mathbb{C}. Then

\displaystyle K^{\mathrm{CM}}_\pi(X_1,X_2,\ldots,X_n)= \begin{cases} \displaystyle \prod_{P \in \mathrm{Outer}(\pi)}K^{\mathrm{CM}}_{\left\lvert P \right\rvert}(X_P) \prod_{P \in \mathrm{Inner}(\pi)}K^{{\mathrm{M}},\psi}_{\left\lvert P \right\rvert}(X_P),& \pi \in \mathcal{M}_n, \\ 0, &\pi \notin \mathcal{M}_n \end{cases}

with \mathrm{Outer}(\pi) and \mathrm{Inner}(\pi) the sets of outer blocks of \pi and inner blocks of \pi, respectively; see Definition A.3 ?.

Proof of Proposition 3.26

We use two spreadability systems: the c-monotone spreadability system \mathcal{S}_{\mathrm{CM}}=(\mathcal{U}, \tilde{\varphi}, (\iota^{(i)})_{i=1}^\infty) for (\mathcal{A},\varphi) and the monotone spreadability system \mathcal{S}_{\mathrm{M}}=(\mathcal{U}, \tilde{\psi}, (\iota^{(i)})_{i=1}^\infty) for (\mathcal{A},\psi). The multiplicative extension \varphi_{(\pi)} of \varphi is now modified to

\begin{aligned} & \varphi_{(\pi), \psi}(Y_1, Y_2,\ldots, Y_n) := \prod_{P \in \mathrm{Outer}(\pi)}\varphi_{\left\lvert P \right\rvert}(Y_P) \prod_{P \in \mathrm{Inner}(\pi)}\psi_{\left\lvert P \right\rvert}(Y_P) \\ & \qquad \text{for} \quad \pi\in\mathcal{ONC}_n, \quad Y_1,Y_2,\dots, Y_n\in \bigcup_{i=1}^\infty \iota^{(i)}(\mathcal{A}) \quad \text{or} \quad Y_1, Y_2,\dots, Y_n \in \mathcal{A}, \end{aligned}

which apparently coincides with \varphi_{(\pi)} if \varphi=\psi. By the definition of c-monotone spreadability system, there exist universal constants u_{\mathrm{CM}}(\sigma, f; \pi) \in \mathbb{C} depending only on \pi \in \mathcal{OP}_n, \sigma \in\mathcal{P}_n with \bar\pi > \sigma and a “2-coloring” f of the blocks of \sigma, i.e. a function f\colon \sigma\to \{1,2\}, such that for any Y_j\in \mathcal{A}

\begin{gathered} \varphi_\pi(Y_1,Y_2,\ldots,Y_n) = \varphi_{(\pi), \psi}(Y_1,Y_2,\ldots,Y_n) \\ + \sum_{\substack{\sigma\in\mathcal{P}_n, \sigma <\bar{\pi} \\ f\colon\, \sigma\to \{1,2\} }}u_{\mathrm{CM}}(\sigma, f; \pi) \Bigl(\prod_{f(S)=1}\varphi_{\left\lvert S \right\rvert}(Y_S)\Bigr) \Bigl(\prod_{f(S)=2}\psi_{\left\lvert S \right\rvert}(Y_S)\Bigr) \end{gathered}

if \pi \in \mathcal{M}_n and

\varphi_\pi(Y_1,Y_2,\ldots,Y_n) = \sum_{\substack{\sigma\in\mathcal{P}_n, \sigma <\bar{\pi} \\ f\colon\, \sigma\to \{1,2\} }} u_{\mathrm{CM}}(\sigma, f; \pi) \Bigl(\prod_{f(S)=1}\varphi_{\left\lvert S \right\rvert}(Y_S)\Bigr) \Bigl(\prod_{f(S)=2}\psi_{\left\lvert S \right\rvert}(Y_S)\Bigr)

if \pi \notin \mathcal{M}_n. Similar to the monotone case, thanks to associativity and the arguments from the proof of Proposition 2.18, this formula holds also for Y_j \in \bigcup_{i=1}^\infty \iota^{(i)}(\mathcal{A}), j\in[n]. Then the arguments from Lemma 3.22 and Proposition 3.24 applies.

Example

Example 3.27

X2HV

Let (\mathcal{A}, \varphi) be a \mathbb{C}-ncps and let \mathcal{U}:= \sqcup_{i=1}^\infty \mathcal{A} be the coproduct in the category of algebras (see (?)) and \iota^{(i)} \colon \mathcal{A} \to \mathcal{U} be the natural embedding as i^{\mathrm{th}} component. We define a linear functional \tilde{\varphi} on \mathcal{U} by, for each X_1,X_2,\dots, X_n \in \mathcal{A} and i_1,i_2,\dots, i_n\in\mathbb{N} with i_j \ne i_{j+1} for all j =1,2,\dots, n-1, setting

\tilde{\varphi}(X_1^{(i_1)} X_2^{(i_2)} \cdots X_n^{(i_n)}) = \begin{cases} 0& \text{~if there are $k \ne \ell$ such that $i_k=i_\ell$}, \\ \prod_{i=1}^n \varphi(X_i) & \text{~otherwise}. \end{cases}

Then we get an exchangeability system (\mathcal{U}, \tilde{\varphi}, (\iota^{(i)})_{i \in \mathbb{N}}) for (\mathcal{A}, \varphi).

For \pi \in \mathcal{P}_n and X_1,X_2,\dots, X_n \in \mathcal{A} it is easy to see that

\varphi_\pi(X_1,X_2,\dots,X_n) = \begin{cases} \varphi_{(\pi)}(X_1,X_2,\dots,X_n) & \text{~if~} \pi \in \mathcal{I}_n, \\ 0 & \text{~otherwise} \end{cases}

holds for any tuple X_i \in \mathcal{A}; however this formula does not extend to mixed tuples Y_i \in \bigcup_{i=1}^\infty \iota^{(i)}(\mathcal{A}). For example, if \pi=\{\{1,3\},\{2\}\} then

\varphi_\pi(X_1^{(1)}, X_2^{(2)}, X_3^{(3)}) = \varphi(X_1) \varphi(X_2) \varphi(X_3) = \varphi(X_1^{(1)}) \varphi(X_2^{(2)}) \varphi(X_3^{(3)})

but

\varphi_\pi(X_1^{(1)}, X_2^{(1)}, X_3^{(1)}) =0 \ne \varphi(X_1^{(1)}) \varphi(X_2^{(1)}) \varphi(X_3^{(1)})

in general.

Next we shall compute some cumulants. Straightforward calculations yield

\begin{aligned} \varphi_{\hat{1}_1} (N. X_1) &= N \varphi(X_1), \\ \varphi_{\hat{1}_2}(N.X_1,N.X_2) &= N \varphi(X_1X_2) + N(N-1) \varphi(X_1) \varphi(X_2), \\ \varphi_{\hat{1}_3}(N.X_1, N.X_2, N.X_3) &= \begin{multlined}[t] N \varphi(X_1 X_2 X_3) + N(N-1) \varphi(X_1 X_2) \varphi(X_3) \\ + N(N-1) \varphi(X_1) \varphi(X_2 X_3) \\ + N(N-1) (N-2) \varphi(X_1)\varphi( X_2) \varphi(X_3), \end{multlined} \end{aligned}

so that the coefficients of N yield the cumulants

\begin{aligned} K_1 (X_1) &= \varphi(X_1), \\ K_2(X_1,X_2) &= \varphi(X_1X_2) - \varphi(X_1) \varphi(X_2), \\ K_3(X_1, X_2, X_3) &= \varphi(X_1 X_2 X_3) - \varphi(X_1 X_2) \varphi(X_3) - \varphi(X_1) \varphi(X_2 X_3) + 2 \varphi(X_1)\varphi( X_2) \varphi(X_3). \end{aligned}

On the other hand, for \pi = \{\{1,3\},\{2\}\} we have

\begin{aligned} \varphi_{\pi}(N.X_1, N.X_2, N.X_3) &= \sum_{\substack{ i_1,i_2,i_3 \\ i_1 = i_3}} \sum_{i_1 = i_3} \varphi_{\pi}(X_1^{(i_1)}, X_2^{(i_2)}, X_3^{(i_3)}) + \sum_{\substack{i_1,i_2,i_3 \\ i_1 \ne i_3}} \varphi_{\pi}(X_1^{(i_1)}, X_2^{(i_2)}, X_3^{(i_3)}) \\ &= \sum_{\substack{i_1,i_2,i_3 \\ i_1 = i_3}} \underbrace{\varphi_{\pi}(X_1, X_2, X_3)}_{\text{$=0$}} + \sum_{\substack{i_1,i_2,i_3 \\ i_1 \ne i_3}} \varphi_{\hat{0}_3}(X_1, X_2, X_3) \\ &= N^2(N-1) \varphi(X_1)\varphi(X_2) \varphi(X_3), \end{aligned}

so that

K_{\pi}(X_1, X_2, X_3) = - \varphi(X_1)\varphi(X_2) \varphi(X_3),

which does not vanish in general although the underlying set partition \pi is not interval. This does not contradict Lemma 3.22 because condition (?) is not satisfied. Indeed, the only possible identity would be \varphi_{\{13/2\}}(Y_1,Y_2,Y_3) = c \varphi_{\{1/2/3\}}(Y_1,Y_2,Y_3) for every Y_j\in\bigcup \iota^{(i)}(\mathcal{A}). However the combinations \varphi_{\{13/2\}}(X_1^{(1)},X_2^{(1)},X_3^{(1)}) = \varphi_{\{13/2\}}(X_1,X_2,X_3) = 0 and \varphi_{\{13/2\}}(X_1^{(1)},X_2^{(2)},X_3^{(3)}) = \varphi_{\{1/2/3\}}(X_1,X_2,X_3) = \varphi(X_1) \varphi(X_2) \varphi(X_3) show that there is no universal constant c satisfying the requirement.

Remark

Remark 3.28

CI9Q

We already verified in Section ? that Dacko’s V-monotone independence [Dacko:2019] satisfies the axioms of a spreadability system, but is non-associative. On the other hand, it also falls into the framework of the tree operad of Jekel and Liu [JekelLiu:2020:operad] with the corresponding cumulants and it is an interesting question how these cumulants are related.

Definition

Definition 4.1

AD81

Let \pi=(P_1,P_2,\dots, P_p) \in \mathcal{OP}_n, \underline{t}=(t_1,\dots, t_p) \in \mathbb{R}^p, j \in[p]. We define the partial cumulant to be the polynomial

K_{\pi, P_j}^{(t_1,\dots, t_{j-1},1,t_{j+1}, \dots, t_p)}(X_1,\dots,X_n) = \left.\frac{\partial }{\partial t_j}\right\rvert_{t_j=0}\varphi_{\pi}^{\underline{t}}(X_1,\dots, X_n).
Proposition

Proposition 4.2

GK2H

Let \pi=(P_1,P_2,\dots, P_p) \in \mathcal{OP}_n, \underline{t}=(t_1,\dots, t_p) \in \mathbb{R}^p, j \in [p]. Then

\begin{gathered} K_{\pi, P_j}^{(t_1,\dots, t_{j-1},1,t_{j+1}, \dots, t_p)}(X_1,\dots,X_n) \\ = \sum_{\sigma \in \mathcal{OP}_{P_j}}\varphi_{(P_1,P_2,\dots,P_{j-1},\sigma,P_{j+1},\dots,P_p)}^{(t_1,\dots, t_{j-1},1, t_{j+1}, \dots, t_p)}(X_1,\dots, X_n)\, \widetilde{\mu}(\sigma, \hat{1}_{P_j}). \end{gathered}
Theorem

Theorem 4.3

16V7

For \pi =(P_1,P_2,\dots, P_p) \in \mathcal{OP}_n, \underline{t} =(t_1,t_2,\dots, t_p) and j \in [p] we have

\begin{aligned} \frac{\partial}{\partial t_j} \varphi^{\underline{t}}_\pi(X_1,\dots, X_n) &= \sum_{\emptyset \neq A \subseteq P_j}K_{(P_1,\dots, P_{j-1}, A, P_j\setminus A, P_{j+1}, \dots, P_p),A}^{(t_1,\dots, t_{j-1}, 1, t_j, t_{j+1}, \dots, t_p)}(X_1,\dots, X_n) \\ &= \sum_{\emptyset \neq A \subseteq P_j}K_{(P_1,\dots, P_{j-1}, P_j\setminus A, A, P_{j+1}, \dots, P_p),A}^{(t_1,\dots, t_{j-1}, t_j,1, t_{j+1}, \dots, t_p)}(X_1,\dots, X_n). \end{aligned}
Proof of Theorem 4.3

The main ingredient here is the invariance principle of Lemma 3.7. Recall the delta/dot operation from Definition 3.3?, tensor notations from Section ? and recall that for any partition \pi=(P_1,P_2,\dots, P_p) \in \mathcal{OP}_n the multilinear map \varphi_\pi^{\underline{t}}\colon \mathcal{A}^n \to \mathbb{C} is identified with the linear lifting \tilde{\varphi}_\pi^{\underline{t}}\colon \mathcal{A}^{\otimes n} \to \mathbb{C}. Similar to (?), for a linear map L\colon \mathcal{A}\to\mathcal{A} we we adopt the notation L^{\otimes P}\colon \mathcal{A}^{\otimes n} \to \mathcal{B}^{\otimes n} for the linear map

L^{\otimes P}=I\otimes I\otimes\dots\otimes L\otimes I\otimes \dots\otimes L\otimes I\otimes\dots\otimes L\otimes\dots\otimes I

with L appearing exactly at position j for every j\in P. Let {\mathbf X}=X_1\otimes X_2\otimes \dots\otimes X_n, let k + [m]:=\{k+1,k+2,\dots, k+m\} and let e_j \in \mathbb{R}^p be the jth unit vector. Then, for each j \in [p],

\varphi_\pi^{\underline{N}+me_j}(X_1,\dots, X_n) = \varphi_\pi^{\underline{N}+me_j}({\mathbf X}) = \varphi_\pi\left( \left(\delta_{[N_j+m]}^{\otimes P_j} \prod_{i\ne j} \delta_{[N_i]}^{\otimes P_i}\right) ({\mathbf X})\right) ;

after expansion we obtain

\delta_{[N_j+m]}^{\otimes P_j} = (\delta_{[N_j]}+\delta_{N_j+[m]})^{\otimes P_j} = \sum_{A\subseteq P_j} \delta_{[N_j]}^{\otimes P_j \setminus A} \delta_{N_j+[m] }^{\otimes A}

and thus (Proof) is equal to

\begin{aligned} &= \sum_{A\subseteq P_j} \varphi_\pi\!\left( \left(\delta_{[N_j]}^{\otimes P_j \setminus A} \delta_{N_j+[m] }^{\otimes A} \prod_{i\ne j} \delta_{[N_i]}^{\otimes P_i}\right) ({\mathbf X}) \right). \intertext{This is a sum of $\varphi_\pi$'s with entries $X_{k}^{(i_k)}, k\in[n]$, where the indices $i_k$ with $k\in A$ are strictly larger than the indices $i_k$ with $k\in P_j \setminus A$, and therefore we may split the block $P_j \in \pi$ into two parts: } &= \sum_{A\subseteq P_j} \varphi_{(P_1,P_2,\dots,P_{j-1},P_j\setminus A, A, P_{j+1},\dots,P_p)}\left( \left( \delta_{[N_j]}^{\otimes P_j \setminus A} \delta_{N_j+[m] }^{\otimes A} \prod_{i\ne j} \delta_{[N_i]}^{\otimes P_i}\right) ({\mathbf X}) \right). \intertext{ Now by Lemma~\ref{lem:quasimeetshift} the local shift in $A$ can be omitted without changing the value and we obtain } &= \sum_{A\subseteq P_j} \varphi_{(P_1,P_2,\dots,P_{j-1},P_j\setminus A, A, P_{j+1},\dots,P_p)}\left( \left( \delta_{[N_j]}^{\otimes P_j \setminus A} \delta_{[m]}^{\otimes A} \prod_{i\ne j} \delta_{[N_i]}^{\otimes P_i}\right) ({\mathbf X}) \right) \\ &= \sum_{A\subseteq P_j} \varphi_{(P_1,P_2,\dots,P_{j-1},P_j\setminus A, A, P_{j+1},\dots,P_p)}^{(N_1,N_2,\dots,N_{j-1},N_j,m,N_{j+1},\dots,N_p)} ({\mathbf X}). \end{aligned}

The analytic extension of this identity is

\begin{aligned} \varphi_\pi^{\underline{t}+se_j}({\mathbf X}) &= \sum_{A\subseteq P_j} \varphi_{(P_1,P_2,\dots,P_{j-1},P_j\setminus A, A, P_{j+1},\dots,P_p)}^{(t_1,t_2,\dots,t_{j-1},t_j,s,t_{j+1},\dots,t_p)} ({\mathbf X})\\ &= \varphi_\pi^{\underline{t}}({\mathbf X}) + \sum_{\emptyset\ne A\subseteq P_j} \varphi_{(P_1,P_2,\dots,P_{j-1},P_j\setminus A,A, P_{j+1},\dots,P_p)}^{(t_1,t_2,\dots,t_{j-1},t_j,s,t_{j+1},\dots,t_p)} ({\mathbf X}), \end{aligned}

and the derivative satisfies the derived identity (4.3):

\frac{\partial}{\partial t_j} \varphi_\pi^{\underline{t}}({\mathbf X}) = \left. \frac{\partial}{\partial s} \right\rvert_{s=0} \varphi_\pi^{\underline{t}+se_j}({\mathbf X}) = \sum_{\emptyset\ne A\subseteq P_j} K_{(P_1,P_2,\dots,P_{j-1},P_j\setminus A,A,P_{j+1},\dots,P_p)}^{(t_1,t_2,\dots,t_{j-1},t_j,1,t_{j+1},\dots,t_p)} ({\mathbf X}).

In order to prove the first differential equation (4.3) we replace (Proof) by the complementary expansion

\delta_{[N_j+m]}^{\otimes P_j} = (\delta_{[m]}+\delta_{m+[N_j]})^{\otimes P_j} = \sum_{A\subseteq P_j} \delta_{[m]}^{\otimes A} \delta_{m+[N_j]}^{\otimes P_j \setminus A},

and following the lines of (Proof)–(Proof) we obtain

\begin{gathered} \varphi_\pi^{\underline{N}+me_j}(X_1,\dots, X_n) \\ \begin{aligned}[t] &= \sum_{A\subseteq P_j} \varphi_\pi\!\left( \left( \delta_{[m]}^{\otimes A} \delta_{m+[N_j]}^{\otimes P_j \setminus A} \prod_{i\ne j} \delta_{[N_i]}^{\otimes P_i}\right) ({\mathbf X}) \right) \\ &= \sum_{A\subseteq P_j} \varphi_{(P_1,P_2,\dots,P_{j-1},A, P_j\setminus A, P_{j+1},\dots,P_p)}\left( \left( \delta_{[m]}^{\otimes A} \delta_{[N_j]}^{\otimes P_j \setminus A}\prod_{i\ne j} \delta_{[N_i]}^{\otimes P_i}\right) ({\mathbf X}) \right) \\ &= \sum_{A\subseteq P_j} \varphi_{(P_1,P_2,\dots,P_{j-1},A,P_j\setminus A, P_{j+1},\dots,P_p)}^{(N_1,N_2,\dots,N_{j-1},m,N_j,N_{j+1},\dots,N_p)} ({\mathbf X}). \end{aligned} \end{gathered}

By analytic continuation, we may replace m and N_i with s and t_i respectively. Taking the derivative with respect to s at 0 we obtain (4.3).

Remark

Remark 4.4

JWL8

In the case of exchangeability both differential equations coincide.

