Remarks on a free analogue of the beta prime distribution

TL;DR

This paper introduces a free probability analogue of the classical beta prime distribution, constructed via free convolution, and explores its properties, moments, and relation to free Meixner distributions.

Contribution

It defines the free beta prime distribution using free convolution and analyzes its moments and connection to free negative binomials, extending classical distribution concepts.

Findings

Explicit calculation of moments using non-crossing linked partitions

Identification of the free beta prime as part of the free Meixner family

Development of free analogues of classical distributions like F, T, and beta distributions

Abstract

We introduce the free analogue of the classical beta prime distribution by the multiplicative free convolution of the free Poisson and the reciprocal of free Poisson distributions, and related free analogues of the classical $F$, $T$, and beta distributions. We show the rationales of our free analogues via the score functions and the potentials. We calculate the moments of the free beta prime distribution explicitly in combinatorial by using non-crossing linked partitions, and demonstrate that the free beta prime distribution belongs to the class of the free negative binomials in the free Meixner family.

Equations

$$\mu_{X}\big{(}p(\mathbb{X})\big{)}=\int_{\sigma(X)}p(t)d\nu(t)\;\mbox{ for }p(\mathbb{X})\in\mathbb{C}[\mathbb{X}].$$

$${\mu }_X⊞{\mu }_Y={\mu }_{X+Y}.$$

$$R_{\mu }(z)=\sum _{n=1}^{\infty }{r}_n{z}^{n-1}$$

$$G_{\mu }(z)={\int }_{\mathbb{R}}\frac{d\mu (x)}{z-x},$$

$$R_{\mu}(z)=G_{\mu}^{\mbox{\tiny$\langle{-1}\rangle$}}(z)-\frac{1}{\,z\,},$$

$$R_{\mu ⊞\nu }(z)=R_{\mu }(z)+R_{\nu }(z).$$

$$\mathcal{T}(\mathcal{H})=\mathbb{C}\Omega\oplus\Big{(}\bigoplus_{n=1}^{\infty}\mathcal{H}^{\otimes n}\Big{)},$$

$$⟨x_1\otimes x_2\otimes \cdots \otimes x_n|y_1\otimes y_2\otimes \cdots \otimes y_m⟩={\delta }_{n,m}\prod _{i=1}^n(x_i|y_i),$$

$$X_1$$

$$X_2$$

$$X_{3}=\ell_{1}+(p_{1}+q_{1}){\bm{1}}+(p_{2}+q_{2})\ell_{1}^{*}+(p_{3}+q_{3})\big{(}\ell_{1}^{*}\big{)}^{2}+\cdots$$

$$X=\ell+R_{\mu}\big{(}\ell^{*}\big{)}$$

$${\mu }_X⊠{\mu }_Y={\mu }_{X^{\frac{1}{2}}Y{X}^{\frac{1}{2}}}.$$

$$\Phi_{\mu}(z)=\sum_{n=1}^{\infty}m_{n}(\mu)z^{n}=\mbox{\small$\dfrac{1}{\,z\,}$}G_{\mu}\Big{(}\mbox{\small$\dfrac{1}{\,z\,}$}\Big{)}-1,$$

$$\Phi_{\mu}\Big{(}\mbox{\small$\dfrac{z}{z+1}$}S_{\mu}(z)\Big{)}=z.$$

$$S_{\mu}(z)=\mbox{\small$\dfrac{z+1}{z}$}\Phi_{\mu}^{\mbox{\tiny$\langle{-1}\rangle$}}(z).$$

$$S_{\mu ⊠\nu }(z)=S_{\mu }(z)S_{\nu }(z).$$

$$T_{\mu }(z)=\frac{1}{S_{\mu }(z)},$$

$$X=({\bm{1}}+\ell)T_{\mu}\big{(}\ell^{*}\big{)}.$$

$$d{\mu }_{\lambda }(x)=\frac{\sqrt{-(x-{\lambda }_{-})(x-{\lambda }_{+})}}{2\pi x}{I}_{[{\lambda }_{-},{\lambda }_{+}]}(x)dx+\max (1-\lambda ,0){\delta }_0(x),$$

$$G_{\mu_{\lambda}}(z)=\frac{z+(1-\lambda)-\sqrt{\big{(}z+(1-\lambda)\big{)}^{2}-4z}}{2z},$$