Example

Example 4.5 (Tensor independence)

L3N9

Consider the tensor spreadability system \mathcal{S}_{\mathrm{T}}. It satisfies the factorization property (?) and therefore

\varphi_{(A,A^c)}^{(t_1,t_2)}(X_1,\dots, X_n) = \varphi^{t_1}(X_A)\, \varphi^{t_2}(X_{A^c}),

so we get

K_{(A,A^c), A}^{(1,t_2)}(X_1,X_2,\dots, X_n) = \left.\frac{\partial}{\partial t_1}\right\rvert_{t_1=0} \varphi^{t_1}(X_A) \,\varphi^{t_2}(X_{A^c}) = K_{|A|}^{\mathrm{T}}(X_A)\, \varphi^{t_2}(X_{A^c}).

The identities in Theorem 4.3 (for \pi=\hat{1}_n) read

\frac{d}{d t} \varphi^t (X_1,\dots, X_n) = \sum_{\emptyset \neq A \subseteq [n]} K_{|A|}^{\mathrm{T}}(X_A) \,\varphi^{t}(X_{A^c}).

This differential equation can be translated to (exponential) generating functions as follows. Given a vector \underline{u}=(u_1,u_2,\dots, u_n) of commuting indeterminates and a vector \underline{X}=(X_1,X_2,\dots, X_n) of random variables we define the exponential moment generating function

\begin{aligned} \mathcal{F}_{\underline{X}}^t(\underline{u}) &:= 1+\sum_{\substack{(p_1,\dots, p_n) \in (\mathbb{N} \cup \{0\})^n \\ (p_1,\dots, p_n)\neq0}} \frac{u_1^{p_1}\cdots u_n^{p_n}}{p_1! \cdots p_n!}\, \varphi^t(\underbrace{X_1,\dots, X_1}_{\text{$p_1$ times}}, \dots, \underbrace{X_n,\dots X_n}_{\text{$p_n$ times}})\\ &= [\varphi (e^{u_1 X_1} \cdots e^{u_n X_n})]^t \end{aligned}

and the exponential cumulant generating function as the logarithm of the previous

\begin{aligned} \mathcal{L}_{\underline{X}}(\underline{u}) &:= \sum_{\substack{(p_1,\dots, p_n) \in (\mathbb{N} \cup \{0\})^n \\ (p_1,\dots, p_n)\neq0}} \frac{u_1^{p_1}\cdots u_n^{p_n}}{p_1! \cdots p_n!} \, K_{p_1+\dots +p_n}^{\mathrm{T}}(\underbrace{X_1,\dots, X_1}_{\text{$p_1$ times}}, \dots, \underbrace{X_n,\dots X_n}_{\text{$p_n$ times}})\\ &= \log[\mathcal{F}_{\underline{X}}^1(\underline{u})]. \end{aligned}

Then one can prove that

\frac{d}{dt} \mathcal{F}_{\underline{X}}^t(\underline{u}) = \mathcal{L}_{\underline{X}}(\underline{u}) \, \mathcal{F}_{\underline{X}}^t(\underline{u}),

which is equivalent to (4.5). Note that the functions \mathcal{L}_{\underline{X}}(\underline{u}) and \mathcal{F}_{\underline{X}}^t(\underline{u}) commute.

Example

Example 4.6 (Boolean independence)

8U7Y

In the Boolean spreadability system \mathcal{S}_{\mathrm{B}} we have

\varphi_{(A,A^c)}^{(t_1,t_2)}(X_1,X_2,\dots, X_n) = \prod_{P \in \operatorname{\mathrm{\iota_{\max}}}(A)}\varphi^{t_1}(X_P) \prod_{Q\in \operatorname{\mathrm{\iota_{\max}}}(A^c)}\varphi^{t_2}(X_Q).

where \operatorname{\mathrm{\iota_{\max}}}(A) is the interval partition constructed in Definition A.5 and consists of the contiguous subintervals of A. It follows that

K_{(A,A^c), A}^{(1,t_2)}(X_1,X_2\dots, X_n) = \begin{cases} K_{|A|}^{\mathrm{B}}(X_A) \prod_{Q\in \operatorname{\mathrm{\iota_{\max}}}(A^c)}\varphi^{t_2}(X_Q), & \text{if $|\operatorname{\mathrm{\iota_{\max}}}(A)|=1$}, \\ 0, & \text{if $|\operatorname{\mathrm{\iota_{\max}}}(A)|>1$}. \end{cases}

The identities in Theorem 4.3 (for \pi=\hat{1}_n) coincide and read

\frac{d}{d t} \varphi^t (X_1,X_2,\dots, X_n) = \sum_{A: \text{~interval of~} [n]} K_{|A|}^{\mathrm{B}}(X_A) \prod_{Q\in \operatorname{\mathrm{\iota_{\max}}}(A^c)}\varphi^{t}(X_Q).

This differential equation can be interpreted in terms of generating functions in noncommuting indeterminates z_1,\dots, z_n. To this end we define noncommutative formal power series

\begin{aligned} M_{\underline{X}}^t (\underline{z}) &=1 + \sum_{m=1}^\infty \sum_{i_1, \cdots, i_m = 1} ^n \varphi^t(X_{i_1},X_{i_2},\dots, X_{i_m}) \,z_{i_1}\cdots z_{i_m} \intertext{and } K_{\underline{X}}^{\mathrm{B}} (\underline{z}) &= \left.\frac{\partial}{\partial t}\right\rvert_0 M_{(X_1,X_2,\dots, X_n)}^t (\underline{z}) \\ &= \sum_{m=1}^\infty \sum_{i_1, \cdots, i_m = 1} ^n K_m^{\mathrm{B}}(X_{i_1},\dots, X_{i_m}) \, z_{i_1}\cdots z_{i_m}. \end{aligned}

Then differential equation (4.6) is equivalent to the identity

\frac{d}{d t} M_{\underline{X}}^t(\underline{z}) = M_{\underline{X}}^t(\underline{z})\, K_{\underline{X}}^{\mathrm{B}} (\underline{z})\, M_{\underline{X}}^t(\underline{z}).

Note that M_{\underline{X}}^t(\underline{z}) and K_{\underline{X}}^{\mathrm{B}} (\underline{z}) do not commute.

Example

Example 4.7 (Monotone independence)

5HFW

Consider the monotone spreadability system \mathcal{S}_{\mathrm{M}}. With the notation \operatorname{\mathrm{\iota_{\max}}}(A) introduced in Definition A.5 we have

\varphi_{(A,A^c)}^{(t_1,t_2)}(X_1,X_2,\dots, X_n) = \varphi^{t_1}(X_A) \prod_{B \in \operatorname{\mathrm{\iota_{\max}}}(A^c)} \varphi^{t_2}(X_B),

using monotone independence. Thus

\begin{aligned} K_{(A,A^c), A}^{(1,t_2)}(X_1,X_2,\dots, X_n) &= \left.\frac{\partial}{\partial t_1}\right\rvert_{t_1=0} \varphi^{t_1}(X_A) \prod_{B \in \operatorname{\mathrm{\iota_{\max}}}(A^c)} \varphi^{t_2}(X_B) \\ &= K_{|A|}^{\mathrm{M}}(X_A)\prod_{B \in \operatorname{\mathrm{\iota_{\max}}}(A^c)} \varphi^{t_2}(X_B). \end{aligned}

The first identity in Theorem 4.3 (for \pi=\hat{1}_n) reads

\frac{d}{d t} \varphi^t (X_1,X_2,\dots, X_n) = \sum_{ \emptyset \neq A \subseteq [n]} K_{|A|}^{\mathrm{M}}(X_A) \prod_{B \in \operatorname{\mathrm{\iota_{\max}}}(A^c)} \varphi^{t_2}(X_B),

which is exactly the first identity in [HS11b, Corollary 5.2].

On the other hand

\begin{aligned} K_{(A^c,A), A}^{(t_1,1)}(X_1,X_2,\dots, X_n) &= \left.\frac{\partial}{\partial t_2}\right\rvert_{t_2=0} \varphi^{t_1}(X_A) \prod_{B \in \operatorname{\mathrm{\iota_{\max}}}(A^c)} \varphi^{t_2}(X_B) \\ &= \begin{cases} \varphi^{t_1}(X_A)\, K_{|A^c|}^{\mathrm{M}}(X_{A^c}), & \text{if $|\operatorname{\mathrm{\iota_{\max}}}(A^c)|=1$}, \\ 0, & \text{if $|\operatorname{\mathrm{\iota_{\max}}}(A^c)|>1$}. \end{cases} \end{aligned}

Condition |\operatorname{\mathrm{\iota_{\max}}}(A^c)|=1 holds if and only if A^c is an interval. Therefore, the second equality in Theorem 4.3 (for \pi=\hat{1}_n) reads

\frac{d}{d t} \varphi^t (X_1,X_2,\dots, X_n) = \sum_{B: \text{~interval of~} [n]} K_{|B|}^{\mathrm{M}}(X_B)\,\varphi^t(X_{B^c}),

which is exactly the second equality in [HS11b, Corollary 5.2].

Results on generating functions in [HS11b] correspond to these differential equations. For noncommutative indeterminates z_1,\dots, z_n, we define the cumulant generating function

\begin{aligned} K_{\underline{X}}^{\mathrm{M}} (\underline{z}) &:= \left.\frac{\partial}{\partial t}\right\rvert_0 M_{\underline{X}}^t (\underline{z}) \\ &= \sum_{m=1}^\infty \sum_{i_1,i_2, \cdots, i_m = 1} ^n K_m^{\mathrm{M}}(X_{i_1}X_{i_2}\cdots X_{i_m}) \, z_{i_1}z_{i_2}\cdots z_{i_m}. \end{aligned}

Then it is shown in [HS11b, Theorem 6.3] that

M_{\underline{X}}^{s+t} (\underline{z}) = M_{\underline{X}}^t (\underline{z})\, M_{\underline{X}}^s (z_1M_{\underline{X}}^t (\underline{z}) , \dots, z_n M_{\underline{X}}^t (\underline{z}) ).

The partial derivatives of (?) regarding s at 0 and t at 0 become (4.7) and (?), respectively.

Example

Example 4.8 (Free independence)

8VZ2

Consider the free spreadability system \mathcal{S}_{\mathrm{F}}. One can use the formula for products of free random variables [NicaSpeicher:2006:lectures, Theorem 14.4] to show that for a nonempty subset A\subseteq [n] we can expand

\varphi_{(A,A^c)}(X_1,X_2,\dots, X_n) = \varphi(X_A)\prod_{P\in \operatorname{\mathrm{\tilde{\iota}_{\max}}}(A^c)}\varphi(X_P) + R,

where \operatorname{\mathrm{\tilde{\iota}_{\max}}}(A^c) is the partition defined in Definition A.5 ? and every term in R has at least two factors from A, i.e., factors of the form \varphi(X_{k_1}X_{k_2}\cdots X_{k_m}), k_1,k_2,\dots, k_m \in A. For example, \varphi_{\{\{2,4\}, \{1,3,5\}\}}(X_1,X_2,\dots, X_5) = \varphi(X_2 X_4)\, \varphi(X_1 X_5)\, \varphi(X_3) + R, where every term in R contains the factor \varphi(X_2)\,\varphi(X_4). This implies that

\varphi_{(A,A^c)}^{(t_1,t_2)}(X_1,X_2,\dots, X_n) = \varphi^{t_1}(X_A)\prod_{P\in \operatorname{\mathrm{\tilde{\iota}_{\max}}}(A^c)}\varphi^{t_2}(X_P) +O(t_1^2) \qquad \text{as $t_1\to0$},

so by taking the partial derivative \left.\frac{\partial}{\partial t_1}\right\rvert_{t_1=0} we get

K_{(A,A^c), A}^{(1,t_2)}(X_1,X_2,\dots, X_n) = K_{|A|}^{\mathrm{F}}(X_A)\prod_{P\in \operatorname{\mathrm{\tilde{\iota}_{\max}}}(A^c)}\varphi^{t_2}(X_P).

This yields the differential equation

\frac{d}{d t} \varphi^t (X_1,X_2,\dots, X_n) = \sum_{\emptyset \neq A \subseteq [n]} K_{|A|}^{\mathrm{F}}(X_A)\prod_{P\in \operatorname{\mathrm{\tilde{\iota}_{\max}}}(A^c)}\varphi^t(X_P),

which is similar to the monotone case (4.7).

Denote by \underline{z}=(z_1,z_2, \dots, z_n) a vector of noncommuting indeterminates and by \underline{X}=(X_1,X_2,\dots,X_n) a random vector. In order to obtain a differential equation we need in addition the two-sided generating function

\tilde{M}^t_{\underline{X}}(\underline{z},w):= \sum_{p,q=0}^\infty\sum_{\substack{i_1,i_2,\dots, i_p \in[n] \\ j_1,j_2,\dots, j_q \in[n]}} \varphi^t(X_{i_1},X_{i_2}, \dots, X_{i_p}, X_{j_1},X_{j_2}, \dots, X_{j_q})\,z_{i_1}z_{i_2}\cdots z_{i_p}w z_{j_1}z_{j_2}\cdots z_{j_q}

as well as the R-transform

R_{\underline{X}}(\underline{z})=\sum_{j=1}^\infty \sum_{i_1, \cdots, i_j = 1} ^n K_j^{\mathrm{F}}(X_{i_1},X_{i_2},\cdots, X_{i_j})\,z_{i_1}z_{i_2}\cdots z_{i_j}.

After some computations we infer from (4.8) that

\frac{\partial }{\partial t} M^t_{\underline{X}}(\underline{z}) = \tilde{M}^t_{\underline{X}}(\underline{z}, R(z_1M^t_{\underline{X}}(\underline{z}),z_2 M^t_{\underline{X}}(\underline{z}) ,\dots, z_n M^t_{\underline{X}}(\underline{z})) (M^t_{\underline{X}}(\underline{z}))^{-1}).

Note that by [NicaSpeicher:2006:lectures, Corollary 16.16] the second argument can be written as

R(z_1 M^t_{\underline{X}}(\underline{z}), z_2 M^t_{\underline{X}}(\underline{z}) ,\dots, z_n M^t_{\underline{X}}(\underline{z}))\,(M^t_{\underline{X}}(\underline{z}))^{-1} =\frac{1- M^t_{\underline{X}}(\underline{z})^{-1}}{t}.

When n=1 the differential equation (?) is equivalent to the generalized complex Burgers equation (see [V86, p. 343])

\frac{\partial }{\partial t} G^t_{X}(z) + \frac{R_X(G_X^t(z))}{G^t_X(z)} \frac{\partial }{\partial z} G^t_{X}(z)=0,

where G^t_X is the Cauchy transform

G^t_X(z)= \sum_{n=0}^\infty \varphi^t(\underbrace{X,\dots,X}_{\text{$n$ fold}}) \, z^{-n-1}.
Proposition

Proposition 5.1

90DC

Let (\mathcal{A},\varphi) be a \mathcal{B}-ncps and \mathcal{E}=(\mathcal{U},\tilde{\varphi},(\iota^{(i)})_{i\ge1}) an exchangeability system for \mathcal{A}. Given a partition \pi\in\mathcal{P}_n and a family X_1,X_2,\dots,X_n\in \mathcal{A} such that there is a block P\in \pi which can be partitioned into P=P_1\dot\cup P_2 such that \{X_i:i\in P_1\} and \{X_i:i\in P_2\} are independent in the sense of Definition 2.12, we have K_\pi(X_1,X_2\dots,X_n)=0.

Proof of Proposition 5.1

Let \rho be the partition obtained from \pi by splitting the block P as indicated in the proposition, then \rho<\pi and by assumption \varphi_\sigma(X_1,X_2,\dots,X_n)=\varphi_{\sigma\wedge\rho}(X_1,X_2,\dots,X_n) for any \sigma\in\mathcal{P}_n; hence

\begin{aligned} K_\pi(X_1,X_2,\dots,X_n) &= \sum_{\substack{\sigma\in\mathcal{P}_n \\\sigma\leq\pi}} \varphi_\sigma(X_1,X_2,\dots,X_n)\, \mu_\mathcal{P}(\sigma,\pi) \\ &= \sum_{\substack{\sigma\in\mathcal{P}_n \\\sigma\leq\pi}} \varphi_{\sigma\wedge\rho}(X_1,X_2,\dots,X_n)\, \mu_\mathcal{P}(\sigma,\pi) \\ & =\sum_{\tau\in\mathcal{P}_n} \varphi_\tau(X_1,X_2,\dots,X_n) \sum_{\substack{\sigma \in\mathcal{P}_n\\ \sigma\wedge\rho =\tau}}\, \mu_\mathcal{P}(\sigma,\pi). \end{aligned}

Now \rho\not\geq \pi and the second case of Weisner’s lemma applies.

Lemma

Lemma 5.2 (Weisner’s Lemma)

S8QM

In any lattice (P,\leq) the Möbius function satisfies the identity

\sum_{\substack{x\\ x\wedge a=c}} \mu(x,b) = \begin{cases} \mu(c,b) & \text{if $a\geq b$}\\ 0 &\text{if $a\not\geq b$}. \end{cases}
Definition

Definition 5.3

ER0Z

For \tau,\eta \in \mathcal{OP}_n let

\begin{aligned} w(\tau,\eta)&:=\sum_{\substack{\sigma \in \mathcal{OP}_n \\ \sigma \curlywedge \eta=\tau}} \widetilde{\mu}(\sigma,\hat{1}_n), && \text{(Weisner coefficients)} \\ g(\tau,\eta)&:=\sum_{\substack{\sigma \in \mathcal{OP}_n \\ \sigma \geq \tau}}\widetilde{\zeta}(\tau,\sigma)\,w(\sigma,\eta), &&\text{(Goldberg coefficients)} \end{aligned}

and more generally for \tau,\eta,\pi \in \mathcal{OP}_n define the partitioned Weisner and Goldberg coefficients

\begin{aligned} w(\tau,\eta,\pi) &:=\sum_{\substack{\sigma \in \mathcal{OP}_n \\ \sigma \curlywedge \eta=\tau\\ \sigma \leq \pi}} \widetilde{\mu}(\sigma,\pi), \\ g(\tau,\eta,\pi) &:=\sum_{\substack{\sigma \in \mathcal{OP}_n \\ \sigma \geq \tau}}\widetilde{\zeta}(\tau,\sigma)\,w(\sigma,\eta,\pi). \end{aligned}
Remark

Remark 5.4

FWGC

It follows from the properties of the quasi-meet operation (cf. Proposition A.21 ? and ?) that both w(\tau,\eta,\pi) and g(\tau,\eta,\pi) vanish unless \bar{\tau}\leq\bar{\eta} and \tau\leq\pi.