$$S_{P_{\lambda }}(z)=\frac{1}{z+\lambda },$$

$$S_{P_{\lambda }^{-1}}(z)=-z+(\lambda -1).$$

$$X(a,b)=P_b^{-\frac{1}{2}}{P}_a{P}_b^{-\frac{1}{2}},$$

$${\mu }_{X(a,b)}={\mu }_{P_a}⊠{\mu }_{P_b^{-1}}.$$

$$G_{X(a,b)}(z)=\frac{(b+1)z+(1-a)-\sqrt{(b-1{)}^2{z}^2-2(ab+a+b-1)z+(a-1{)}^2}}{2z(1+z)},$$

$$S_{X(a,b)}(z)=S_{P_{a}}(z)S_{P_{b}^{\mbox{\tiny$-1\!$}}}(z)=\frac{-z+(b-1)}{z+a}.$$

$$\displaystyle\Phi_{X(a,b)}^{\mbox{\tiny$\langle-1\rangle$}}(z)$$

$${\Phi }_{X(a,b)}(z)$$

$$M_{X(a,b)}(z)$$

$$=\frac{(b+1)+(1-a)z-\sqrt{((b-1)-(1+a)z{)}^2-4az(z+1)}}{2(1+z)}.$$

$$G_{X(a,b)}(z)$$

Full text

**Remarks on a free analogue of the beta prime distribution

**

Hiroaki Yoshida

Department of Information Sciences,

Ochanomizu University,

Otsuka, Bunkyo, Tokyo 112-8610 Japan.

yoshida@ocha.ac.jp

**Abstract. We introduce the free analogue of the classical beta prime distribution by the multiplicative free convolution of the free Poisson and the reciprocal of free Poisson distributions, and related free analogues of the classical $F$, $T$, and beta distributions. We show the rationales of our free analogues via the score functions and the potentials. We calculate the moments of the free beta prime distribution explicitly in combinatorial by using non-crossing linked partitions, and demonstrate that the free beta prime distribution belongs to the class of the free negative binomials in the free Meixner family.

**

**Keywords: Free Beta prime distribution; Free $F$-distribution; Free $T$-distribution; Free Beta distribution; Non-crossing linked partitions; Free Meixner family

Mathematics Subject Classification (2010) 46L54, 60E05 **

1 Introduction

A non-commutative or quantum probability space is a unital algebra $\mathcal{A}$ (possibly non-commutative) together with a linear functional, $\varphi:\mathcal{A}\to\mathbb{C}$ such that $\varphi({\bm{1}})=1$. If $\mathcal{A}$ is a $C^{*}$-algebra and $\varphi$ is a state then we call a non-commutative probability space $(\mathcal{A},\varphi)$ a $C^{*}$-probability space.

An element in $\mathcal{A}$ is regarded as a non-commutative random variable. The distribution of $X\in\mathcal{A}$ under $\varphi$ is the linear functional $\mu_{X}:\mathbb{C}[\mathbb{X}]\to\mathbb{C}$ where $\mathbb{C}[\mathbb{X}]$ is the algebra of polynomials in the variable $\mathbb{X}$ and $\mu_{X}$ is determined by \mu_{X}\big{(}p(\mathbb{X})\big{)}=\varphi\big{(}p(X)\big{)} for all polynomials $p(\mathbb{X})\in\mathbb{C}[\mathbb{X}]$. One should note that the distribution of a random variable $X\in\mathcal{A}$ is nothing more than a way of describing the moments $\varphi(X^{k})$.

If $(\mathcal{A},\varphi)$ is a $C^{*}$-probability space and $X\in\mathcal{A}$ is self-adjoint then the distribution of $X$ is given by the compactly supported probability measure $\nu$ on $\sigma(X)$, the spectrum of $X$, by

[TABLE]

In this case, we will call the measure $\nu$ the distribution of $X$ and we will write $\mu_{X}$ instead of $\nu$. This is the case we treat, that is, the probability measures in this paper are compactly supported on ${\mathbb{R}}$.