Used by Proof
Proposition

Proposition 5.5

9FFB

Let \mathcal{S}=(\mathcal{U},\tilde{\varphi}, (\iota^{(i)})_{i\ge 1}) be a spreadability system for a given \mathcal{B}-ncps (\mathcal{A},\varphi). Let \pi\in\mathcal{OP}_n and (i_1,i_2,\ldots,i_n) \in \mathbb{N}^n be a tuple with kernel \eta=\kappa(i_1,i_2,\ldots,i_n)\in\mathcal{OP}_n. Then

\begin{aligned} K_\pi(X_1^{(i_1)}, X_2^{(i_2)}, \ldots, X_n^{(i_n)}) &= \sum_{\substack{\tau \in \mathcal{OP}_n \\ \tau\leq\pi\\ \bar{\tau}\leq\bar{\eta}}} \varphi_\tau(X_1,X_2,\ldots,X_n) \, w(\tau,\eta,\pi) \\ &= \sum_{\substack{\tau \in \mathcal{OP}_n \\ \tau\leq\pi\\ \bar{\tau}\leq\bar{\eta}}} K_\tau(X_1,X_2,\ldots,X_n) \, g(\tau,\eta,\pi) . \end{aligned}
Proof of Proposition 5.5

We proceed as in the proof of Proposition 5.1. Expressing cumulants in terms of moments (see Theorem 3.11), we have

\begin{aligned} K_\pi(X_1^{(i_1)},X_2^{(i_2)}, \ldots, X_n^{(i_n)}) &= \sum_{\substack{\sigma \in \mathcal{OP}_n \\ \sigma \leq \pi}} \varphi_\sigma(X_1^{(i_1)}, X_2^{(i_2)},\ldots,X_n^{(i_n)})\,\widetilde{\mu}(\sigma,\pi) \\ &= \sum_{\substack{\sigma \in \mathcal{OP}_n \\ \sigma \leq \pi}} \varphi_{\sigma\curlywedge\eta}(X_1,X_2,\ldots,X_n)\, \widetilde{\mu}(\sigma,\pi) \\ &=\sum_{\tau \in \mathcal{OP}_n} \varphi_\tau(X_1,X_2,\ldots,X_n)\,\biggl(\sum_{\substack{\sigma \in \mathcal{OP}_n \\ \sigma\curlywedge\eta =\tau, \sigma \leq \pi }} \widetilde{\mu}(\sigma,\pi) \biggr) \\ &= \sum_{\tau \in \mathcal{OP}_n} \varphi_\tau(X_1,X_2,\ldots,X_n) \, w(\tau,\eta,\pi) \\ \intertext{which is \eqref{eq:cum-mom}. Now substitute the moment-cumulant formula \eqref{eq:phipi=sumKsi} to obtain } &= \sum_{\tau \in \mathcal{OP}_n} \sum_{\substack{\sigma \in\mathcal{OP}_n\\ \sigma \leq \tau}}K_\sigma(X_1,X_2,\ldots,X_n) \, \widetilde{\zeta}(\sigma,\tau)\,w(\tau,\eta,\pi) \\ &= \sum_{\sigma \in \mathcal{OP}_n} K_\sigma(X_1,X_2,\ldots,X_n) \,\biggl(\sum_{\substack{\tau \in\mathcal{OP}_n\\ \tau\geq\sigma}}\widetilde{\zeta}(\sigma,\tau)\,w(\tau,\eta,\pi)\biggr) \\ &= \sum_{\sigma \in \mathcal{OP}_n} K_\sigma(X_1,X_2,\ldots,X_n)\,g(\sigma,\eta,\pi). \end{aligned}

Finally Remark 5.4 applies and the sums (5.5) and (5.5) can be restricted to \tau\leq\pi and \bar{\tau}\leq\bar{\eta}.

Theorem

Theorem 5.6

9BPI

Let \mathcal{S}=(\mathcal{U},\tilde{\varphi},(\iota^{(i)})_{i\ge1}) be a spreadability system for a given \mathcal{B}-ncps (\mathcal{A},\varphi). A sequence of subalgebras (\mathcal{A}_i)_{i\in I} of \mathcal{A}, where I\subseteq \mathbb{N}, is \mathcal{S}-independent if and only if for any tuple (i_1,i_2,\dots,i_n)\in I^n, any random variables (X_1,X_2,\dots,X_n)\in \mathcal{A}_{i_1} \times \mathcal{A}_{i_2} \times \cdots \times \mathcal{A}_{i_n} and any ordered set partition \pi\in\mathcal{OP}_n, we have

K_\pi(X_1,X_2, \ldots, X_n) - \sum_{\substack{\tau \in \mathcal{OP}_n\\ \tau\leq\pi}} K_\tau(X_1,X_2,\ldots,X_n) \, g(\tau,\kappa(i_1,\dots,i_n),\pi) = 0.
Remark

Remark 5.7

B7O3

We can also formulate the theorem in terms of moments; we only need to replace (5.6) by the equation

K_\pi(X_1,X_2, \ldots, X_n) - \sum_{\substack{\tau \in \mathcal{OP}_n\\ \tau\leq\pi}} \varphi_\tau(X_1,X_2,\ldots,X_n) \, w(\tau,\kappa(i_1,\dots,i_n),\pi) = 0.
Proof of Remark 5.7

Fix a tuple (i_1,\dots, i_n) with kernel \eta=\kappa(i_1,\dots,i_n). Independence means that

\begin{aligned} \varphi_\pi(X_1,X_2,\dots,X_n) &= \varphi_{\pi\curlywedge\eta}(X_1,X_2,\dots,X_n) \\ &= \varphi_\pi(X_1^{(i_1)},X_2^{(i_2)},\dots,X_n^{(i_n)}) \end{aligned}

for all \pi \in \mathcal{OP}_n, and by Möbius inversion this is equivalent to the identity

\begin{aligned} K_\pi(X_1,X_2,\dots,X_n) &= K_\pi(X_1^{(i_1)},X_2^{(i_2)},\dots,X_n^{(i_n)}) \\ &= \sum_{\substack{\tau \in \mathcal{OP}_n\\ \tau\leq\pi}} K_\tau(X_1,X_2,\ldots,X_n) \, g(\tau,\kappa(i_1,\dots,i_n),\pi) \end{aligned}

for all \pi\in\mathcal{OP}_n.

Remark

Remark 5.8

R79V

It is not obvious that Theorem 5.6 generalizes Proposition 5.1 (vanishing of mixed cumulants for \mathcal{E}-independent subalgebras). In fact it implies a nontrivial identity: for any \tau,\eta,\pi \in \mathcal{OP}_n such that \eta\rfloor_P \neq \hat{1}_P for some P\in\pi,

\sum_{h \in \mathfrak{S}_{|\tau|}} g(h(\tau), \eta, \pi)=0,

where h(\tau) is the action of the permutation h on the blocks of \tau. Similarly, Remark 5.7 (or Proposition 5.5) generalizes the vanishing of mixed cumulants for \mathcal{E}-independent subalgebras. Consequently we must have

\sum_{h \in \mathfrak{S}_{|\tau|}} w(h(\tau), \eta, \pi)=0

under the same assumptions on \tau,\eta,\pi.

Definition

Definition 5.9

HMLS

A multiset is a pair (A,f) where A is the underlying set and f:A\to\mathbb{N} is a function. The value f(a) is called the multiplicity of the element a\in A. Informally, a multiset is a set which contains multiple indistinguishable copies of each of its elements. In the present paper multisets will always be based on integer segments A=[n] and in this case the tuple (f(1),f(2),\dots,f(n)) is called the type of the multiset. A permutation of a multiset is a rearrangement of all its elements where all multiplicities are preserved, i.e., a word with a prescribed total number of occurrences of each letter. The number of distinct permutations of a multiset of type (f_1,f_2,\dots,f_n) is given by the multinomial coefficient

\binom{f_1+f_2+\dotsm+ f_n}{f_1,f_2,\dots,f_n}.

A proper set is a multiset of type (1,1,\dots,1) and we recover the number of its permutations as n!.

Definition

Definition 5.10

MJR5

Let \sigma=w_1w_2\dots w_s be a permutation of a multiset of type (f_1,f_2,\dots,f_n) where s=f_1+f_2+\dotsm+ f_n. An index 1\leq i\leq s-1 is called a

  1. descent (or drop or fall) if w_i>w_{i+1}.

  2. plateau (or level) if w_i=w_{i+1}.

  3. ascent (or rise) if w_i<w_{i+1}.

We denote these sets by

\begin{aligned} \operatorname{Des}(\sigma)&=\operatorname{Des}((w_i)_{i=1}^s)=\{j\in [s-1]\mid w_j > w_{j+1} \},\\ \operatorname{Pla}(\sigma)&=\operatorname{Pla}((w_i)_{i=1}^s)=\{j\in [s-1]\mid w_j = w_{j+1} \},\\ \operatorname{Asc}(\sigma)&=\operatorname{Asc}((w_i)_{i=1}^s)=\{j\in [s-1]\mid w_j < w_{j+1} \} \end{aligned}

and the respective cardinalities, i.e., the number of descents (ascents, plateaux respectively) of the multiset permutation \sigma, by

\begin{aligned} \operatorname{\mathrm{des}}(\sigma)=\operatorname{\mathrm{des}}((w_i)_{i=1}^s)&=\left\lvert \operatorname{Des}((w_i)_{i=1}^s) \right\rvert,\\ \operatorname{\mathrm{pla}}(\sigma)=\operatorname{\mathrm{pla}}((w_i)_{i=1}^s)&=\left\lvert \operatorname{Pla}((w_i)_{i=1}^s) \right\rvert, \\ \operatorname{\mathrm{asc}}(\sigma)=\operatorname{\mathrm{asc}}((w_i)_{i=1}^s)&=\left\lvert \operatorname{Asc}((w_i)_{i=1}^s) \right\rvert. \end{aligned}
Remark

Remark 5.11

I26T
  1. Note that some authors also count i=s as a descent and i=0 as an ascent.

  2. Counting permutations by descents and ascents is a classic subject in combinatorics. In the case of descents of multiset permutations this is also known as Simon Newcomb’s problem [Riordan:1958:combinatorial, DillonRoselle:1969:simon].

  • Clearly every element except the last is either a descent, an ascent, or a plateau, and therefore

    \left\lvert \sigma \right\rvert = \operatorname{\mathrm{des}}(\sigma)+\operatorname{\mathrm{pla}}(\sigma)+\operatorname{\mathrm{asc}}(\sigma) + 1 .
  • Definition

    Definition 5.12

    4EXT

    To any pair of ordered set partitions \tau,\eta\in\mathcal{OP}_n such that \bar{\tau} \leq \bar{\eta}, we associate a multiset on [\left\lvert \eta \right\rvert] by setting the multiplicity of k\in \{1,2,\dots,\left\lvert \eta \right\rvert\} to be equal to the number of blocks of \tau contained in the k-th block of \eta. Replacing every block of \tau by the label of the block of \eta containing it we obtain a permutation of this multiset. We denote by \operatorname{\mathrm{des}}_\eta(\tau), \operatorname{\mathrm{asc}}_\eta(\tau) and \operatorname{\mathrm{pla}}_\eta(\tau) its statistics as defined in Definition 5.10. More precisely, let \tau,\eta\in\mathcal{OP}_n such that \bar{\tau} \leq \bar{\eta}. If \eta= (E_1,E_2,\ldots,E_e), then the blocks of \bar{\tau} can be arranged

    \bar{\tau} = \{ E_{1,1},E_{1,2},\ldots, E_{1,l_1}, \ldots, E_{e,1},\ldots, E_{e,l_e}\},

    where each E_i = \bigcup_{j=1}^{l_i} E_{i,j} is a disjoint union taken in canonical order of the subsets (i.e., sorted according to their minimal elements). So \tau can be written as

    \tau=(E_{m_1,n_1},E_{m_2,n_2}, \ldots, E_{m_s, n_s}),

    which is a permutation of the blocks of \bar{\tau}. Then we denote by \operatorname{Des}_\eta(\tau), \operatorname{Pla}_\eta(\tau) and \operatorname{Asc}_\eta(\tau) respectively, the sets

    \begin{aligned} \operatorname{Des}_\eta(\tau) &=\operatorname{Des}((m_i)_{i=1}^s), \\ \operatorname{Pla}_\eta(\tau) &= \operatorname{Pla}((m_i)_{i=1}^s), \\ \operatorname{Asc}_\eta(\tau) &= \operatorname{Asc}((m_i)_{i=1}^s) \end{aligned}

    and by \operatorname{\mathrm{des}}_\eta(\tau), \operatorname{\mathrm{pla}}_\eta(\tau) and \operatorname{\mathrm{asc}}_\eta(\tau) the respective cardinalities.

    Example

    Example 5.13

    D072

    Consider the partitions \eta=(E_1,E_2,E_3,E_4,E_5)\in \mathcal{OP}_{10} with blocks E_1=\{5,8\}, E_2=\{9,10\}, E_3=\{3,6\}, E_4=\{1,2,4\}, E_5=\{7\}, and \tau=(T_1, T_2, T_3,T_4,T_5,T_6,T_7,T_8)\in\mathcal{OP}_{10} with blocks T_1=\{3\}, T_2=\{6\}, T_3=\{1,4\}, T_4=\{7\}, T_5=\{10\}, T_6=\{5,8\}, T_7=\{9\}, T_8=\{2\}. Then \bar{\tau}\leq \bar{\eta}. Now T_1\subseteq E_3 and thus w_1=3, T_2 \subseteq E_3 and thus w_2=3, etc.; the multiset permutation thus induced on \tau by \eta is

    (w_i)_{i=1}^{8}=(3,3,4,5,2,1,2,4),

    and the statistics are \operatorname{\mathrm{des}}((w_i)_{i=1}^8)=2, \operatorname{\mathrm{pla}}((w_i)_{i=1}^8)=1 and \operatorname{\mathrm{asc}}((w_i)_{i=1}^8)=4.

    Definition

    Definition 5.14

    IFCS

    Let \sigma=(w_i)_{i=1}^s be a multiset permutation. An ascending (resp. descending) run is a maximal contiguous subsequence (w_i)_{i=l}^k which is strictly increasing (resp. decreasing). Similarly, a level run is a maximal subsequence (w_i)_{i=j}^k such that w_j=w_{j+1}=\cdots=w_k.

    Proposition

    Proposition 5.15

    VLBS

    Let \sigma=(w_i)_{i=1}^s be a multiset permutation.

    1. \sigma can be decomposed uniquely into ascending runs separated by descents and plateaux. The number of ascending runs is equal to s-\operatorname{\mathrm{asc}}(\sigma).

    2. \sigma can be decomposed uniquely into descending runs separated by ascents and plateaux. The number of descending runs is equal to s-\operatorname{\mathrm{des}}(\sigma).

    3. \sigma can be decomposed uniquely into level runs separated by ascents and descents. The number of level runs is equal to s-\operatorname{\mathrm{pla}}(\sigma).

    Proof of Proposition 5.15

    We only prove (i). The claim is clearly true when \operatorname{\mathrm{asc}}\sigma = s-1, i.e., when the sequence is monotone increasing. Otherwise replacing any ascent by a descent or plateau splits an ascending run into two, i.e., increases the number of ascending runs by one.

    Remark

    Remark 5.16

    OTC1

    Suppose \bar{\tau}=\bar{\eta}, then the blocks of \tau are a permutation of the blocks of \eta. More precisely, if \eta=(E_1,E_2,\dots,E_e) then there is a permutation h\in \mathfrak{S}_e such that \tau=h(\eta)=(E_{h(1)},E_{h(2)}, \dots,E_{h(e)}). Then \operatorname{\mathrm{des}}_\eta(\tau) and \operatorname{\mathrm{asc}}_\eta(\tau) coincide with \operatorname{\mathrm{des}}(h) and \operatorname{\mathrm{asc}}(h), where the latter quantities are the numbers of descents and ascents of the permutation h, respectively. See [Bona:2012:combinatorics, Brenti:1993:permutation, Reutenauer:1993:freeLiealg] and [Stanley:1986:enumerative1, p. 25] for the uses of \operatorname{\mathrm{des}}(h) and \operatorname{\mathrm{asc}}(h) in the context of symmetric groups.

    Example

    Example 5.17

    CTP4

    The statistics of the sequence (m_i)_{i=1}^9=(1,1,3,5,5,5,4,1,4) are as follows: \operatorname{\mathrm{des}}((m_i)_{i=1}^9)=2, \operatorname{\mathrm{pla}}((m_i)_{i=1}^9)=3 and \operatorname{\mathrm{asc}}((m_i)_{i=1}^9)=3. The decomposition into ascending runs is given by

    (1), (1,3,5), (5), (5), (4), (1,4),

    the decomposition into level runs is

    (1,1), (3), (5,5,5), (4), (1), (4)

    and the decomposition into descending runs is

    (1),(1),(3),(5),(5),(5,4,1),(4).
    Lemma

    Lemma 5.18

    64HY

    Let \eta, \tau \in\mathcal{OP}_n such that \bar{\tau} \leq \bar{\eta}, then there is an ordered set partition \sigma_{\mathrm{max}}^{\mathrm{asc}}(\tau,\eta) such that

    \{\sigma \in\mathcal{OP}_n \mid \sigma \curlywedge \eta =\tau \} =[\tau, \sigma_{\mathrm{max}}^{\mathrm{asc}}(\tau,\eta)].

    and the restriction of the mapping \Psi in Proposition A.24 establishes a poset isomorphism

    \{\sigma \in\mathcal{OP}_n \mid \sigma \curlywedge \eta =\tau \} \to \mathcal{I}_{p_1} \times \mathcal{I}_{p_2} \times \cdots \times \mathcal{I}_{p_t},

    where 1 \leq p_i \leq n are the lengths of the ascending runs of the sequence (m_i)_{i=1}^s from Definition 5.12 and t=\left\lvert \tau \right\rvert-\operatorname{\mathrm{asc}}_\eta(\tau).

    Uses Definition 5.12 · Used by Proof
    Proof of Lemma 5.18

    Let \eta=(E_1,E_2,\ldots,E_e) and write \tau=(E_{m_1,n_1},E_{m_2,n_2}, \ldots, E_{m_s, n_s}) as in Definition 5.12. From Proposition A.21 we infer that any partition \sigma\in\mathcal{OP}_n such that \sigma\curlywedge\eta=\tau satisfies \sigma \geq \tau. From Proposition A.24 we infer that there is an interval partition \lambda=(L_1,L_2,\ldots, L_l) \in \mathcal{I}_{s} such that

    \sigma=\biggl(\bigcup_{i\in L_1} E_{m_i,n_i},\bigcup_{i\in L_2} E_{m_i,n_i}, \ldots, \bigcup_{i\in L_l} E_{m_i,n_i}\biggr).

    In order that \sigma \curlywedge \eta =\tau it is necessary and sufficient that every block L\in\lambda, say L=\{a+1, a+2, \ldots,a+b\}, induces a strictly increasing sequence m_{a+1} < m_{a+2}< \cdots <m_{a+b}.

    Let (m_i)_{i=1}^{i_1}, (m_i)_{i=i_1+1}^{i_2},\ldots, (m_i)_{i=i_{t-1}+1}^s be the decomposition of (m_i)_{i=1}^s into ascending runs. This decomposition defines the interval blocks A_j:= \{i_{j-1}+1, i_{j-1}+2,\ldots, i_j\} (i_0=0,i_t=s) and hence defines an interval partition \alpha = (A_1,A_2,\ldots,A_t). The interval partition \lambda consists of increasing intervals and therefore is finer than \alpha. Thus we have an isomorphism

    \{\sigma \in\mathcal{OP}_n \mid \sigma \curlywedge \eta =\tau \} \to \mathcal{I}_{A_1} \times\mathcal{I}_{A_2} \times \cdots \times \mathcal{I}_{A_t}

    via the restriction of the map \Psi\rfloor_{[\tau,\hat{1}_n]} from Proposition A.24. The number t=\left\lvert \alpha \right\rvert is equal to \left\lvert \tau \right\rvert-\operatorname{\mathrm{asc}}_\eta(\tau), the integers p_j are the cardinalities of A_j and \sigma_{\mathrm{max}}^{\mathrm{asc}}(\tau,\eta) is the ordered set partition corresponding to \lambda =\alpha.