The notion of independence can be understood as a rule of calculating mixed moments. In classical (usual) probability such a rule leads to only one meaningful notion of (commutative) independence. In non-commutative case, however, one can consider several notions of independence. Among them, the free independence, first introduced by Voiculescu in [38], seems to be the most interesting and important notion in non-commutative probability.

Presently, the non-commutative probability equipped with free independence is called free probability, and its theoretical framework is very similar to that of classical probability. With a notion of independence, one can consider the corresponding convolution of probability measures. Namely, for freely independent random variables $X$ and $Y$ with respective distributions $\mu_{X}$ and $\mu_{Y}$, we introduce the additive free convolution of $\mu_{X}$ and $\mu_{Y}$ as the distribution of the sum of the random variables $X+Y$.

In classical probability the product of independent random variables is rather trivial. But in free probability it has a much richer structure because free independence is highly non-commutative. Hence it is worth defining the multiplicative free convolution as the distribution of the product of freely independent random variables. The precise definitions of free independence and the corresponding convolutions are mentioned in Section 2 below. The same section introduces some analytic tools for the calculation of the additive free and the multiplicative free convolutions.

It is a natural approach in free probability to look for the free analogue of the classical distributions with the same probabilistic framework as in classical probability. For example, Wigner’s semicircle law plays the Gaussian law role in free probability, which can be obtained as the free central limit distribution, and the Marchenko-Pastur distribution appears as the limit distribution of the free Poisson limit, hence it is often called the free Poisson distribution.

Surprisingly many classical characterization problems of the probability measures by independence have free probability counterparts. For example, the free analogue of Bernstein’s theorem was proved in [28], which says that for freely independent random variables $X$ and $Y$, the random variables $X+Y$ and $X-Y$ are freely independent if and only if $X$ and $Y$ are distributed according to the semicircle law. Moreover, the characterization of the semicircle law by free independence of linear and quadratic statistics in freely i.i.d. random variables was investigated in [17] (see also [23]).

Another free analogue of characterization problem by independence, in [34], [35], that is, a free version of Lukacs theorem, the classical version states that, for independent random variables $X$ and $Y$, the random variables $X+Y$ and $\dfrac{X}{X+Y}$ are independent if and only if $X$ and $Y$ are gamma distributed with the same scale parameter [25]. Here one should note that in the free analogue of Lukacs theorem, the free Poisson distribution plays the classical gamma distribution role. However, it is not very strange because the free Poisson distribution can be thought as the free $\chi^{2}$-distribution and the classical $\chi^{2}$-distribution is in a family of the classical gamma distributions. This phenomenon can be found in yet another characterization problem by free independence, namely, in the study of the free analogue of Matsumoto-Yor property in [36], which elucidated that the role of the classical gamma distribution is taken again by the free Poisson distribution.

Furthermore, in the free probability literature there are investigations on the free analogue of the characterization problems of the probability measure by regression conditions. A well-known example of such a direction is the characterization of the free Meixner distributions defined in [1] (see also [31]). Particularly, in the papers [8] and [14], the authors studied the free analogue of Laha-Lukacs regression problem, and found that the free Meixner distributions can be classified in six classes: semicircle, free Poisson, free Pascal (free negative binomial), free gamma, free binomial, and pure free Meixner distributions, which corresponds to the result in the classical case. Inspired by this result, free analogue of characterization problems by regression were further investigated in [15], [16], and [37].

In this paper we look for the free analogue of the classical distribution as in the literature of free probability to date.

In particular, the free beta prime distribution is investigated. In classical probability a certain dilation of the beta prime distribution gives the $F$-distribution, thus the free $F$-distribution can be also obtained immediately. Broadly speaking we define the free beta prime distribution as the distribution of the ratio of freely independent pair of free Poisson random variables that are based on the property that the distribution of the ratio of independent classical gamma distributed random variables yields the beta prime distribution. In other words, we will make the free Poisson random variables play the free gamma random variables role again. The free $T$-distribution and the free beta distribution can be also derived from the specialized free $F$-distribution and the free beta prime distribution, respectively, via some transformations.

Furthermore, by using the combinatorial object, the non-crossing linked partitions, first introduced in [12] with some set partition statistics, we will investigate the moments of the free beta prime distributions, and show that the free beta prime distributions can be classified into the free negative binomials of the free Meixner family.

The paper is organized as follows, in Section 2 we recall the definitions of free independence and of the corresponding additive and multiplicative convolutions, and we provide the analytic tools for calculation of these convolutions. We introduce the free beta prime distribution and some related analogues in Section 3; the rationales of which are discussed by the score functions in Section 4. In Section 5, we review the non-crossing linked partitions and some set partition statistics, and we give the combinatorial moment formula related to the multiplicative free convolution by using these set partition statistics. We investigate the moments of the free beta prime distributions in Section 6 and determine the class of the free beta prime distributions.

2 Preliminaries on free probability

In this section, we recall the definition of free independence and analytic tools for the calculation of the corresponding additive and multiplicative convolutions. The introduction here is far from being detailed. A comprehensive and good introduction to the theory of free probability can be found in the monographs [26], [30], and [43].

2.1 Free independence and additive free convolution

A family of subalgebras \big{(}\mathcal{A}_{i}\big{)}_{i\in I}, where ${\bm{1}}\in\mathcal{A}_{i}$, in the non-commutative probability space \big{(}\mathcal{A},\varphi\big{)} is freely independent if $\varphi(X_{1}\,X_{2}\cdots X_{n})=0$ whenever $\varphi(X_{k})=0$, $1\leq k\leq n$ and $X_{k}\in\mathcal{A}_{j(k)}$ where $j(1)\neq j(2)\neq\cdots\neq j(n)$, that is, consecutive indices are distinct.

We say that a family of random variables \big{(}X_{i}\big{)}_{i\in I} are freely independent (simply, free) if the subalgebras generated by $\{X_{i},{\bm{1}}\}$ are freely independent.

Let $X$ and $Y$ be free independent random variables in $(\mathcal{A},\varphi)$ with respective distributions $\mu_{X}$ and $\mu_{Y}$. Then the distribution of $X+Y$ is completely determined by the distributions $\mu_{X}$ and $\mu_{Y}$. Hence we define the additive free convolution operation $\boxplus$ of $\mu_{X}$ and $\mu_{Y}$ by

[TABLE]

Here we should note that the additive free convolution of distributions depends only on the distributions, not on the choice of particular random variables having those distributions, thus the operation $\boxplus$ is well-defined.

In order to calculate the additive free convolution, that is, to find the higher moments \varphi\big{(}(X+Y)^{n}\big{)}, Voiculescu invented the $R$-transform in [39].

For compactly supported probability measure $\mu$ on ${\mathbb{R}}$, the $R$-transform of $\mu$ is the power series

[TABLE]

defined as follows. Consider the Cauchy transform $G_{\mu}$ of the compactly supported probability measure $\mu$ on ${\mathbb{R}}$

[TABLE]

which is an analytic function in $\mathbb{C}\setminus\mathrm{supp}(\mu)$. We take the inverse of $G_{\mu}(z)$ in a neighborhood of $\infty$ because $G_{\mu}(z)$ is univalent and analytic there.

Then the $R$-transform of $\mu$ can be obtained as

[TABLE]

which is an analytic function in a neighborhood of [math], where G_{\mu}^{\mbox{\tiny\langle{-1}\rangle}} denotes the inverse of $G_{\mu}$.

The $R$-transform is the linearizing functor for the additive free convolution, that is,

[TABLE]

Hence the coefficients $\{r_{n}\}$ of the $R$-transforms are called the free cumulant.

2.2 Full Fock space and canonical random variables

An example of free independence which plays an important role in the theory of free probability is that of the creation and the annihilation operators on full Fock space.

Let $\mathcal{H}$ be a Hilbert space and let $\mathcal{T}(\mathcal{H})$ denote the full Fock space of $\mathcal{H}$

[TABLE]

where $\displaystyle{\mathcal{H}^{\otimes n}=\underbrace{\mathcal{H}\otimes\cdots\otimes\mathcal{H}}_{\footnotesize n-\mbox{times}}}$

and $\Omega$ is the distinguished unit vector called vacuum. The full Fock space $\mathcal{T}(\mathcal{H})$ is endowed with the inner product

[TABLE]

where $(\,\cdot\,|\,\cdot\,)$ is the inner product on the Hilbert space $\mathcal{H}$.