    Proposition

    Proposition 5.19 (Weisner coefficients)

    IHIA
    1. For \tau,\eta \in \mathcal{OP}_n the Weisner coefficient (5.3) is

      w(\tau,\eta) = \begin{cases}\displaystyle \int_{-1}^0 x^{\left\lvert \tau \right\rvert -\operatorname{\mathrm{asc}}_\eta(\tau)-1}(1+x)^{\operatorname{\mathrm{asc}}_\eta(\tau)}\,dx=\frac{(-1)^{\left\lvert \tau \right\rvert-\operatorname{\mathrm{asc}}_\eta(\tau)-1}}{\left\lvert \tau \right\rvert\binom{\left\lvert \tau \right\rvert-1}{\operatorname{\mathrm{asc}}_\eta(\tau)}}, &\bar{\tau} \leq \bar{\eta},\\ 0, & \bar{\tau} \not\leq \bar{\eta}. \end{cases}
    2. For \tau,\eta,\pi \in \mathcal{OP}_n the partitioned Weisner coefficient (5.3) is

      w(\tau,\eta,\pi) = \begin{cases}\displaystyle \prod_{P\in\pi} w(\tau\rfloor_P,\eta\rfloor_P), &\bar{\tau} \leq \bar{\eta},~\tau \leq \pi,\\ 0, & \text{otherwise}. \end{cases}
    Used by Proof, Proof
    Proof of Proposition 5.19

    ?   If \bar{\tau} \not\leq \bar{\eta}, then there is no \sigma such that \sigma \curlywedge \eta=\tau and the sum is empty. Let us therefore assume henceforth that \bar{\tau} \leq \bar{\eta}. We take up the end of the proof of Lemma 5.18 where we established the poset isomorphism

    \begin{aligned} \{\sigma \in\mathcal{OP}_n \mid \sigma \curlywedge \eta=\tau \} & \cong\mathcal{I}_{p_1} \times \mathcal{I}_{p_2} \times \cdots \times \mathcal{I}_{p_t} \\ & \cong \mathcal{B}_{p_1-1} \times \mathcal{B}_{p_2-1} \times \cdots \times \mathcal{B}_{p_t-1} \\ & \cong \mathcal{B}_{p_1+\cdots+p_t -t}, \end{aligned}

    where the second isomorphism follows from Proposition A.4 and p_i denotes the length of the i-th ascending run. The latter contains p_i-1 rises and therefore the total number of ascents is \operatorname{\mathrm{asc}}_\eta(\tau) = p_1+p_2+\cdots +p_t-t. In the identification above, a partition \sigma is mapped to a subset A\subseteq [p_1+p_2+\cdots+p_t-t] with \left\lvert A \right\rvert = \left\lvert \sigma \right\rvert-t elements and we have

    \widetilde{\mu}(\sigma,\hat{1}_n)= \frac{(-1)^{\left\lvert \sigma \right\rvert-1}}{\left\lvert \sigma \right\rvert} = \frac{(-1)^{\left\lvert A \right\rvert+t-1}}{\left\lvert A \right\rvert+t}.

    Performing the sum we obtain

    \begin{aligned} \sum_{\substack{\sigma \in \mathcal{OP}_n \\ \sigma \curlywedge \eta=\tau}}\widetilde{\mu}(\sigma,\hat{1}_n) &= \sum_{A \subseteq [\operatorname{\mathrm{asc}}_\eta(\tau)]}\frac{(-1)^{\left\lvert A \right\rvert+t-1}}{\left\lvert A \right\rvert+t}\\ &= \sum_{k=0}^{\operatorname{\mathrm{asc}}_\eta(\tau)}\binom{\operatorname{\mathrm{asc}}_\eta(\tau)}{k} \frac{(-1)^{k+t-1}}{k+t} \\ &= \int_{-1}^0 \sum_{k=0}^{\operatorname{\mathrm{asc}}_\eta(\tau)}\binom{\operatorname{\mathrm{asc}}_\eta(\tau)}{k} x^{k+t-1}\,dx \\ &= \int_{-1}^0 x^{t-1} (1+x)^{\operatorname{\mathrm{asc}}_\eta(\tau)}\,dx \\ &= (-1)^{t-1}B(t, \operatorname{\mathrm{asc}}_\eta(\tau)+1), \end{aligned}

    where B is the beta function B(a,b)= \frac{\Gamma(a )\Gamma(b )}{\Gamma(a+b)} which can be written in terms of binomial coefficients as desired.

    ?   In order for the set \{\sigma\in\mathcal{OP}_n \mid \sigma \leq \pi,~ \sigma \curlywedge \eta=\tau\} to be nonempty, it is necessary that \tau \leq \pi and \bar{\tau}\leq \bar{\eta}. We adopt the notations from Definition 5.12 and infer from Proposition A.24 that there exists an interval partition \rho=(R_1,R_2,\dots,R_p) \in \mathcal{I}_{s} such that

    \pi = (P_1,P_2,\dots,P_p)= \biggl(\bigcup_{i\in R_1} E_{m_i,n_i},\bigcup_{i\in R_2} E_{m_i,n_i}, \dots, \bigcup_{i\in R_p} E_{m_i,n_i}\biggr).

    If follows from Lemma 5.18 that \sigma belongs to [\tau, \sigma_{\max{}}^{\mathrm{asc}}(\tau,\eta)]. In addition, \sigma must satisfy \sigma \leq \pi. Hence, the ascending runs considered in Lemma 5.18 are split by the blocks of \pi. More precisely, for each k\in[p] we decompose (m_i)_{i\in R_k} into ascending runs, which give rise to the interval partition \gamma_k=(G_{k,1},\dots, G_{k,u_k}) \in \mathcal{I}_{R_k} where G_{k,j} consists of the indices i of the j-th ascending run of (m_i)_{i\in R_k}. Then

    \sigma_{\max{}}^{\mathrm{asc}}(\tau\rfloor_{P_k},\eta\rfloor_{P_k})= \biggl(\bigcup_{i\in G_{k,1}} E_{m_i,n_i},\bigcup_{i\in G_{k,2}} E_{m_i,n_i},\dots, \bigcup_{i\in G_{k,u_k}} E_{m_i,n_i}\biggr)

    and for each P\in\pi, we pick an arbitrary \sigma_P \in\mathcal{OP}_P from the interval [\tau\rfloor_P, \sigma_{\max{}}^{\mathrm{asc}}(\tau\rfloor_P,\eta\rfloor_P)], concatenate them and obtain \sigma=\sigma_{P_1} \sigma_{P_2}\cdots\sigma_{P_p} \in \mathcal{OP}_n. Since \widetilde{\mu}(\sigma,\pi) is the product of \frac{(-1)^{\#(\sigma\rfloor_P)-1}}{\#(\sigma\rfloor_P)} over P\in\pi, the conclusion follows.

    Lemma

    Lemma 5.20

    BOXT

    Let \eta, \tau \in\mathcal{OP}_n such that \bar{\tau} \leq \bar{\eta}, then the restriction of the map \Psi\rfloor_{[\tau,\hat{1}_n]} from Proposition A.24 establishes a poset isomorphism

    \{\sigma \in\mathcal{OP}_n \mid \sigma \geq \tau, \bar{\sigma} \leq \bar{\eta} \} \to \mathcal{I}_{q_1} \times \mathcal{I}_{q_2} \times \cdots \times \mathcal{I}_{q_r},

    where 1 \leq q_i \leq n are the lengths of the level runs and the number r is equal to \left\lvert \tau \right\rvert-\operatorname{\mathrm{pla}}_\eta(\tau). In particular, there is an ordered set partition \sigma_{\max{}}^{\mathrm{pla}}(\tau,\eta) such that

    \{\sigma \in\mathcal{OP}_n \mid \sigma \geq \tau, \bar{\sigma} \leq \bar{\eta} \} =[\tau, \sigma_{\max{}}^{\mathrm{pla}}(\tau,\eta)].
    Proof of Lemma 5.20

    The proof is similar to that of Lemma 5.18. Write \tau=(E_{m_1,n_1},E_{m_2,n_2}, \ldots, E_{m_s, n_s}) and \eta=(E_1,E_2,\ldots,E_e) as in Definition 5.12. Let \sigma \geq \tau, then from Proposition A.24 we infer that there is an interval partition \lambda=(L_1,L_2,\ldots, L_l) \in \mathcal{I}_{s} such that

    \sigma=\biggl(\bigcup_{i\in L_1} E_{m_i,n_i},\bigcup_{i\in L_2} E_{m_i,n_i}, \ldots, \bigcup_{i\in L_l} E_{m_i,n_i}\biggr).

    Let (m_i)_{i=1}^{i_1}, (m_i)_{i=i_1+1}^{i_2},\ldots, (m_i)_{i=i_{r-1}+1}^s be the decomposition of (m_i)_{i=1}^s into level runs. This decomposition determines interval blocks B_j:= (i_{j-1}+1,i_{j-1}+2,\ldots, i_j) (i_0=0,i_r=s) and hence gives rise to an interval partition \beta = (B_1,B_2,\ldots,B_r). In order that \bar{\sigma} \leq \bar{\eta}, each L_i connects only plateaux, which is equivalent to the condition that \lambda\leq \beta. Denoting by q_j=\left\lvert B_j \right\rvert=i_j-i_{j-1}, we get the isomorphism (5.20). The ordered set partition \sigma_{\max}^{\operatorname{\mathrm{pla}}}(\tau,\eta) corresponds to the choice \lambda=\beta.

    Proposition

    Proposition 5.21 (Goldberg coefficients)

    0XZF
    1. For \tau,\eta \in \mathcal{OP}_n the Goldberg coefficient (5.3) evaluates to

      g(\tau,\eta) = \begin{cases}\displaystyle \frac{1}{q_1! q_2! \cdots q_r!}\int_{-1}^0 x^{\operatorname{\mathrm{des}}_\eta(\tau)}(1+x)^{\operatorname{\mathrm{asc}}_\eta(\tau)}\prod_{j=1}^r P_{q_j}(x)\,dx, &\bar{\tau} \leq \bar{\eta},\\ 0, & \bar{\tau} \not\leq \bar{\eta}, \end{cases}

      where r, q_1,q_2,\dots, q_r are the integers from Lemma 5.20,

      P_q(x)=\sum_{k=1}^q k! \, S(q,k)\, x^{k-1} = E_q(x,x+1),

      S(q,k) are the Stirling numbers of the second kind and

      E_n(x,y)=\sum_{\sigma\in\mathfrak{S}_n}x^{\operatorname{\mathrm{des}}\sigma}y^{\operatorname{\mathrm{asc}}\sigma}

      are the homogeneous Eulerian polynomials [Reutenauer:1993:freeLiealg, p. 62]. The first few polynomials are

      P_1(x)=1, \quad P_2(x)=2x +1,\quad P_3(x) = 6x^2 + 6x +1,\quad\dots
    2. For \tau,\eta,\pi \in \mathcal{OP}_n the partitioned Goldberg coefficient coefficient (5.3) is

      g(\tau,\eta,\pi) = \begin{cases}\displaystyle \prod_{P\in\pi} g(\tau\rfloor_P,\eta\rfloor_P), &\bar{\tau} \leq \bar{\eta},~\tau \leq \pi,\\ 0, & \text{otherwise}. \end{cases}
    Uses Lemma 5.20 · Used by Proof
    Remark

    Remark 5.22

    FXY5

    The name “Goldberg coefficients” originates from the Campbell-Baker-Hausdorff formula, see Section ? below, in particular Theorem 6.17. Some examples of g(\tau,\eta) will be computed in Example 5.26.

    Remark

    Remark 5.23

    X5AR

    The expansion of P_q(x) in terms of Stirling coefficients was first proved by Frobenius [Comtet:1974:advanced, Theorem E, p. 244].

    Proof of Remark 5.23

    ?   If \bar{\tau} \not\leq \bar{\eta}, then w(\sigma,\eta)=0 for all \sigma \geq \tau and so g(\tau,\eta)=0. Assume hereafter that \bar{\tau} \leq \bar{\eta}. From Lemma 5.20 we have the isomorphism

    \begin{aligned} \{\sigma \in\mathcal{OP}_n \mid \sigma \geq \tau, \bar{\sigma} \leq \bar{\eta} \} &\cong \mathcal{I}_{q_1} \times \mathcal{I}_{q_2} \times \cdots \times \mathcal{I}_{q_r}, \\ \sigma &\mapsto (\sigma_1,\sigma_2,\ldots,\sigma_r). \end{aligned}

    Since r=\left\lvert \tau \right\rvert-\operatorname{\mathrm{pla}}_\eta(\tau)=\operatorname{\mathrm{des}}_\eta(\tau)+\operatorname{\mathrm{asc}}_\eta(\tau)+1, we have \left\lvert \sigma \right\rvert-\operatorname{\mathrm{asc}}_\eta(\tau)-1 = \sum_{i=1}^r (\left\lvert \sigma_i \right\rvert-1)+\operatorname{\mathrm{des}}_\eta(\tau). Note also that [\tau:\sigma]! =\prod_{i=1}^r \prod_{S\in\sigma_i}\left\lvert S \right\rvert!. Since \sigma just connects the blocks of a level run of (m_i)_{i=1}^{\left\lvert \tau \right\rvert}, it does not change the number of ascents: \operatorname{\mathrm{asc}}_\eta(\sigma)=\operatorname{\mathrm{asc}}_\eta(\tau) and we have

    \begin{gathered} \sum_{\substack{\sigma \in \mathcal{OP}_n \\ \sigma \geq \tau}} \widetilde{\zeta}(\tau,\sigma) \, w(\sigma,\eta) \\ \begin{aligned}[t] &=\sum_{\substack{\sigma \in \mathcal{OP}_n \\ \sigma \geq \tau, \bar{\sigma} \leq \bar{\eta}}}\frac{1}{[\tau:\sigma]!} \int_{-1}^0 x^{\left\lvert \sigma \right\rvert -\operatorname{\mathrm{asc}}_\eta(\tau)-1}(1+x)^{\operatorname{\mathrm{asc}}_\eta(\tau)}\,dx \\ &= \sum_{(\sigma_1,\sigma_2,\ldots,\sigma_r) \in\mathcal{I}_{q_1} \times\mathcal{I}_{q_2} \times \cdots \times \mathcal{I}_{q_r}} \int_{-1}^0 x^{\operatorname{\mathrm{des}}_\eta(\tau)}(1+x)^{\operatorname{\mathrm{asc}}_\eta(\tau)}\prod_{i=1}^r\frac{x^{\left\lvert \sigma_i \right\rvert-1}}{\prod_{S\in\sigma_i }\left\lvert S \right\rvert!} dx \\ &= \int_{-1}^0 x^{\operatorname{\mathrm{des}}_\eta(\tau)}(1+x)^{\operatorname{\mathrm{asc}}_\eta(\tau)}\prod_{i=1}^r\left(\sum_{\rho \in \mathcal{I}_{q_i}}\frac{x^{\left\lvert \rho \right\rvert-1}}{\prod_{R\in\rho}\left\lvert R \right\rvert!}\right) dx. \end{aligned} \end{gathered}

    For q\in\mathbb{N} we have

    \sum_{\rho \in \mathcal{I}_q}\frac{x^{\left\lvert \rho \right\rvert-1}}{\prod_{R\in\rho}\left\lvert R \right\rvert!} = \sum_{\substack{n_1+\cdots+n_k=q\\ n_i\geq1, k\geq1}}\frac{x^{k-1}}{n_1! \cdots n_k!}.

    For each fixed k, the sum \sum_{\substack{n_1+\cdots+n_k=q\\ n_i\geq1}}\frac{q!}{n_1! \cdots n_k!} is the number of ways of distributing q distinct objects among k nonempty urns, so it equals k! S(q,k) and the proof is complete.

    ?   The idea of the proof is similar to Proposition 5.19? and we omit the proof.

    Proposition

    Proposition 5.24

    99KC

    Suppose \tau,\eta\in\mathcal{OP}_n are such that \bar{\tau}\leq \bar{\eta}.

    1. If \operatorname{\mathrm{des}}_\eta(\tau)=\operatorname{\mathrm{asc}}_\eta(\tau) and \left\lvert \tau \right\rvert is even then g(\tau,\eta)=0.

    2. If |\tau| is prime then g(\tau,\eta) \neq0.

    Used by Remark 5.25
    Proof of Proposition 5.24

    ? By [Reutenauer:1993:freeLiealg, Corollary 3.15], the coefficient g(\tau,\eta) of K_\tau(X_1,\dots,X_n) vanishes if q_1+\cdots +q_r +\operatorname{\mathrm{des}}_\eta(\tau)+ \operatorname{\mathrm{asc}}_\eta(\tau)-r is odd and \operatorname{\mathrm{des}}_\eta(\tau)=\operatorname{\mathrm{asc}}_\eta(\tau). Since \sum_{i=1}^r q_i = r +\operatorname{\mathrm{pla}}_\eta(\tau) and \left\lvert \tau \right\rvert= \operatorname{\mathrm{des}}_\eta(\tau)+\operatorname{\mathrm{asc}}_\eta(\tau)+\operatorname{\mathrm{pla}}_\eta(\tau)+1, the conclusion follows.

    ? The idea of the proof is taken from [BV15]. Let p:=\left\lvert \tau \right\rvert. By definition and Proposition 5.19 we have

    \begin{aligned} g(\tau,\eta) &=\sum_{\substack{\sigma \in \mathcal{OP}_n \\ \sigma \geq \tau}} \widetilde{\zeta}(\tau,\sigma)\,w(\sigma,\eta) =\sum_{\substack{\sigma \in \mathcal{OP}_n \\ \sigma \geq \tau}} \frac{1}{[\tau:\sigma]!} \frac{(-1)^{\left\lvert \sigma \right\rvert-\operatorname{\mathrm{asc}}_\eta(\sigma)-1}}{\left\lvert \sigma \right\rvert\binom{\left\lvert \sigma \right\rvert-1}{\operatorname{\mathrm{asc}}_\eta(\sigma)}} \\ &= \frac{(-1)^{p-\operatorname{\mathrm{asc}}_\eta(\tau)-1}}{p\binom{p-1}{\operatorname{\mathrm{asc}}_\eta(\tau)}} + \sum_{\substack{\sigma \in \mathcal{OP}_n \\ \sigma > \tau}} \frac{1}{[\tau:\sigma]!} \frac{(-1)^{\left\lvert \sigma \right\rvert-\operatorname{\mathrm{asc}}_\eta(\sigma)-1}}{\left\lvert \sigma \right\rvert\binom{\left\lvert \sigma \right\rvert-1}{\operatorname{\mathrm{asc}}_\eta(\sigma)}}. \end{aligned}

    Since the number [\tau:\sigma]! \left\lvert \sigma \right\rvert\binom{\left\lvert \sigma \right\rvert-1}{\operatorname{\mathrm{asc}}_\eta(\sigma)} never contains p as a factor for any \sigma >\tau, g(\tau,\eta) is nonzero.

    Remark

    Remark 5.25

    WJ6F

    It is a difficult problem to characterize vanishing Goldberg coefficients. The criterion ? from Proposition 5.24 above does not cover all cases, for example one can show that g(\tau,\eta)=0 for \tau=(\{3\},\{4\},\{2\},\{1\}), \eta=(\{1,2,3\},\{4\}), although the pair (\tau,\eta) does not satisfy the assumption of the criterion. More information about Goldberg coefficients can be found in [Reutenauer:1993:freeLiealg, Thompson82] and in particular [BV15, Section IV] concerning the question of vanishing coefficients.

    Example

    Example 5.26 (Goldberg coefficients and partial vanishing of cumulants)

    ZHWL

    We will write the ordered kernel set partition \kappa(i_1,\dots, i_n) simply as i_1i_2\dots i_n.

    1. Take \eta= 12. We compute g(12,12). Now \tau=\eta so m_1=1, m_2=2, r=2, q_1=q_2=1. Hence \operatorname{\mathrm{des}}_\eta(\tau)=0, \operatorname{\mathrm{asc}}_\eta(\tau)=1 and so

      g(12,12)= \frac{1}{1!1!}\int_{-1}^0 (1+x) \,dx=\frac{1}{2}.

      Similarly we get g(21,12)=-\frac{1}{2}, g(12,21)=-\frac{1}{2}, g(21,21)=\frac{1}{2} and hence

      \begin{aligned} &K_{11}(X^{(1)},Y^{(2)}) = \frac{1}{2}K_{12}(X,Y)-\frac{1}{2} K_{21}(X,Y), \\ & K_{11}(X^{(2)}, Y^{(1)}) = -\frac{1}{2}K_{12}(X,Y)+\frac{1}{2} K_{21}(X,Y). \end{aligned}

      In the most popular spreadability systems, like the tensor, free, Boolean or monotone spreadability systems, partitioned cumulants factorize, e.g., K_{12}(X,Y)=K_1(X)K_1(Y). Hence we get K_{11}(X^{(1)},Y^{(2)}) =K_{11}(X^{(2)},Y^{(1)})=0.

    2. We then consider the case \eta=112= (\{1,2\},\{3\}). The Goldberg coefficients can be nonzero only when \bar{\tau} \leq \bar{\eta}, so \tau is one of

      112, 221, 123, 132, 213, 231, 312, 321.