For $h\in\mathcal{H}$, let \ell(h)\in\mathcal{B}\big{(}\mathcal{T}(\mathcal{H})\big{)} be the left creation operator so that $\ell(h)\xi=h\otimes\xi$, and the adjoint operator $\ell(h)^{*}$ is called the left annihilation operator. We consider the state $\varphi(T)=\langle T\Omega,\Omega\rangle$ on \mathcal{B}\big{(}\mathcal{T}(\mathcal{H})\big{)} given by the vacuum vector $\Omega$, which is called the vacuum expectation. The following important example of a free family of subalgebras and the distinguished operators in \mathcal{B}\big{(}\mathcal{T}(\mathcal{H})\big{)} can be found in [39].

Theorem 2.1**.**

Let $(e_{i})_{i\in I}$ be a family of orthonormal vectors in $\mathcal{H}$. Then the family of subalgebras \big{(}\mbox{alg}\{\ell(e_{i}),\ell(e_{i})^{*}\}\big{)}_{i\in I} is free with respect to the vacuum expectation.

Proposition 2.2**.**

Let $e_{1}$ and $e_{2}$ be orthonormal unit vectors in $\mathcal{H}$, and $\ell_{1}=\ell(e_{1})$ and $\ell_{2}=\ell(e_{2})$ be the left creation operators in \mathcal{B}\big{(}\mathcal{T}(\mathcal{H})\big{)}. Let further

[TABLE]

Then the random variables $X_{1}+X_{2}$ and

[TABLE]

have the same distribution in \big{(}\mathcal{B}\big{(}\mathcal{T}(\mathcal{H})\big{)},\varphi\big{)}.

Since $\ell_{1}$ and $\ell_{2}$ are free with respect to the vacuum expectation, the above proposition gives the model for the additive free convolution. Namely, we denote the $R$-transform of a compactly supported probability measure $\mu$ on ${\mathbb{R}}$ by $R_{\mu}(z)$ then the random variable

[TABLE]

on $\mathcal{T}(\mathcal{H})$ is called the canonical operator for the additive free convolution. In this case, of course, the probability distribution of $X$ is given by $\mu$.

2.3 Multiplicative free convolution

As we have mentioned in Introduction that by high non-commutativity of freeness, the product of freely independent random variables has a richer structure than the classical case. Here one can introduce the multiplicative free convolution operation on compactly supported provability measures on $\mathbb{R}_{\geq 0}=[0,\infty)$.

Let $X$ and $Y$ be freely independent random variables in $(\mathcal{A},\varphi)$ with respective distributions $\mu_{X}$ and $\mu_{Y}$ compactly supported on $\mathbb{R}_{\geq 0}$. Then the distribution of $X^{\frac{1}{\,2\,}}YX^{\frac{1}{\,2\,}}$ is determined only by the distributions $\mu_{X}$ and $\mu_{Y}$. Thus we can define the multiplicative free convolution operation $\boxtimes$ of $\mu_{X}$ and $\mu_{Y}$ by

[TABLE]

The multiplicative free convolution is well-defined and becomes the commutative operation on compactly supported probability measures on ${\mathbb{R}}_{\geq 0}$.

An analytic tool for calculation of the multiplicative free convolution is the $S$-transform invented by Voiculescu in [40], which is multiplicative map and plays the Mellin transform role in classical probability.

Let $\mu$ be a compactly supported probability measure on $[0,\infty)$ with non-zero mean $m_{1}(\mu)\neq 0$. We shall first introduce the $\Phi$-series of $\mu$, which is essentially the moment generating function but without constant

[TABLE]

where $m_{n}(\mu)$ stands for the $n$th moment of the probability measure $\mu$ and we denote by $G_{\mu}$ the Cauchy transform of $\mu$.

Then the $S$-transform and the $\Phi$-series are related by the functional relation

[TABLE]

Thus the $S$-transform can be obtained as

[TABLE]

The $S$-transform is multiplicative map for the multiplicative free convolution, that is

[TABLE]

Like in the additive case, the canonical operator for the multiplicative free convolution is known as follows. Let $\mu$ be a compactly supported probability measure on $[0,\infty)$ with non-zero mean, we introduce the $T$-transform of $\mu$ by the reciprocal of the $S$-transform of $\mu$, that is,

[TABLE]

the remarkable combinatorial properties of which were investigated by Dykema in [12]. Some of detailed results are discussed in Section 5 below.

Then the canonical operators for the multiplicative case can be given by the operator on $\mathcal{T}(\mathcal{H})$ of the form

[TABLE]

3 The free beta prime distribution

and some derived free analogues

3.1 The free beta prime distribution

Let $P_{\lambda}$ be a free Poisson random variable of parameter $\lambda>0$, which has the probability distribution $\mu_{\lambda}$, the Marchenko-Pastur law of parameter $\lambda$,

[TABLE]

where \lambda_{\pm}=\big{(}1\pm\sqrt{\lambda}\big{)}^{2} and $I$ stands for the indicator function.

Then it is well-known that the Cauchy transform of $\mu_{\lambda}$ is given by

[TABLE]

where the branch of the square root should be chosen so that $\displaystyle{\lim_{z\to\infty}G_{\mu_{\lambda}}(z)=0}$.

For a self-adjoint random variable $X$ with the probability distribution $\mu$ on ${\mathbb{R}}$, we denote the Cauchy transform of $\mu$ by an abbreviated notation $G_{X}$. We shall apply such a notation to other characteristic series of the distribution of a non-commutative random variable.

Since we will calculate a multiplicative free convolution involving $P_{\lambda}$, we need the $S$-transform of $P_{\lambda}$, and it is given by

[TABLE]

which can be found in, for instance, [3].

The following Proposition direct consequence of Proposition 3.13 in [20].

Proposition 3.1**.**

Let $P_{\lambda}$ be a free Poisson random variable of parameter $\lambda$. If the parameter $\lambda>1$, then $P_{\lambda}$ is strictly positive and invertible, and the $S$-transform of the probability distribution of $P_{\lambda}^{-1}$ is given by

[TABLE]

Now we can introduce the free analogue of the beta prime distribution by the free multiplicative convolution of the distributions of free Poisson and the reciprocal of free Poisson random variables.

Let $P_{a}$ and $P_{b}$ be freely independent free Poisson random variables of parameter $a$ and $b$, respectively. We assume parameter $b>1$ while $a>0$. Since $b>1$ the random variable $P_{b}$ is strictly positive, we can define the self-adjoint operator

[TABLE]

and its probability distribution is given by the multiplicative free convolution of the distributions of $P_{a}$ and $P_{b}^{-1}$,

[TABLE]

Proposition 3.2**.**

The Cauchy transform of the distribution of $X(a,b)$ is given by

[TABLE]

where the branch of the square root is chosen so that $\displaystyle{\lim_{z\to\infty}G_{X(a,b)}(z)=0}$.

Proof.

By the multiplicativity, the $S$-transform of $X(a,b)$ is given by

[TABLE]

Thus we can calculate the moment generating function of $X(a,b)$ as follow.

[TABLE]

Eventually, the Cauchy transform is given by

[TABLE]

∎

Applying Stieltjes inversion formula, one can obtain the probability measure of the distribution of $X(a,b)$.

Concerning the point mass of the measure, the function $G_{X(a,b)}(z)$ has the simple poles at $z=0$ and $z=-1$ and the residues are calculated as

[TABLE]

respectively. Thus the measure has a point mass $(1-a)$ at [math] if $0<a<1$, while $z=-1$ is removable singularity.

The density $f_{\mbox{\scriptsize f}{\bm{\beta}}^{\prime}}(x;a,b)$ of the Lebesgue absolutely continuous part is given by

[TABLE]

which is supported on the interval $[\gamma_{-},\gamma_{+}]$, where \displaystyle{\gamma_{\pm}=\bigg{(}\mbox{\small\dfrac{\sqrt{ab}\pm\sqrt{a+b-1}}{b-1}}\bigg{)}^{\!2}} are the roots of the quadratic equation $(b-1)^{2}x^{2}-2(ab+a+b-1)x+(a-1)^{2}=0$.