      If \tau=112 then g(112,112)=\frac{1}{2} by the same calculation as g(12,12). If \tau=221 then again g(221,112)= -\frac{1}{2}. If \tau=123 then m_1=1,m_2=1,m_3=2, r=2, q_1=2, q_2=1. So

      g(123,112) = \frac{1}{1!2!}\int_{-1}^0 (1+x) P_2(x) \, dx = \frac{1}{12}.

      If we take \tau= 132 then m_1=1,m_2=2, m_3=1, so r=3,q_1=q_2=q_3=1, \operatorname{\mathrm{asc}} = 1, \operatorname{\mathrm{des}}=1. Thus we get

      g(132,112) = \frac{1}{1!1!1!}\int_{-1}^0 x (1+x) \, dx = -\frac{1}{6}.

      If we take \tau= 231 then m_1=2,m_2=1,m_3=1. So r=2, q_1=1, q_2=2, \operatorname{\mathrm{asc}}=0,\operatorname{\mathrm{des}}=1. Hence

      g(231,112) = \frac{1}{1!2!}\int_{-1}^0 x P_2(x) \, dx = \frac{1}{2}\int_{-1}^0 x (2x+1)\, dx=\frac{1}{12}.

      Similarly we can compute the remaining Goldberg coefficients and get

      \begin{gathered} K_{111}(X^{(1)},Y^{(1)},Z^{(2)}) = \frac{1}{2} K_{112} -\frac{1}{2} K_{221} + \frac{1}{12} K_{123} -\frac{1}{6} K_{132} \\ +\frac{1}{12}K_{213}+\frac{1}{12} K_{231} -\frac{1}{6}K_{312} + \frac{1}{12}K_{321}, \end{gathered}

      where X,Y,Z are omitted for simplicity. We can see that (5.8) holds (now \pi=111):

      \frac{1}{2} - \frac{1}{2}=0, \qquad \frac{1}{12} - \frac{1}{6} + \frac{1}{12} + \frac{1}{12} - \frac{1}{6} + \frac{1}{12} =0.

      Again if the cumulants factorize then K_{112}=K_{221}=K_{11}(X,Y)K_1(Z) and K_{123}=\cdots= K_{321} = K_1(X)K_1(Y)K_1(Z), so the mixed cumulant K_{111}(X^{(1)},Y^{(1)},Z^{(2)}) vanishes.

      Similarly one can compute g(\tau, 121) for all \tau such that \bar{\tau} \leq \overline{121} and get

      \begin{gathered} K_{111}(X^{(1)},Y^{(2)},Z^{(1)}) = \frac{1}{2} K_{121} -\frac{1}{2} K_{212} - \frac{1}{6} K_{123} +\frac{1}{12} K_{132} \\ +\frac{1}{12}K_{213}+\frac{1}{12} K_{231} +\frac{1}{12}K_{312} - \frac{1}{6}K_{321}. \end{gathered}

      Now in the tensor, free or Boolean spreadability system the mixed cumulant vanishes. However in the monotone spreadability system it does not: K_{212} vanishes identically since 212 is not a monotone partition (see Proposition 3.24) and therefore, in the monotone case we have

      K_{111}(X^{(1)},Y^{(2)},Z^{(1)}) = \frac{1}{2}K_{121}(X,Y,Z)=\frac{1}{2} K_{11}(X,Z)K_1(Y)

      which does not vanish in general.

    Example

    Example 5.27 (Weisner coefficients and partial vanishing of cumulants)

    DXRL

    We can reuse some results from Example 5.26.

    1. If we take \tau=\eta=12 then \operatorname{\mathrm{asc}}_\eta(\tau)=1 and so

      w(12,12)= \frac{(-1)^{2-1-1}}{2 \binom{1}{1}}=\frac{1}{2}.
    2. Similarly if \tau=21 and \eta=12 then m_1=2, m_1=1 so \operatorname{\mathrm{asc}}_\eta(\tau)=0. Therefore we get w(21,12)=-\frac{1}{2}. Similarly, w(12,21)=-\frac{1}{2}, w(21,21)=\frac{1}{2}. So

      \begin{aligned} K_{11}(X^{(1)},Y^{(2)}) &= \frac{1}{2}\varphi_{12}(X,Y)-\frac{1}{2} \varphi_{21}(X,Y), \\ K_{11}(X^{(2)}, Y^{(1)}) &= -\frac{1}{2}\varphi_{12}(X,Y)+\frac{1}{2} \varphi_{21}(X,Y). \end{aligned}

      In factorizing spreadability systems we have \varphi_{12}(X,Y)=\varphi_{21}(X,Y)=\varphi_1(X)\,\varphi_1(Y) and hence K_{11}(X^{(1)},Y^{(2)}) =K_{11}(X^{(2)},Y^{(1)})=0.

    3. One can show that w(112,112)=\frac{1}{2} = - w(221,112) by the same calculation as w(12,12) and w(21,12). If \tau=123,\eta=112 then m_1=1,m_2=1,m_3=2, so \operatorname{\mathrm{asc}} =1 and

      w(123,112) = \frac{(-1)^{3-1-1}}{3\binom{2}{1}} = -\frac{1}{6}.

      If we take \tau= 231 then m_1=2,m_2=1,m_3=1 and so \operatorname{\mathrm{asc}}=0. Hence

      w(231,112) = \frac{(-1)^{3-0-1}}{3\binom{2}{0}}=\frac{1}{3}.

      Similarly we can compute the remaining Weisner coefficients and get

      \begin{gathered} K_{111}(X^{(1)},Y^{(1)},Z^{(2)}) = \frac{1}{2} \varphi_{112} -\frac{1}{2} \varphi_{221} - \frac{1}{6} \varphi_{123} -\frac{1}{6} \varphi_{132} \\ -\frac{1}{6}\varphi_{213}+\frac{1}{3} \varphi_{231} -\frac{1}{6}\varphi_{312} + \frac{1}{3}\varphi_{321}, \end{gathered}

      where X,Y,Z are omitted for simplicity. We can see that (5.8) holds (now \pi=111):

      \frac{1}{2} - \frac{1}{2}=0, \qquad -\frac{1}{6} - \frac{1}{6} - \frac{1}{6} + \frac{1}{3} - \frac{1}{6} + \frac{1}{3} =0.

      In factorizing spreadability systems the mixed cumulant K_{111}(X^{(1)},Y^{(1)},Z^{(2)}) vanishes by using the factorization of partitioned moments.

    4. Similarly one can compute w(\tau, 121) for all \tau such that \bar{\tau} \leq \overline{121} and get

      \begin{gathered} K_{111}(X^{(1)},Y^{(2)},Z^{(1)}) = \frac{1}{2} \varphi_{121} -\frac{1}{2} \varphi_{212} - \frac{1}{6} \varphi_{123} -\frac{1}{6} \varphi_{132} \\ +\frac{1}{3}\varphi_{213}-\frac{1}{6} \varphi_{231} +\frac{1}{3}\varphi_{312} - \frac{1}{6}\varphi_{321}. \end{gathered}

      In the tensor, free or Boolean spreadability system mixed cumulants vanishes, but in the monotone case in general they don’t: indeed \varphi_{212}=\varphi(X)\varphi(Y)\varphi(Z) while \varphi_{121}=\varphi(X Z)\varphi(Y). Therefore, in the monotone case we have

      K_{111}(X^{(1)},Y^{(2)},Z^{(1)}) = \frac{1}{2}(\varphi_{121}-\varphi_{212})=\frac{1}{2}(\varphi(X Z)- \varphi(X)\varphi(Z)) \varphi(Y)

      which does not vanish in general.

    Proposition

    Proposition 6.1

    E73H

    The triple \mathcal{S}_{\mathrm{\shuffle}}=\mathcal{S}_{\mathrm{\shuffle}}(\mathcal{A})=(\mathcal{U}, \operatorname{\mathrm{conc}}_\infty,(\iota^{(j)})_{j=1}^\infty) defined above constitutes a spreadability system for (\mathcal{A},\operatorname{\mathrm{Id}}).

    Proof of Proposition 6.1

    Clearly \tilde{\varphi}\circ \iota^{(i)}=\operatorname{\mathrm{Id}} on \mathcal{A}. The symmetry condition (?) holds too since the value of \tilde{\varphi} only depends on the ordered kernel set partition of the upper indices.

    Proposition

    Proposition 6.2 (Multiplicativity of partitioned cumulants)

    SRCO

    For \pi = (P_1,P_2,\dots, P_p) \in \mathcal{OP}_n, we have

    K_\pi^{\mathrm{\shuffle}}(X_1,X_2,\dots, X_n) = K_{|P_1|}^{\mathrm{\shuffle}}(X_{P_1})\, K_{|P_2|}^{\mathrm{\shuffle}}(X_{P_2}) \dotsm K_{|P_p|}^{\mathrm{\shuffle}}(X_{P_p}).
    Proof of Proposition 6.2

    For \pi=(P_1,P_2,\dots, P_p)\in\mathcal{OP}_n,

    \begin{gathered} \varphi_\pi(N.X_1,N.X_2, \dots, N.X_n) \\ \begin{aligned} &= \sum_{i_1, i_2,\dots, i_n \in [N]} \varphi_{\pi\curlywedge \kappa(i_1,i_2,\dots, i_n)}(X_1,X_2, \dots, X_n) \\ &= \sum_{i_1,i_2, \dots, i_n \in [N]} X_{\kappa(i_1,i_2,\dots, i_n) \rfloor_{P_1}} X_{\kappa(i_1,i_2,\dots, i_n) \rfloor_{P_2}} \cdots X_{\kappa(i_1,i_2,\dots, i_n) \rfloor_{P_p}}\\ &= \biggl( \sum_{\substack{i_k \in [N] \\ k\in P_1}} X_{\kappa(i_1,i_2,\dots, i_n) \rfloor_{P_1}} \biggr) \biggl( \sum_{\substack{i_k \in [N] \\ k\in P_2}} X_{\kappa(i_1,i_2,\dots, i_n) \rfloor_{P_2}} \biggr) \,\cdots\, \biggl( \sum_{\substack{i_k \in [N] \\ k\in P_p}} X_{\kappa(i_1,i_2,\dots, i_n) \rfloor_{P_p}} \biggr) \\ &= \tilde{\varphi}\biggl(\prod_{i\in P_1}N.X_i\biggr)\, \tilde{\varphi}\biggl(\prod_{i\in P_2}N.X_i\biggr)\, \cdots\, \tilde{\varphi}\biggl(\prod_{i\in P_p}N.X_i\biggr). \end{aligned} \end{gathered}

    We conclude by observing that the coefficient of N^{p} is the product of the linear coefficients of the factors.

    Proposition

    Proposition 6.3

    58G2

    For n \in \mathbb{N} we have

    K_n^{\mathrm{\shuffle}} (X_1,X_2,\dots, X_n) = \sum_{\pi =(P_1,P_2,\dots,P_k)\in \mathcal{OP}_n} \frac{(-1)^{|\pi|-1}}{|\pi|}X_{P_1} X_{P_2}\cdots X_{P_k}.
    Example

    Example 6.4

    26DE

    The reader can easily verify that

    \begin{aligned} K_1^{\mathrm{\shuffle}}(X)&=X, \\ K_2^{\mathrm{\shuffle}}(X_1,X_2) &= \frac{1}{2}[X_1,X_2], \\ K_3^{\mathrm{\shuffle}}(X_1,X_2,X_3) &= \frac{1}{3} (X_1 X_2 X_3+X_3 X_2 X_1) \\ &\qquad- \frac{1}{6} (X_1 X_3 X_2 + X_2 X_1 X_3 + X_2 X_3 X_1 + X_3 X_1 X_2). \end{aligned}
    Remark

    Remark 6.5

    E9BE

    We will see in Section ? that K_n^{\mathrm{\shuffle}}~(n\geq2) can be expressed as a sum of commutators, i.e., a Lie polynomial.

    Theorem

    Theorem 6.6 (CBH formula)

    GRJ9

    As formal power series on \mathcal{A} we have the identity

    \begin{gathered} \log(e^{a_1} e^{a_2}\cdots e^{a_n}) \\ = \sum \frac{1}{p_1!p_2! \cdots p_n!} K^{\mathrm{\shuffle}}_{p_1+p_2+\cdots+p_n}(\underbrace{a_{1}, a_{1},\ldots, a_{1}}_{\text{$p_1$ times}}, \underbrace{a_{2}, a_{2},\ldots, a_{2}}_{\text{$p_2$ times}}, \ldots, \underbrace{a_{n},a_n,\ldots,a_{n}}_{\text{$p_n$ times}}). \end{gathered}

    where the sum runs over all n-tuples (p_1,p_2,\dots,p_n) \in (\mathbb{N}\cup\{0\})^n with the exception of the tuple (0,0,\dots,0).

    Used by Proof
    Proof of Theorem 6.6

    First observe that the right hand side of the claimed identity is the coefficient of N in the series

    \begin{gathered} 1+ \sum_{\substack{(p_1,p_2,\dots,p_n) \in (\mathbb{N}\cup \{0\})^n,\\(p_1,p_2,\dots,p_n) \neq (0,0,\dots,0) }} \frac{1}{p_1!p_2! \cdots p_n!}\tilde{\varphi}((N.a_{1})^{p_1}(N.a_2)^{p_2}\cdots\,(N.a_{n})^{p_n})\\ \begin{aligned} &= \tilde{\varphi}(e^{N.a_1}e^{N.a_2} \cdots \, e^{N.a_n}) \\ &= \tilde{\varphi} (e^{a_1^{(1)}+a_1^{(2)}+\cdots + a_1^{(N)}} e^{a_2^{(1)}+a_2^{(2)}+\cdots + a_2^{(N)}} \cdots\, e^{a_n^{(1)}+a_n^{(2)}+\cdots + a_n^{(N)}}). \end{aligned} \end{gathered}

    Our construction implies that a_i^{(j)}, j=1,2,3,\dots mutually commute for different j and so

    e^{a_i^{(1)}+a_i^{(2)}+\cdots + a_i^{(N)}} = e^{a_i^{(1)}} e^{a_i^{(2)}}\cdots e^{a_i^{(N)}} .

    Thus (Proof) can be factorized as follows:

    \begin{gathered} \tilde{\varphi}(e^{a_1^{(1)}+a_1^{(2)}+\cdots + a_1^{(N)}} e^{a_2^{(1)}+a_2^{(2)}+\cdots + a_2^{(N)}} \cdots e^{a_n^{(1)}+a_n^{(2)}+\cdots + a_n^{(N)}}) \\ = \tilde{\varphi}((e^{a_1^{(1)}}e^{a_1^{(2)}}\cdots e^{a_1^{(N)}})( e^{a_2^{(1)}} e^{a_2^{(2)}}\cdots e^{a_2^{(N)}}) \cdots ( e^{a_n^{(1)}} e^{a_n^{(2)}}\cdots e^{a_n^{(N)}}) ). \end{gathered}

    Now \tilde{\varphi} simply rearranges the factors according to the upper index and the last expression equals

    (e^{a_1} e^{a_2}\cdots e^{a_n})^N.

    On the other hand,

    \begin{aligned} (e^{a_1} e^{a_2}\cdots e^{a_n})^N &= e^{N\log(e^{a_1} e^{a_2}\cdots e^{a_n})} \\ &= 1 + N \log(e^{a_1} e^{a_2}\cdots e^{a_n}) + \frac{N^2}{2!}( \log(e^{a_1} e^{a_2}\cdots e^{a_n}))^2 + \dots \end{aligned}

    and we conclude by comparing the coefficient of N in this series with the one in (Proof).

    Remark

    Remark 6.7

    EJJV

    Theorem 6.6 is similar to the well-known formula for the generating function of multivariate cumulants from classical probability theory,

    \begin{gathered} \log \operatorname{\mathbf{E}}[e^{z_1 X_1 +z_2 X_2 +\cdots +z_n X_n}] \\ = \sum \frac{z_1^{p_1}z_2^{p_2}\cdots z_n^{p_n}}{p_1!p_2! \cdots p_n!} K^{\mathrm{T}}_{p_1+p_2+\cdots+p_n}( \underbrace{X_{1}, X_{1},\ldots, X_{1}}_{\text{$p_1$ times}}, \underbrace{X_{2}, X_{2},\ldots, X_{2}}_{\text{$p_2$ times}}, \ldots, \underbrace{X_{n},X_n,\ldots,X_{n}}_{\text{$p_n$ times}}), \end{gathered}

    where X_1,X_2,\dots,X_n are \mathbb{C}-valued classical random variables and z_1,z_2,\dots, z_n are commuting indeterminates.

    Proposition

    Proposition 6.8

    XAKG
    1. A sequence (\mathcal{A}_i)_{i=1}^\infty of subalgebras of \mathcal{A} is \mathcal{S}_{\mathrm{\shuffle}}-independent if and only if the subalgebras \mathcal{A}_1, \mathcal{A}_2, \dots commute mutually. This is obviously equivalent to K_2^{\mathrm{\shuffle}}(X,Y)=0 whenever X \in \mathcal{A}_i, Y\in \mathcal{A}_j with i\neq j.

    2. For fixed n \geq2, if \{X_1,X_2,\dots, X_n\} \subseteq \mathcal{A} splits into two mutually commuting families then

      K_n^{\mathrm{\shuffle}}(X_1,X_2,\dots, X_n)=0 .
    Used by Remark 6.16
    Proof of Proposition 6.8

    ?   Suppose that (\mathcal{A}_i)_{i=1}^\infty is \mathcal{S}_{\mathrm{\shuffle}}-independent. Let \pi =\hat{1}_2, \rho = (\{2\},\{1\}) \in\mathcal{OP}_2 and let X \in \mathcal{A}_i, Y \in \mathcal{A}_j for fixed i > j. Then \kappa(i,j)=\rho and

    \varphi_\pi(X,Y) = X Y, \qquad \varphi_{\pi\curlywedge \rho}(X,Y) = Y X,

    and by \mathcal{S}_{\mathrm{\shuffle}}-independence these two must coincide, so X Y = Y X. This shows that [\mathcal{A}_i,\mathcal{A}_j]=0. Conversely, if the subalgebras \mathcal{A}_1, \mathcal{A}_2, \dots mutually commute then for any i_1,i_2,\dots, i_n \in \mathbb{N}, any X_k \in \mathcal{A}_{i_k}, k=1,2,\dots, n and any \pi=(P_1,P_2,\dots, P_p) \in\mathcal{OP}_n we have

    \varphi_{\pi\curlywedge \rho}(X_1,X_2,\dots, X_n) = (X_{P_1 \cap R_1} X_{P_1 \cap R_2} \cdots X_{P_1 \cap R_r} ) \cdots (X_{P_p \cap R_1} X_{P_p \cap R_2} \cdots X_{P_p \cap R_r})

    where \rho=\kappa(i_1,i_2,\dots, i_n)=(R_1,R_2,\dots, R_r). Here X_\emptyset is understood as the unit \hat{1}_\mathcal{A}. Now if \{X_k \mid k \in R_i\} and \{X_k \mid k \in R_j\} mutually commute for distinct i,j then for each i =1,2,\dots, p we have

    X_{P_i \cap R_1} X_{P_i \cap R_2} \cdots X_{P_i \cap R_r} = X_{P_i},

    which shows that \varphi_{\pi\curlywedge \rho}(X_1,X_2,\dots, X_n) =\varphi_{\pi}(X_1,X_2\dots, X_n).

    ?   Suppose that \{X_i \mid i\in I\} and \{X_i \mid i\in I^c\} commute with each other and \emptyset \subsetneq I \subsetneq \{1,2,\dots,n\}. Then, for commuting indeterminates z_1,\dots, z_n, we have

    \log(e^{z_1 X_1} e^{z_2 X_2} \cdots e^{z_n X_n}) = \log\Bigl(\prod_{i \in I} e^{z_ i X_i}\Bigr) + \log\Bigl(\prod_{i \in I^c} e^{z_ iX_i}\Bigr),

    where the products \prod_{i \in I}, \prod_{i \in I^c} preserve the natural orders on I,I^c, respectively. On the other hand, by Theorem 6.6 we have

    \begin{gathered} \log(e^{z_1 X_1}e^{z_2 X_2} \cdots e^{z_n X_n}) \\ = \sum \frac{z_1^{p_1} z_2^{p_2} \cdots z_n^{p_n}}{p_1!p_2! \cdots p_n!} K^{\mathrm{\shuffle}}_{p_1+p_2+\cdots+p_n}( \underbrace{X_{1}, X_{1},\ldots, X_{1}}_{\text{$p_1$ times}}, \underbrace{X_{2}, X_{2},\ldots, X_{2}}_{\text{$p_2$ times}}, \ldots, \underbrace{X_{n},X_n,\ldots,X_{n}}_{\text{$p_n$ times}}). \end{gathered}

    Comparing the coefficients of z_1z_2\cdots z_n in (?) with those of (?) we conclude that K_n^{\mathrm{\shuffle}}(X_1,X_2,\dots, X_n)=0.

    Remark

    Remark 6.9

    BE9I
    1. Additivity of Lie polynomials in commuting variables is well known, see [Reutenauer:1993:freeLiealg, p. 20].

    2. We have thus shown that

      \text{$\mathcal{S}_{\mathrm{\shuffle}}$-independence $\iff$ commutativity $\iff$ vanishing of mixed cumulants}.

      This example illustrates that in general vanishing of mixed cumulants does not imply exchangeability.

    Definition

    Definition 6.10

    7VNQ

    The specialization \mathcal{S}_{\mathrm{\shuffle}} (\mathbb{C}\langle{\mathbf{A}}\rangle) of the unshuffle spreadability system from Section ? for the ncps (\mathbb{C}\langle{\mathbf{A}}\rangle,\text{Id}) is called the free Lie spreadability system and denoted by \mathcal{S}_{\mathrm{FL}}({\mathbf{A}}).

    Definition

    Definition 6.11

    DH4C

    Let \Pi: \mathbb{C}\langle{\mathbf{A}}\rangle \to \mathbb{C}\langle{\mathbf{A}}\rangle be the map defined by

    \Pi=\sum_{k=1}^\infty \frac{(-1)^{k-1}}{k} (\operatorname{\mathrm{Id}}-\epsilon)^{\ast k},

    where \epsilon is regarded as an endomorphism of \mathbb{C}\langle{\mathbf{A}}\rangle.

    Proposition

    Proposition 6.12

    YL1W

    Let a_1,a_2,\ldots,a_n \in {\mathbf{A}}. Then

    \Pi(a_1a_2\cdots a_n) = \sum_{\pi = (P_1, P_2, \dots) \in \mathcal{OP}_n} \frac{(-1)^{\left\lvert \pi \right\rvert-1}}{\left\lvert \pi \right\rvert} a_{P_1} a_{P_2}\cdots a_{P_{\left\lvert \pi \right\rvert}}.
    Proof of Proposition 6.12

    From (?) we have

    \begin{aligned} \Pi(a_1a_2\cdots a_n) &= \sum_{k=1}^\infty \frac{(-1)^{k-1}}{k} (\operatorname{\mathrm{Id}}-\epsilon)^{\ast k}(a_1a_2\cdots a_n)\\ &=\sum_{k=1}^\infty \frac{(-1)^{k-1}}{k} (\operatorname{\mathrm{conc}}_k \circ (\operatorname{\mathrm{Id}}-\epsilon)^{\otimes k}\circ \delta_k)(a_1a_2\cdots a_n) \\ &=\sum_{k=1}^\infty \frac{(-1)^{k-1}}{k} \sum_{(P_1,P_2,\dots,P_k) \in \mathcal{OPP}_n} \!\!(\operatorname{\mathrm{Id}}-\epsilon)(a_{P_1}) (\operatorname{\mathrm{Id}}-\epsilon)(a_{P_2}) \cdots (\operatorname{\mathrm{Id}}-\epsilon)(a_{P_k}). \end{aligned}

    Note that (\operatorname{\mathrm{Id}}-\epsilon)(1)=0 and (\operatorname{\mathrm{Id}}-\epsilon)(w)=w for w\in{\mathbf{A}}^\ast\setminus\{1\}. If P_i=\emptyset for some i, then the product (\operatorname{\mathrm{Id}}-\epsilon)(a_{P_1}) (\operatorname{\mathrm{Id}}-\epsilon)(a_{P_2}) \cdots (\operatorname{\mathrm{Id}}-\epsilon)(a_{P_k}) vanishes. Therefore only proper ordered set partitions contribute to the sum, and in particular the sum extends over k \leq n only. This proves the claim.

    Theorem

    Theorem 6.13

    GLKH

    Let a_1,a_2,\ldots,a_n \in {\mathbf{A}} and let K^{\mathrm{FL}}_\pi be the cumulants associated to the spreadability system \mathcal{S}_{\mathrm{FL}}. Then

    \Pi(a_1a_2\cdots a_n) = K^{\mathrm{FL}}_n(a_1,a_2,\dots,a_n).
    Remark

    Remark 6.14

    FVAK
    1. The space of Lie polynomials \mathcal{L}({\mathbf{A}}) is the smallest subspace of \mathbb{C}\langle{\mathbf{A}}\rangle that contains {\mathbf{A}} and is closed with respect to the Lie bracket [X,Y]=XY-YX. It is well known that \Pi is a Lie projector, i.e., \mathcal{L}({\mathbf{A}})=\Pi(\mathbb{C}\langle{\mathbf{A}}\rangle) [Reutenauer:1993:freeLiealg, Theorem 3.7]. So in the above setting cumulants with entries from {\mathbf{A}} are exactly Lie polynomials.

    2. The values at words of higher order are given by convolution powers of \Pi, i.e.,

      \Pi_k=\frac{1}{k!}\Pi^{\ast k},\qquad k\in\mathbb{N}

      (see [Reutenauer:1993:freeLiealg]), and we can show that

      \Pi_k(a_1a_2\cdots a_n) =\frac{1}{k!}\sum_{\substack{\pi\in \mathcal{OP}_n\\\left\lvert \pi \right\rvert=k}} K^{\mathrm{FL}}_\pi(a_1,a_2,\dots,a_n),\qquad a_i\in{\mathbf{A}},~ i\in[n].

      is the coefficient of N^{k} appearing in \tilde{\varphi}((N.a_1) (N.a_2)\cdots (N.a_n)) by Theorem 3.11 and Proposition 3.16. In other words,

      \tilde{\varphi}((N.a_1) (N.a_2)\cdots (N.a_n))= \sum_{k=1}^n N^k \Pi_k(a_1a_2\cdots a_n),\qquad a_i\in{\mathbf{A}},~ i\in[n].
    Lemma

    Lemma 6.15

    H7L8

    Let \eta=(E_1,E_2,\dots,E_e) be an interval partition of [n] where the blocks are in canonical order, i.e. if s<t then i<j for all i\in E_s, j\in E_t. Then for any X_i \in \mathbb{C}\langle{\mathbf{A}}\rangle and any \pi\in \mathcal{OP}_n, we have

    \varphi_{\pi \curlywedge \eta}(X_1,X_2,\dots,X_n)=\varphi_{\pi}(X_1,X_2,\dots,X_n).
    Used by Proof
    Proof of Lemma 6.15

    The statement holds by definition because every block P\in\pi is the concatenation of P\cap E_1, P\cap E_2, \dots, P\cap E_e.

    Remark

    Remark 6.16

    W667

    Thus a sequence of distinct letters satisfies some partial independence, but it is not \mathcal{S}_{\mathrm{FL}}-independent. Indeed, (a_1,a_2)\in {\mathbf{A}} \times {\mathbf{A}} is \mathcal{S}_{\mathrm{FL}}-independent if and only if a_1=a_2 by Proposition 6.8.

    Theorem

    Theorem 6.17

    2655

    Let a_1,a_2,\ldots,a_n be distinct letters from the alphabet {\mathbf{A}}, let r\in\mathbb{N} and q_j,i_j \in \mathbb{N} for j\in [r] such that i_j \neq i_{j+1} for j\in[r-1]. Then the coefficient of the monomial a_{i_1}^{q_1}a_{i_2}^{q_2}\cdots a_{i_r}^{q_r} appearing in \log(e^{a_1}\cdots e^{a_n}) is given by

    \frac{1}{q_1!q_2!\cdots q_r!}\int_{-1}^0 x^{\operatorname{\mathrm{des}}(\underline{i})} (1+x)^{\operatorname{\mathrm{asc}}(\underline{i})} \prod_{j=1}^r P_{q_j}(x)\,dx,

    where P_q(x) are the homogeneous Euler polynomials already encountered in Proposition 5.21.

    Proof of Theorem 6.17

    In order for a monomial v=a_{i_1}^{q_1}a_{i_2}^{q_2}\cdots a_{i_r}^{q_r} to occur as a term in \Pi(a_1^{p_1}a_2^{p_2}\cdots a_n^{p_n}) it necessarily has to be a rearrangement of the word (= permutation of the multiset) w=a_1^{p_1}a_2^{p_2}\dotsm a_n^{p_n}, since the projector in the CBH formula (?) does not change multiplicities and therefore every letter must occur the same number of times in v and w.

    Thus p_k=\sum_{j\in B_k}q_j where B_k=\{j : i_j=k\}. Let p=p_1+p_2+\dots+p_n be the total length of w and \eta=(A_1,A_2,\dots,A_n)\in \mathcal{OI}_p be the ordered interval partition corresponding to the composition (p_1,p_2,\dots,p_n), i.e., A_j=\{p_1+p_2+\dots+p_{j-1}+1, p_1+p_2+\dots+p_{j-1}+2,\dots,p_1+p_2+\dots+p_j\} and \left\lvert A_j \right\rvert=p_j.

    From Theorem 6.13 we infer

    \begin{aligned} \Pi(a_{1}^{p_1}a_{2}^{p_2} \cdots a_{n}^{p_n}) &= K^{\mathrm{FL}}_{p}(\underbrace{a_1, a_1,\ldots, a_1}_{\text{$p_1$ times}}, \underbrace{a_2, a_2,\ldots, a_2}_{\text{$p_2$ times}}, \ldots, \underbrace{a_{n},a_{n},\ldots,a_{n}}_{\text{$p_n$ times}}) \\ & =:K_p^{\mathrm{FL}}(a_1^{\dots p_1},a_2^{\dots p_2}, \dots, a_n^{\dots p_n}). \end{aligned}

    Note that by Lemma 6.15 we have

    \varphi_{\pi}(a_1^{\dots p_1},a_2^{\dots p_2}, \dots, a_n^{\dots p_n}) = \varphi_{\pi \curlywedge \eta}(a_1^{\dots p_1},a_2^{\dots p_2}, \dots,a_n^{\dots p_n})

    for any \pi \in \mathcal{OP}_p and thus

    \begin{aligned} K^{\mathrm{FL}}_{p}(a_1^{\dots p_1},a_2^{\dots p_2}, \dots, a_n^{\dots p_n}) &= \sum_{\sigma \in \mathcal{OP}_p} \varphi_\sigma(a_1^{\dots p_1},a_2^{\dots p_2}, \dots, a_n^{\dots p_n}) \, \widetilde{\mu}(\sigma, \hat{1}_p) \\ &= \sum_{\sigma \in \mathcal{OP}_p} \varphi_{\sigma\curlywedge\eta} (a_1^{\dots p_1},a_2^{\dots p_2}, \dots, a_n^{\dots p_n}) \, \widetilde{\mu}(\sigma, \hat{1}_p)\\ &= \sum_{\substack{\tau \in \mathcal{OP}_p \\ \bar{\tau}\leq \bar{\eta}}} \varphi_{\tau}(a_1^{\dots p_1},a_2^{\dots p_2}, \dots, a_n^{\dots p_n}) \,w(\tau, \eta) \end{aligned}

    by Proposition 5.19. Thus in order to determine the coefficient of v we must collect all ordered set partitions \tau such that

    T_v =\{\tau\in\mathcal{OP}_p : \bar{\tau}\leq\bar{\eta}, \varphi_{\tau}(a_1^{\dots p_1},a_2^{\dots p_2}, \dots, a_n^{\dots p_n}) = v \}

    and then sum up the corresponding values of w(\tau,\eta). Let us now investigate the structure of this set. First note that T_v is an order ideal: if \tau\in T_v and \tau'\leq \tau, then \tau'\in T_v, because every block of \tau contains only repetitions of one letter and further refinement of \tau leaves the end result v invariant. Moreover T_v is the disjoint union of the principal ideals \operatorname{\downarrow\!}\tau_0=\{\tau : \tau\leq \tau_0\} where \tau_0\in T_v is maximal.

    The maximal partition \tau_0 arises as follows: After application of \varphi_\tau for \tau\in T_v each factor a_j^{p_j} is divided into pieces a_j^{q_l}, l\in B_j and the number of such subdivisions is the multinomial coefficient

    \binom{p_j}{q_l : l\in B_j} .

    These subdivisions are in one-to-one correspondence with ordered set partitions of A_j which, pieced together in the order of j, give rise to a maximal ordered set partition \tau_0\in T_v. Thus in total there are as many maximal ordered set partitions as there are subdivisions, namely

    \binom{p_1}{q_l:l\in B_1} \binom{p_2}{q_l:l\in B_2} \dotsm \binom{p_n}{q_l:l\in B_n} = \frac{p_1!p_2!\cdots p_n!}{q_1!q_2!\cdots q_r!} .

    By Proposition A.23 the principal ideal \operatorname{\downarrow\!}\tau_0 generated by a maximal ordered set partition \tau_0 is isomorphic to

    \mathcal{OP}_{q_1}\times\mathcal{OP}_{q_2}\times\dotsm\mathcal{OP}_{q_r}

    and in particular, all principal ideals are isomorphic. Moreover the number of ascents is \operatorname{\mathrm{asc}}_\eta(\tau)=\operatorname{\mathrm{asc}}(\underline{i}) and does not depend on the choice of \tau\in T_v. Thus all ideals \operatorname{\downarrow\!}\tau_0 deliver the same contribution and as a consequence of the discussion above we are left with

    \sum_{\tau\in T_v} w(\tau,\eta) = \frac{p_1!p_2!\cdots p_n!}{q_1!q_2!\cdots q_r!} \sum_{\tau\leq\tau_0} w(\tau,\eta)

    for one fixed maximal element \tau_0\in T_v.

    A canonical representative \tau_0 is obtained by concatenating consecutive subintervals of length q_j from A_{i_j}, j=1,2,\dots,r. Let us now turn to the value of w(\tau,\eta). As seen above, the numbers of both ascents and descents of \tau only depend on those of the sequence \underline{i} and since \underline{i} has no plateaux we infer from (5.11) that r=\operatorname{\mathrm{des}}(\underline{i}) + \operatorname{\mathrm{asc}}(\underline{i})+1. On the other hand, if we denote by (\tau_1,\tau_2,\dots,\tau_r) the image of \tau\in\operatorname{\downarrow\!}\tau_0 under the isomorphism of Proposition A.23 then the first exponent in formula (?) of Proposition 5.19 becomes

    \left\lvert \tau \right\rvert-\operatorname{\mathrm{asc}}_\eta(\tau) - 1 = \sum_{j=1}^r \left\lvert \tau_j \right\rvert-\operatorname{\mathrm{asc}}(\underline{i}) - 1 = \sum_{j=1}^r (\left\lvert \tau_j \right\rvert-1) + \operatorname{\mathrm{des}}(\underline{i})

    and thus

    w(\tau,\eta) = \int_{-1}^0 x^{\operatorname{\mathrm{des}}(\underline{i})}(1+x)^{\operatorname{\mathrm{asc}}(\underline{i})} \prod_{j=1}^r x^{\left\lvert \tau_j \right\rvert-1} \, d x

    and summing over the cartesian product yields the total value

    \sum_{\tau\in T_v} w(\tau,\eta) = \frac{p_1!p_2!\cdots p_n!}{q_1!q_2!\cdots q_r!} \int_{-1}^0 x^{\operatorname{\mathrm{des}}(\underline{i})}(1+x)^{\operatorname{\mathrm{asc}}(\underline{i})} \prod_{j=1}^r \left( \sum_{\rho\in\mathcal{OP}_{q_j}} x^{\left\lvert \rho \right\rvert-1} \right) d x .

    Finally note that

    \sum_{\rho\in\mathcal{OP}_{q}}x^{\left\lvert \rho \right\rvert-1}= \sum_{\sigma \in \mathcal{P}_{q}} \left\lvert \sigma \right\rvert!\, x^{\left\lvert \sigma \right\rvert-1} =\sum_{k=1}^{q} k!\, S(q,k) x^{k-1}.

    is indeed the homogeneous Euler polynomial as claimed, see Remark 5.23.

    Remark

    Remark 6.18

    L9XJ

    The authors were not able to prove Theorem 6.17 as a corollary of Propositions 5.5 and (5.21) although Goldberg coefficients appear in both formulas.

    Definition

    Definition 7.1

    RA0M
    1. An element k \in[n] is called a singleton of \pi \in \mathcal{OP}_n if \{k\} \in \pi.

    2. Let \mathcal{S}=(\mathcal{U}, \tilde{\varphi},(\iota^{(i)})_{i \geq 1}) be a spreadability system for a \mathcal{B}-ncps (\mathcal{A}, \varphi). We will say that the singleton condition holds for \mathcal{S} if \varphi_\pi(X_1,X_2,\ldots,X_n)=0 for every tuple (X_1,X_2,\dots,X_n) and every partition \pi containing a singleton \{k\} such that \varphi(X_k)=0.

    Theorem

    Theorem 7.2

    OHIW

    Assume that a spreadability system (\mathcal{U}, \tilde{\varphi},(\iota^{(i)})_{i \geq 1}) for a \mathcal{B}-ncps (\mathcal{A}, \varphi) satisfies the singleton condition. Assume \varphi(X)=0 and let Y_N:=\frac{N.X}{\sqrt{N}}. Then, for each n \in\mathbb{N} and \rho \in \mathcal{P}_n,

    \lim_{N\to\infty}\varphi_\rho(Y_N, Y_N, \dots, Y_N)= \begin{cases} \sum_{\substack{\pi\in\mathcal{OP}_n^{(2)}\\ \pi\leq \rho}} \frac{1}{\left\lvert \pi \right\rvert!}\varphi_\pi(X,X,\ldots,X), &n \text{~is even},\\ 0,&n\text{~is odd}. \end{cases}

    where \mathcal{OP}_n^{(2)} is the set of pair ordered set partitions, i.e. every block of \pi \in \mathcal{OP}_n^{(2)} contains exactly 2 entries.

    Proof of Theorem 7.2

    Note that the following holds: if \pi has a singleton at k and \varphi(X_k)=0, then

    K_\pi(X_1,X_2,\ldots,X_n)=0.

    This holds because K_\pi(X_1,X_2,\ldots,X_n) is the coefficient of N in \varphi_\pi(N.X_1,N.X_2,\ldots,N.X_n), and

    \varphi_\pi(N.X_1,N.X_2,\ldots,N.X_n)=\sum_{i_1,i_2,\ldots,i_n\in[n]}\varphi_{\pi\curlywedge \kappa(i_1,i_2,\ldots,i_n)}(X_1,X_2,\ldots,X_n)=0

    because each \pi\curlywedge \kappa(i_1,i_2,\ldots,i_n) has a singleton at k.

    Now multilinearity and extensivity of cumulants imply

    K_\pi(Y_N,Y_N,\ldots,Y_N)= N^{-\frac{n}{2}+\left\lvert \pi \right\rvert}K_\pi(X,X,\ldots,X).

    If \pi has a singleton, this is zero. If \pi does not have a singleton nor \pi is not a pair ordered set partition, then \left\lvert \pi \right\rvert<\frac{n}{2}. Therefore,

    \lim_{N\to\infty}K_\pi(Y_N,Y_N,\ldots,Y_N)= \begin{cases} K_\pi(X,X,\ldots,X),&\text{if~$n$ is even and~}\pi\in\mathcal{OP}_n^{(2)}, \\ 0, &\text{otherwise}. \end{cases}

    If \pi\in\mathcal{OP}_n^{(2)}, then K_\pi(X,X,\ldots, X)= \varphi_\pi(X,X,\ldots,X) from Theorem 3.11 because the expectations \varphi_\sigma(X,X,\ldots,X) all vanish for \sigma <\pi from the singleton condition. Finally, from Theorem 3.11, we obtain the conclusion.

    Definition

    Definition A.1 (Set Partitions)

    E49Y

    A set partition, or simply partition, of a finite set A is a set of mutually disjoint subsets \pi=\{P_1,P_2,\ldots,P_k\} such that \cup_{i=1}^k P_i = A. The number k is the size of the partition and denoted by \left\lvert \pi \right\rvert. The elements P\in\pi are called blocks of \pi. The set of partitions of A is denoted by \mathcal{P}_A. We are mostly concerned with the case A=[n] and in this case \mathcal{P}_{[n]} is abbreviated to \mathcal{P}_n and called (set) partitions of order n. As is well known there is a one-to-one correspondence between set partitions \pi of [n] and equivalence relations on [n] by defining for \pi\in\mathcal{P}_n and i,j \in [n] the relation i \sim_\pi j to hold if and only if there is a block P \in \pi such that both i,j \in P.

    Definition

    Definition A.2 (Refinement Order)

    1BM3

    For partitions \pi and \sigma we write \pi \leq \sigma if for any block P \in \pi, there exists a block S \in \sigma such that P \subseteq S. In other words, every block of \sigma is a union of blocks of \pi. The minimal element of this lattice is \hat{0}_n=\{\{1\},\{2\},\dots,\{n\}\} and the maximal element \hat{1}_n=\{[n]\}.

    Definition

    Definition A.3 (Classes of Set Partitions)

    E9V5

    Let \pi\in\mathcal{P}_n be a set partition.

    1. Two (distinct) blocks B and B'\in\pi are said to be crossing if there are elements i<i'<j<j' such that i,j\in B and i',j'\in B'. \pi is called noncrossing if there are no crossing blocks, i.e., if there is no quadruple of elements i<j<k<l s.t. i\sim_\pi k, j\sim_\pi l and i\not\sim_\pi j. The noncrossing partitions of order n form a sublattice which we denote by \mathcal{NC}_n.

    2. Two blocks B,B' of a noncrossing partition \pi are said to form a nesting if there are i,j \in B such that i<k<j for any k \in B'. In this case B is called the outer block of the nesting and B' is called the inner block of the nesting.

    3. A block B of a noncrossing partition \pi is inner if B is the inner block of a nesting of \pi. If this is not the case B is called an outer block of \pi. The set of inner blocks of \pi is denoted by \mathrm{Inner}(\pi) and the set of outer blocks of \pi by \mathrm{Outer}(\pi).

    4. An interval partition is a partition \pi for which every block is an interval. Equivalently, this means that \pi is noncrossing and has no nestings. The set of interval partitions of [n] is denoted by \mathcal{I}_n.

    Analogous definitions apply to any finite totally ordered set A and the corresponding sets of partitions are denoted \mathcal{P}_A, \mathcal{NC}_A, \mathcal{I}_A etc.

    Proposition

    Proposition A.4

    B72C

    The lattice of interval partitions \mathcal{I}_n is anti-isomorphic to the Boolean lattice \mathcal{B}_{n-1} via the lattice anti-isomorphism

    (I_1, I_2, \ldots, I_p) \mapsto\{r_1, r_2, \ldots,r_{p-1}\} \subseteq [n-1],

    where the blocks I_i are uniquely determined by their maximal elements r_i; note that always r_p =n.

    Definition

    Definition A.5

    AXOS

    Fix a number n\in\mathbb{N} and a subset A\subseteq [n].

    1. Among all noncrossing partitions containing A^c as an outer block there is a maximal one, which we denote by \operatorname{\mathrm{\nu_{\max}}}(A). Removing A^c we obtain an interval partition of A which we denote by \operatorname{\mathrm{\iota_{\max}}}(A); in other words, the blocks of \operatorname{\mathrm{\iota_{\max}}}(A) consist of the maximal contiguous subintervals of A. Yet in another interpretation, the blocks of \operatorname{\mathrm{\iota_{\max}}}(A) are the connected components of the graph induced on A from the integer line.

    2. Among all noncrossing partitions containing A^c as a block there is a maximal one, which we denote by \operatorname{\mathrm{\tilde{\nu}_{\max}}}(A). Removing A^c we obtain a noncrossing partition of A which we denote by \operatorname{\mathrm{\tilde{\iota}_{\max}}}(A); in other words, the blocks of \operatorname{\mathrm{\tilde{\iota}_{\max}}}(A) consist of the maximal contiguous subintervals of A when we consider it on the circle, i.e., the blocks of \operatorname{\mathrm{\tilde{\iota}_{\max}}}(A) are the connected components of the graph induced on A from the Cayley graph of \mathbb{Z}_n.

    See Figure ? for examples.

    Remark

    Remark A.6

    FYYG
    1. Construction ? occurs in some examples, see, e.g., Examples 4.6 and 4.7. It gives rise to the unshuffle coproduct of Ebrahimi-Fard and Patras [EbrahimiFardPatras:2015:cumulants].

    2. Construction ? occurs in the recursion (?) for free cumulants and Example 4.8.

    Proposition

    Proposition A.7

    DYKY

    Let P_n be one of \mathcal{P}_n, \mathcal{NC}_n and \mathcal{I}_n. Then for any pair of elements \sigma,\pi\in P_n such that \sigma\leq\pi there are uniquely determined numbers k_j such that the interval [\sigma,\pi] is isomorphic (as a lattice) to the direct product

    P_1^{k_1}\times P_2^{k_2} \times \dotsm \times P_n^{k_n} .
    Definition

    Definition A.8 (Ordered set partitions)

    3KAY
    1. An ordered set partition of a set A is a sequence (P_1, P_2, \ldots,P_p) of distinct blocks such that \{P_1, P_2, \ldots,P_p\} is a set partition of A. In other words, it is a set partition with a total ordering of its blocks. The set of ordered set partitions of A is denote by \mathcal{OP}_A and \mathcal{OP}_{[n]} is abbreviated to \mathcal{OP}_n.

    2. Ordered set partitions with a fixed number p of blocks are in bijection with surjective functions from [n] to [p]. This bijection is implemented by identifying an ordered set partition \pi=(P_1,P_2,\dots, P_p) \in \mathcal{OP}_n with the function mapping an element i \in [n] to the label k of the block P_k such that i\in P_k, which we denote by \pi(i):=k. This function can be represented by a packed word [NovelliThibon:2006:construction], i.e., the sequence of the images \pi(1)\pi(2)\dots\pi(n). When \pi consists of singletons only, i.e., \# P_i=1 for all i\in [p] then \pi can be identified with a permutation of [n], and the notations introduced above are consistent with the familiar notations for permutations.

    Remark

    Remark A.9

    ED29

    Ordered set partitions are also known under the name of set compositions, see, e.g., [BergeronZabrocki:2009:hopf], pseudopermutations [KrobLatapy:2000:pseudo]. In particular, packed words can be represented as multiset permutations and this point of view is crucial in Section ?.

    Definition

    Definition A.10

    EDQB
    1. The ordered kernel set partition \kappa(i_1,i_2,\ldots,i_n) of a multiindex (i_1,i_2,\ldots,i_n) is defined as follows. First, pick the smallest value, say p_1, from i_1, i_2,\ldots, i_n and define the block P_1=\{k \in[n] \mid i_k=p_1\}. Next, pick the second smallest value p_2 from i_1, i_2,\ldots, i_n and define the block P_2=\{k \in[n]\mid i_k=p_2 \}. By repeating this procedure, we obtain an ordered set partition (P_1, P_2, \ldots), which we denote by \kappa(i_1, i_2,\ldots,i_n).

    2. The kernel set partition \bar{\kappa}(i_1, i_2,\ldots,i_n) of a sequence of indices is defined as the underlying set partition \overline{\kappa(i_1, i_2,\ldots,i_n)} of the corresponding ordered kernel set partition. In other words, it is the equivalence relation such that by p\sim q if and only if i_p=i_q.

    Definition

    Definition A.11 (Order Dropping Map)

    W4RT

    Let \pi \mapsto \bar{\pi} be the map from \mathcal{OP}_n onto \mathcal{P}_n which drops the order on blocks, that is,

    (P_1,P_2,\ldots) \mapsto \{P_1,P_2,\ldots\}.

    We say that an ordered set partition is in canonical order if the blocks are sorted in ascending order according to their minimal elements.

    Definition

    Definition A.12 (Refinement Order)

    2RQG

    Given two ordered set partitions \pi=(P_1, P_2,\dots,P_p) and \sigma=(S_1, S_2,\dots,S_s) \in \mathcal{OP}_n, we define the order relation \pi \leq \sigma by the following requirements:

    1. \bar{\pi}\leq\bar{\sigma} as set partitions.

    2. If P_i \subseteq S_k, P_j \subseteq S_l for i<j, then k \leq l.

    Remark

    Remark A.13

    6MZZ

    In other words, \pi\leq\sigma if every block of \sigma is a union of a contiguous sequence of blocks of \pi. The elements dominated by a given ordered set partition \sigma are obtained as follows:

    1. pick a label i

    2. split the block with label i into two

    3. label one of the pieces with i and the other one with i+1

    4. increment by one all other labels larger than i.

    See Figure ? for examples.

    The paper has a figure here, which this page does not show.

    This order makes (\mathcal{OP}_A,\leq) a poset (but not a lattice as will be seen shortly). The Hasse diagram of \mathcal{OP}_3 is shown in Figure ?; the order is read from outside to inside.

    The paper has a figure here, which this page does not show.

    Note that \pi\leq\sigma implies \bar\pi\leq\bar\sigma, but not vice versa.

    Definition

    Definition A.14 (Classes of Ordered Set Partitions)

    D98A
    1. An ordered set partition \pi\in\mathcal{OP}_n is called noncrossing if the underlying set partition \bar{\pi} has this property. The set of ordered noncrossing partitions of [n] is denoted by \mathcal{ONC}_n. Outer blocks and inner blocks of an ordered noncrossing partition are defined according to the case of noncrossing partitions.

    2. An interval ordered set partition of [n] is an ordered set partition \pi such that \bar{\pi} \in \mathcal{I}_n. The set of interval ordered set partitions of [n] is denoted by \mathcal{OI}_n.

    3. A noncrossing ordered set partition \pi=(P_1, P_2, \dots,P_{\left\lvert \pi \right\rvert}) is called monotone partition if for every nesting the outer block precedes the inner block; in other words, the order of the blocks implements a linearization of the partial order given by the nesting relation. The set of monotone partitions is denoted by \mathcal{M}_n.

    Example

    Example A.15

    Y9ML

    Figure ? shows some examples of monotone and non-monotone partitions whose underlying noncrossing partition is \bar{\pi}=\begin{picture}(32,6.5)(1,0) \put(2,0){\line(0,1){7.5}} \put(8,0){\line(0,1){4.5}} \put(14,0){\line(0,1){4.5}} \put(20,0){\line(0,1){7.5}} \put(26,0){\line(0,1){7.5}} \put(32,0){\line(0,1){4.5}} \put(8,4.5){\line(1,0){6}} \put(2,7.5){\line(1,0){24}} \put(32,4.5){\line(1,0){0}} \end{picture}.

    The paper has a figure here, which this page does not show.
    Remark.

    Interval partitions are characterized by the property that the canonical order defined in Definition A.11 extends to the entire blocks, i.e., the blocks of an interval partition \{I_1,I_2,\dots,I_p\} can be uniquely ordered so that i<j whenever i\in I_s, j\in I_t, s<t. This ordering provides a natural embedding \mathcal{I}_n \subseteq \mathcal{OI}_n and we may write (I_1, I_2, \ldots, I_p) \in \mathcal{I}_n rather than \{I_1, I_2, \dots,I_p\}\in\mathcal{I}_n.

    Remark

    Remark A.16

    6Q31

    Interval partitions are characterized by the property that the canonical order defined in Definition A.11 extends to the entire blocks, i.e., the blocks of an interval partition \{I_1,I_2,\dots,I_p\} can be uniquely ordered so that i<j whenever i\in I_s, j\in I_t, s<t. This ordering provides a natural embedding \mathcal{I}_n \subseteq \mathcal{OI}_n and we may write (I_1, I_2, \ldots, I_p) \in \mathcal{I}_n rather than \{I_1, I_2, \dots,I_p\}\in\mathcal{I}_n.

    Definition

    Definition A.17

    XLR0

    An ordered pseudopartition of [n] is a sequence (P_1, P_2, \dots,P_p) of disjoint subsets of [n] such that \cup_{i=1}^p P_i=[n] with empty blocks allowed. We keep the notation \left\lvert \pi \right\rvert=p for the length, now including empty blocks. The set of ordered pseudopartitions of [n] is denoted by \mathcal{OPP}_n.

    Lemma

    Lemma A.18

    E5HU

    The poset of ordered set partitions \mathcal{OP}_n is isomorphic to the poset of nonempty chains in the boolean lattice 2^n with the reverse refinement order, i.e., a chain \emptyset= A_0\subset A_1\subset \dotsm \subset A_k= [n] is smaller than a chain \emptyset= B_0\subset B_1\subset \dotsm \subset B_l= [n] if it is finer, i.e., as sets \{B_0,B_1,\dots,B_l\} \subseteq \{A_0,A_1,\dots,A_k\}.

    Proof of Lemma A.18

    The bijection is given by the map

    \Phi_n:(A_1\subset A_2\subset \dotsm \subset A_k) \mapsto (A_1,A_2\setminus A_1,A_3\setminus A_2,\dots,A_k\setminus A_{k-1}).
    Definition

    Definition A.19

    OKWF
    1. For an ordered set partition \sigma=(S_1, S_2, \dots,S_s) \in \mathcal{OP}_n and a nonempty subset P\subseteq [n], let the restriction \sigma\rfloor_P\in\mathcal{OP}_P be the ordered set partition (P \cap S_1,P\cap S_2, \dots,P\cap S_s)\in \mathcal{OP}_P, where empty sets are dropped; see [PS2008], where this operation arises in the context of Hopf algebras.

    2. For \pi =(P_1, P_2,\ldots,P_p), \sigma=(S_1,S_2,\ldots,S_s) \in \mathcal{OP}_n, we define the quasi-meet operation \pi \curlywedge \sigma \in \mathcal{OP}_n to be (P_1 \cap S_1, P_1\cap S_2 , \ldots, P_1 \cap S_s, P_2 \cap S_1, P_2\cap S_2,\ldots, P_2\cap S_s, \dots, P_p \cap S_s), where empty sets are skipped. In other words, the quasi-meet is the concatenation of the restrictions \pi\curlywedge\sigma = \sigma\rfloor_{P_1} \sigma\rfloor_{P_2}\dotsm \sigma\rfloor_{P_p}.

    Remark

    Remark A.20

    XNCK

    The quasi-meet operation is associative and coincides with the multiplication operation in the Solomon-Tits algebra of the symmetric group [Solomon:1976:mackey] which resurfaced recently in the domain of Markov chains [Brown:2000:semigroups, Brown:2004:semigroup] and Hopf algebras of noncommutative quasi-symmetric functions [BergeronZabrocki:2009:hopf].

    Proposition

    Proposition A.21

    X04H
    1. The quasi-meet operation on \mathcal{OP}_n is compatible with the meet operation on \mathcal{P}_n, in the sense that \overline{\pi\curlywedge\sigma} = \bar\pi \wedge\bar\sigma.

    2. \pi\curlywedge \sigma \leq \pi for any \pi,\sigma \in\mathcal{OP}_n.

    3. If \sigma \leq \rho \in \mathcal{OP}_n, then \sigma \curlywedge \pi \leq \rho \curlywedge \pi for any \pi \in \mathcal{OP}_n.

    4. \pi\curlywedge\sigma = \pi \iff \bar\pi\leq\bar\sigma.

    5. \pi\curlywedge\sigma = \sigma \iff \sigma\leq\pi.

    Uses Definition A.19 · Used by Proof
    Proof of Proposition A.21

    The first three items are immediate from the definition. To see ?, assume first \pi\curlywedge\sigma=\pi. Then by ? also \bar\pi=\bar\pi\wedge\bar\sigma \leq\bar\sigma. Conversely, if \bar\pi\leq\bar\sigma, then \bar\pi\wedge\bar\sigma=\bar\pi and by ? \pi\curlywedge\sigma\leq \pi. Since the number of blocks of \bar\pi\wedge\bar\sigma and \pi\curlywedge\sigma are equal, we must have \pi\curlywedge\sigma=\pi.

    As for ?, if \pi\curlywedge\sigma=\sigma then it follows from ? that \sigma\leq\pi. On the other hand, if \sigma\leq\pi and \pi=(B_1,B_2,\dots,B_k), then \pi\curlywedge\sigma=\sigma\rfloor_{B_1}\sigma\rfloor_{B_2}\dotsm \sigma\rfloor_{B_k} and the order of the blocks of \sigma remains unchanged, so \pi\curlywedge\sigma=\sigma.

    Definition

    Definition A.22

    MM0W

    Let P be a poset. A down-set or order ideal is a subset I such that x\in I and y\leq x implies y\in I. The principal ideal generated by x, denoted by \operatorname{\downarrow\!} x, is the smallest down-set containing x, i.e.,

    \operatorname{\downarrow\!} x = \{y\in P:y\leq x\} .
    Proposition

    Proposition A.23

    5C41

    The principal ideal generated by an element \pi=(P_1,P_2,\dots,P_p)\in\mathcal{OP}_n is canonically isomorphic to

    \mathcal{OP}_{P_1}\times \mathcal{OP}_{P_2}\times\dotsm\times \mathcal{OP}_{P_p}

    via the map

    \sigma \mapsto (\sigma\rfloor_{P_1},\sigma\rfloor_{P_2},\dots,\sigma\rfloor_{P_p})
    Used by Proof
    Proposition

    Proposition A.24

    R8EQ

    Let \sigma,\pi \in \mathcal{OP}_n be ordered set partitions of size \left\lvert \sigma \right\rvert=s and \left\lvert \pi \right\rvert=p respectively such that \sigma\leq\pi. If \pi=(P_1,P_2,\dots P_p), let k_j be the number of blocks of \sigma contained in P_j, j\in\{1,2,\dots,p\}. Then k_1+k_2+\dots+k_p=s and as a poset the interval [\sigma,\pi] is canonically isomorphic to \mathcal{I}_{k_1} \times \mathcal{I}_{k_2}\times \cdots \times \mathcal{I}_{k_p}. More precisely, if \sigma=(S_1, S_2,\ldots, S_s) \leq \pi \in \mathcal{OP}_n, then there exists a unique \tau =(T_1, T_2, \ldots,T_p) \in \mathcal{I}_s such that

    \pi= \biggl(\bigcup_{i\in T_1} S_i, \bigcup_{i\in T_2} S_i, \ldots, \bigcup_{i\in T_p} S_i\biggr).

    The map \Phi: \mathcal{I}_{T_1}\times \mathcal{I}_{T_2}\times \cdots \times \mathcal{I}_{T_p} \to [\sigma,\pi] by

    (\tau_1,\tau_2,\ldots,\tau_p) \mapsto \biggl(\bigcup_{i\in T_{1,1}} S_i, \bigcup_{i\in T_{1,2}} S_i,\ldots, \bigcup_{i\in T_{1,k_1}} S_i, \ldots, \bigcup_{i\in T_{p,1}} S_i, \ldots, \bigcup_{i\in T_{p,k_p}} S_i \biggr),

    where \tau_i= (T_{i,1}, T_{i,2}, \ldots, T_{i,k_i}), is a bijection, and so its inverse establishes a bijection

    \Psi:= \Phi^{-1}: [\sigma,\pi]\to \mathcal{I}_{k_1} \times \mathcal{I}_{k_2} \times \cdots \times \mathcal{I}_{k_p}

    with k_i:=\left\lvert T_i \right\rvert. The composition (k_1,k_2,\dots,k_p) is called the type of the interval [\sigma,\pi].

    Proof

    Proof

    Z0IY

    Let \pi=(P_1, P_2,\ldots,P_p). Each block of \pi is the union of blocks of \sigma, and so we can find A \subseteq [s] such that P_1=\cup_{i\in A} S_i. We show that there exists k such that A = [k]. Suppose that there are 1\leq u<v \leq p such that u\notin A and v\in A. Then there is j\geq2 such that S_u \subseteq P_j. This contradicts the fact that u<v, S_v \subseteq P_1 and \sigma\leq \pi. Hence A = [k] for some k. Removing the first block of \pi and the first k blocks of \sigma we can repeat the argument with the ordered set partitions (S_{k+1},S_{k+2},\ldots,S_s) \leq (P_2,P_3,\ldots,P_p) and find P_2=\cup_{k+1 \leq i \leq k+l}S_i for some l. After a finite number of iterations we thus construct a unique interval partition \tau =(T_1,T_2,\ldots,T_p) \in \mathcal{I}_s such that

    \pi=\biggl(\bigcup_{i\in T_1} S_i, \bigcup_{i\in T_2} S_i, \ldots, \bigcup_{i\in T_p} S_i\biggr).

    Clearly the image of \Phi is contained in [\sigma,\pi] and \Phi is injective and it remains to show surjectivity. To this end pick an arbitrary \rho \in [\sigma,\pi]. From the first part of the proposition we infer that \rho \geq \sigma is of the form

    \rho=\biggl(\bigcup_{i \in G_1} S_i, \bigcup_{i \in G_2} S_i, \ldots, \bigcup_{i\in G_t} S_i\biggr)

    for some \gamma =(G_1,G_2,\ldots, G_t) \in \mathcal{I}_s. Since \rho \leq \pi, the partition \gamma must be finer than \tau. Hence, each restriction \gamma\rfloor_{T_i} \in \mathcal{I}_{T_i}, i\in[p], consists of sequence of consecutive blocks of \gamma without splitting any original block G_1,G_2,\dots,G_t \in\gamma. Hence we obtain \Phi((\gamma_1,\gamma_2, \ldots, \gamma_p)) = \rho and the map \Phi is indeed bijective.

    Definition

    Definition A.25

    3ISX

    Denote by \mathbb{N}_{\mathrm{fin}}^\infty the set of finite sequences of positive integers:

    \mathbb{N}_{\mathrm{fin}}^\infty= \bigcup_{p=1}^\infty \mathbb{N}^p.

    For m \in \mathbb{N} let F_{\mathcal{OP}^m} be the set of \mathbb{C}-valued functions f on the set of m-chains

    \{(\sigma_1, \dots, \sigma_m) \in \cup_{n=1}^\infty \underbrace{(\mathcal{OP}_n\times \cdots \times \mathcal{OP}_n)}_{\text{$m$ fold}} \mid \sigma_1 \leq \sigma_2 \leq \cdots \leq \sigma_m\}.

    We are concerned with different levels of reduced incidence algebras.

    1. A function f \in F_{\mathcal{OP}^2} is said to be adapted if there is a family (f_{\underline{k}})_{\underline{k}\in \mathbb{N}_{\mathrm{fin}}^\infty} \subseteq \mathbb{C} such that

      f(\sigma,\pi) = f_{k_1,\dots,k_p},

      where (k_i)_{i=1}^p is the type of the interval [\sigma,\pi] as defined in Proposition A.24. The family (f_{\underline{k}})_{\underline{k}\in \mathbb{N}_{\mathrm{fin}}^\infty} is called the defining family.

    2. To any adapted function f\in F_{\mathcal{OP}^2} we associate its multivariate generating function

      \underline{Z}_f (\underline{z})= \sum_{\underline{k}\in\mathbb{N}_{\mathrm{fin}}^\infty} f_{\underline{k}} \underline{z}^{\underline{k}},\qquad \underline{z}=(z_1,z_2,z_3,\dots)

      with the usual multiindex convention \underline{z}^{\underline{k}}= z_1^{k_1}z_2^{k_2} \cdots, where z_1,z_2,\dots are commuting indeterminates.

    3. A function f \in F_{\mathcal{OP}^2} is said to be multiplicative if it is adapted and moreover the defining family satisfies

      f_{k_1,\dots, k_p} = \prod_{i=1}^p f_{k_i}.

      If f is multiplicative, then the sequence of values f_n=f(\hat{0}_n,\hat{1}_n) is called the defining sequence of f.

    4. To a multiplicative function f\in F_{\mathcal{OP}^2} we associate the (univariate) generating function

      Z_f (z)= \sum_{n=1}^\infty f_n z^n.
    5. A function f \in F_{\mathcal{OP}^3} is said to be quasi-multiplicative if there is an array of coefficients (f_{j k})_{j,k =1}^\infty \subseteq \mathbb{C} such that

      f(\sigma,\rho,\pi) = \prod_{i=1}^p \prod_{G\in\gamma_i} f_{i, \left\lvert G \right\rvert},

      where (\gamma_1,\dots, \gamma_p) is the image of \rho under the map \Psi in Proposition A.24. The array (f_{j k})_{j,k =1}^\infty is called the defining array of f.

    6. To a quasi-multiplicative function f\in F_{\mathcal{OP}^3} we associate a sequence of (univariate) generating functions

      Z_{f}^{(j)}( z)= \sum_{k=1}^\infty f_{j k} z^k,\qquad j \in \mathbb{N}.
    7. For f \in F_{\mathcal{OP}^3} and g \in F_{\mathcal{OP}^2} we define the convolution

      (f\otriangle g)(\sigma, \pi):= \sum_{\rho\in[\sigma,\pi]} f(\sigma,\rho,\pi)\,g(\rho,\pi),\qquad \sigma \leq \pi.
    Proposition

    Proposition A.26

    SLD0

    If a function f\in F_{\mathcal{OP}^3} is quasi-multiplicative and g \in F_{\mathcal{OP}^2} is adapted then f\otriangle g \in F_{\mathcal{OP}^2} is adapted and

    \underline{Z}_{f\otriangle g}(\underline{z}) = \underline{Z}_g(Z_f^{(1)}(z_1), Z_f^{(2)}(z_2), \dots),\qquad \underline{z} =(z_1,z_2,\dots).
    Uses Definition A.25 · Used by Proof, Proof
    Proof of Proposition A.26

    We use the notations in the statement of Proposition A.24. Pick any \rho \in[\sigma,\pi] and let (\gamma_1,\gamma_2, \ldots, \gamma_p):=\Psi(\rho) be its image under \Psi. Thus \gamma_i=(G_{i,1}, G_{i,2}, \ldots, G_{i,r_i})\in \mathcal{I}_{k_i} is an interval partition and we have the bijective images

    \begin{aligned} \Psi([\sigma,\rho])&= \mathcal{I}_{\left\lvert G_{1,1} \right\rvert}\times \mathcal{I}_{\left\lvert G_{1,2} \right\rvert}\times \cdots \times \mathcal{I}_{\left\lvert G_{1,r_1} \right\rvert} \times \mathcal{I}_{\left\lvert G_{2,1} \right\rvert} \times \cdots \times \mathcal{I}_{\left\lvert G_{p,r_p} \right\rvert}, \\ \Psi([\rho,\pi])&= \mathcal{I}_{\left\lvert \gamma_1 \right\rvert}\times \mathcal{I}_{\left\lvert \gamma_2 \right\rvert}\times \cdots \times \mathcal{I}_{\left\lvert \gamma_p \right\rvert}. \end{aligned}

    Hence

    \begin{aligned} (f\otriangle g)\,(\sigma,\pi) &= \sum_{\rho \in [\sigma,\pi]} f(\sigma,\rho,\pi)\,g(\rho, \pi) \\ &= \sum_{(\gamma_1,\gamma_2,\ldots,\gamma_p)\in \mathcal{I}_{k_1}\times \mathcal{I}_{k_2}\times \cdots \times \mathcal{I}_{k_p}}\prod_{i=1}^p\left(\prod_{G\in\gamma_i} f_{i,\left\lvert G \right\rvert}\right) g_{\left\lvert \gamma_1 \right\rvert, \left\lvert \gamma_2 \right\rvert, \dots, \left\lvert \gamma_p \right\rvert} \\ &= \sum_{r_i\in[k_i], 1 \leq i \leq p}\sum_{\substack{(n_{i k})_{i \in [p], k\in[r_i]} \\ n_{i1}+n_{i2}+\cdots + n_{i r_i}=k_i, 1\leq i \leq p}}\prod_{i=1}^p\left(\prod_{k=1}^{r_i} f_{i, n_{ik}}\right) g_{r_1, r_2, \dots, r_p} \\ &=: (f\otriangle g)_{k_1,k_2,\dots, k_p}. \end{aligned}

    This shows that for every pair (\sigma,\pi) the value f\otriangle g(\sigma,\pi) is determined by the structural sequence (k_1,\dots, k_p) and thus f\otriangle g is adapted. Now multiplying the terms with z^{\underline{k}} and summing over \underline{k} we obtain

    \begin{aligned} \underline{Z}_{f\otriangle g}(\underline{z}) &= \sum_{\underline{k}=(k_1,k_2,\dots) \in \mathbb{N}_{\mathrm{fin}}^\infty} (f\otriangle g)_{\underline{k}} \underline{z}^{\underline{k}} \\ &= \sum_{p\geq1} \sum_{\underline{r}=(r_1,r_2,\dots,r_p) \in\mathbb{N}^p} \sum_{\substack{(n_{i1}, \dots, n_{i r_i}) \in \mathbb{N}^{r_i} \\ 1 \leq i \leq p}}\prod_{i=1}^p\left(\prod_{k=1}^{r_i} f_{i, n_{ik}}z_i^{n_{ik}}\right) g_{r_1, r_2, \dots, r_p} \\ &= \sum_{p\geq1} \sum_{\underline{r}=(r_1,r_2,\dots,r_p) \in\mathbb{N}^p} g_{r_1, r_2, \dots, r_p}\prod_{i=1}^p Z_f^{(i)}(z_i)^{r_i} \\ &= \underline{Z}_g(Z_f^{(1)}(z_1), Z_f^{(2)}(z_2),\dots). \end{aligned}
    Corollary

    Corollary A.27

    82AN

    If f,g \in F_{\mathcal{OP}^2} are multiplicative, then so is f\ast g and

    Z_{f\ast g}(z) = Z_g(Z_f(z)).
    Uses Definition A.25 · Used by Proof
    Proof of Corollary A.27

    Given a multiplicative function f\in F_{\mathcal{OP}^2}, we lift it to a quasi-multiplicative function \tilde{f}\in F_{\mathcal{OP}^3} via its defining family \tilde{f}_{j k}= f_k, i.e., \tilde{f}(\sigma,\rho,\pi):=f(\sigma,\rho) and the generating functions are Z_{\tilde{f}}^{(j)}(z)= Z_{\tilde{f}}(z) for all j \geq 1. On the other hand, g is multiplicative, therefore adapted with g_{k_1,\dots, k_p}= g_{k_1} \cdots g_{k_p} and has generating function

    \underline{Z}_g(z_1,z_2,\dots)= \sum_{p=1}^\infty \sum_{(k_1,k_2,\dots, k_p) \in \mathbb{N}^p} g_{k_1}g_{k_2} \dotsm g_{k_p}z_1^{k_1}z_2^{k_2}\cdots z_p^{k_p} = \sum_{p=1}^\infty Z_g(z_1) Z_g(z_2) \dotsm Z_g(z_p).

    By Proposition A.26

    \underline{Z}_{\tilde{f}\otriangle g}(z_1,z_2,\dots)= \sum_{p=1}^\infty Z_g(Z_f(z_1))Z_g(Z_f(z_2)) \dotsm Z_g(Z_f(z_p)).

    This shows that (\tilde{f}\otriangle g)_{k_1,k_2,\dots, k_p}= \prod_{i=1}^p h_{k_i}, where h_k:= \frac{1}{k!}\frac{d^k}{dz^k}\big|_{z=0}Z_g(Z_f(z)). Therefore \tilde{f} \otriangle g = f \ast g is multiplicative and Z_{f\ast g}= Z_g(Z_f(z)).

    Definition

    Definition A.28

    79Y1

    Given ordered set partitions \sigma\leq\rho\leq \pi =(P_1,P_2,\dots,P_p)\in \mathcal{OP}_n, a sequence \underline{t}=(t_1,t_2,\dots)\in\mathbb{R}^\mathbb{N} and a number t\in \mathbb{R} we define

    \begin{aligned} \beta_{\underline{t}}(\sigma,\pi)&=\prod_{i=1}^p\binom{t_i}{\#(\sigma\rfloor_{P_i})}, \\ \beta_t(\sigma,\pi) &= \beta_{(t,t,\dots)}(\sigma,\pi), \\ \gamma_{\underline{t}}(\sigma,\rho,\pi) &= \prod_{i=1}^p \prod_{G\in\gamma_i} \binom{t_i}{\left\lvert G \right\rvert}, \end{aligned}

    where \binom{t}{n} is the generalized binomial coefficient and (\gamma_1,\gamma_2,\dots, \gamma_p) is the image of \rho by the map \Psi in Proposition A.24. Moreover, for \sigma\leq\pi we define

    \begin{aligned} [\sigma:\pi] &= \prod_{P \in \bar{\pi}} \#(\sigma\rfloor_P), \\ [\sigma:\pi]!&= \prod_{P \in \bar\pi} \#(\sigma\rfloor_P)!, \\ \widetilde{\zeta}(\sigma,\pi) &= \frac{1}{[\sigma:\pi]!}, \\ \widetilde{\mu}(\sigma,\pi)&= \frac{(-1)^{\left\lvert \sigma \right\rvert-\left\lvert \pi \right\rvert}}{[\sigma:\pi]}. \end{aligned}
    Remark

    Remark A.29

    O0EU
    1. The combinatorial significance of the numbers (A.28) is

      [\sigma:\pi]! = \#\{\rho~ |~ \rho\leq \pi \text{ and } \bar{\rho}=\bar{\sigma}\}
    2. The values \beta_{\underline{t}}(\sigma,\pi) and \gamma_{\underline{t}}(\sigma,\rho,\pi) depend only on the first \left\lvert \pi \right\rvert elements of \underline{t}.

    3. \beta_{\underline{t}}, \beta_t, \gamma_{\underline{t}} are related via

      \beta_{\underline{t}}(\sigma,\pi) = \gamma_{\underline{t}}(\sigma, \pi,\pi) \quad \text{and}\quad \gamma_{\underline{t}}(\sigma, \rho,\pi) = \prod_{i=1}^p \beta_{t_i}(\sigma\rfloor_{P_i}, \rho\rfloor_{P_i}).
    Used by Proof
    Proposition

    Proposition A.30

    9Y7I

    Let f:\mathcal{OP}\to \mathbb{C} be a function which is invariant under permutations of the blocks, i.e., f(\pi)=f(\pi') whenever \overline{\pi}=\overline{\pi'} and denote by \bar{f}:\mathcal{P}\to\mathbb{C} the corresponding function on \mathcal{P}, then the convolutions f*_\mathcal{OP}\tilde{\zeta} and f*_\mathcal{OP}\tilde{\mu} are invariant as well and are given by

    \begin{aligned} f*_\mathcal{OP}\tilde{\zeta}(\pi) &= \bar{f}*_\mathcal{P}\zeta_\mathcal{P}(\overline{\pi}) \\ f*_\mathcal{OP}\tilde{\mu}(\pi) &= \bar{f}*_\mathcal{P}\mu_\mathcal{P}(\overline{\pi}) , \intertext{ i.e., } \sum_{\substack{\rho\in\mathcal{OP}\\ \rho\leq\pi}}f(\rho)\,\tilde{\zeta}(\rho,\pi) &= \sum_{\substack{\sigma\in\mathcal{P}\\ \sigma\leq\overline{\pi}}}\bar{f}(\sigma) \\ \sum_{\substack{\rho\in\mathcal{OP}\\ \rho\leq\pi}}f(\rho)\,\tilde{\mu}_{\mathcal{OP}}(\rho,\pi) &= \sum_{\substack{\sigma\in\mathcal{P}\\ \sigma\leq\overline{\pi}}}\bar{f}(\sigma)\,\mu(\sigma,\overline{\pi}) . \end{aligned}
    Proposition

    Proposition A.31

    IJ4Z
    1. The function \beta_{\underline{t}} \in F_{\mathcal{OP}^2} is adapted with defining family

      (\beta_{\underline{t}})_{\underline{k}} = \prod_{i=1}^p \binom{t_i}{k_i}.
    2. The function \gamma_{\underline{t}}\in F_{\mathcal{OP}^3} is quasi-multiplicative with defining array

      (\gamma_{\underline{t}})_{jk}= \binom{t_j}{k}.
    3. The functions \beta_t,\widetilde{\mu},\widetilde{\zeta}\in F_{\mathcal{OP}^2} are multiplicative with defining sequences

      \begin{aligned} &\beta_t(\hat{0}_n,\hat{1}_n)= \binom{t}{n}, \\ &\widetilde{\zeta}(\hat{0}_n,\hat{1}_n)= \frac{1}{n!}, \\ &\widetilde{\mu}(\hat{0}_n,\hat{1}_n) = \frac{(-1)^{n-1}}{n}. \end{aligned}
    Proof of Proposition A.31

    The claims follow by definition and by Proposition A.24.

    Corollary

    Corollary A.32

    MOB9
    1. The inverse function (with respect to the convolution \ast) of \widetilde{\mu} is \widetilde{\zeta}.

    2. For \underline{s},\underline{t} \in \mathbb{R}^\infty we have

      \gamma_{\underline{s}}\otriangle\beta_{\underline{t}} = \beta_{\underline{s} \circ \underline{t}},

      where \underline{s} \circ \underline{t}=(s_1 t_1, s_2 t_2, \dots).

    3. \beta_t satisfies the semigroup property \beta_s \ast \beta_t = \beta_{s t} for s,t \in\mathbb{R}.

    Used by Proof, Proof
    Proof of Corollary A.32

    ?   Since Z_{\widetilde{\mu}}(z)=\log (1+z) and Z_{\widetilde{\zeta}}(z)= e^z-1, we have Z_{\widetilde{\mu}}(Z_{\widetilde{\zeta}}(z))=z.

    ?   We have

    \begin{aligned} &Z_{\gamma_{\underline{s}}}^{(j)}(z) = \sum_{k=1}^\infty \binom{s_j}{k} z^k = (1+z)^{s_j}-1, \\ &\underline{Z}_{\beta_{\underline{t}}}(\underline{z}) = \sum_{\underline{k}} \prod_{i=1}^p \binom{t_i}{k_i} z_i^{k_i} = \sum_{p\geq1}\prod_{i=1}^p\left((1+z_i)^{t_i}-1\right). \end{aligned}

    By Proposition A.26, \underline{Z}_{\gamma_{\underline{s}}\otriangle\beta_{\underline{t}}}(\underline{z}) = \underline{Z}_{\beta_{\underline{t}}}(Z_{\gamma_{\underline{s}}}^{(1)}(z_1), Z_{\gamma_{\underline{s}}}^{(2)}(z_2),\dots), which equals \underline{Z}_{\beta_{\underline{s}\circ \underline{t}}}(\underline{z}).

    ?   We can use Z_{\beta_t}(z)=\sum_{n=1}^\infty \binom{t}{n}z^n = (1+z)^t-1 and Corollary A.27.

    Source and licence

    The text on this page is from arXiv:1711.00219v3 by Takahiro Hasebe and Franz Lehner, under the licence CC BY 4.0. The corpus made these changes: it read the items out of the LaTeX source, it expanded the macros of the paper, it put each proof after its statement, and it converted the text to HTML. The formulas are the TeX of the paper, and KaTeX renders them. A formula or a reference can look different from the PDF. The page is published under CC BY 4.0.

    The record of each item is in analysis/objects/1711/1711.00219.jsonl.