Definition 3.3**.**

For $a>0$ and $b>1$, the free beta prime distribution ${\mathrm{f}{\bm{\beta}}}^{\prime}(a,b)$ of parameters $a$ and $b$ is the probability measure

[TABLE]

where \displaystyle{\gamma_{\pm}=\bigg{(}\mbox{\small\dfrac{\sqrt{ab}\pm\sqrt{a+b-1}}{b-1}}\bigg{)}^{\!2}}. **

3.2 The free $F$-distribution

Let $X_{1}$ and $X_{2}$ be the classical $\chi^{2}$-distributed random variables of degree of freedom $d_{1}$ and $d_{2}$, respectively, and assume $X_{1}$ and $X_{2}$ are independent. Then it is well-known that the probability distribution of random variable of the scaled ratio Y=\mbox{\small\dfrac{X_{1}/d_{1}}{X_{2}/d_{2}}}=\Big{(}\mbox{\small\dfrac{d_{2}}{d_{1}}}\Big{)}\mbox{\small\dfrac{X_{1}}{X_{2}}} is the $F$-distribution $F(d_{1},d_{2})$.

As mentioned in the next section with respect to the rationales of free analogue, the free Poisson law of parameter $\lambda$ can be regarded as the free $\chi^{2}$-distribution of degree of freedom $\lambda$, and we have introduced the free beta prime distribution ${\mathrm{f}{\bm{\beta}}^{\prime}}(a,b)$ essentially based on the ratio $\dfrac{P_{a}}{P_{b}}$ of freely independent pair of free Poisson random variables. With these facts in mind, it is natural to introduce the free $F$-distribution in the following manner:

Definition 3.4**.**

For $a>0$ and $b>1$, let $X(a,b)$ be a free beta prime ${\mathrm{f}{\bm{\beta}}^{\prime}}(a,b)$-distributed random variable. The free $F$-distribution $\mathrm{f}F(a,b)$ is defined as the distribution of the scaled random variable \mbox{\small\dfrac{b}{,a,}}\,X(a,b), the probability measure of which is given by D_{\frac{b}{\,a\,}}\big{(}\mu_{a,b}\big{)}, where $D_{\frac{b}{\,a\,}}$ stands for the dilation. **

Here we shall give the measure \nu_{a,b}=D_{\frac{b}{\,a\,}}\big{(}\mu_{a,b}\big{)} explicitly. If $0<a<1$ then $\nu_{a,b}$ has the point mass at [math] with mass $(1-a)$. The density for the Lebesgue absolutely continuous part of $\nu_{a,b}$ is given by

[TABLE]

Hence we have

[TABLE]

where

[TABLE]

Remark 3.5**.**

The free $F$-distribution introduced above is not very new one but it has been known as the limit spectral distribution of a multivariate $F$-matrix (random Fisher matrix) which is one of the important models of random matrices, see Chapter 4 in [2]. In our setting, their result can be stated as follows: Let the random matrix ${\bm{F}}={\bm{S}}_{n_{1}}{\bm{S}}_{n_{2}}^{-1}$, where ${\bm{S}}_{n_{i}}$ $(i=1,2)$ is a sample covariance matrix with dimension $p$ and sample size $n_{i}$ with an underlying distribution of mean [math] and variance $1$. If ${\bm{S}}_{n_{1}}$ and ${\bm{S}}_{n_{2}}$ are independent and we take the limit with the asymptotic ratios $p/n_{1}\to 1/a\in(0,\infty)$ and $p/n_{2}\to 1/b\in(0,1)$. Then the limit spectral distribution of the random matrix ${\bm{F}}$ is given by the free $F$-distribution $\mathrm{f}F(a,b)$. **

3.3 The free $T$-distribution

For the free $F$-distribution $\mathrm{f}F(a,b)$, we consider the critical case that $a=1$, which corresponds to the free analogue of one-dimensional $T^{2}$-distribution. Because in classical probability, it is known that if a random variable $Y$ is distributed according to the $T$-distribution of degree of freedom $m$, then the random variable $X=Y^{2}$ is $F(1,m)$-distributed.

Hence it is natural to understand that the distribution of a symmetric random variable $Y$ such that $Y^{2}$ has the free $F$-distribution $\mathrm{f}F(1,m)$ is the free $T$ -distribution of degree of freedom $m$. Here the fact that $Y$ is symmetric means that all the odd moments $m_{2k-1}(Y)$ $=E(Y^{2k-1})$, $k\geq 1$ vanish.

Proposition 3.6**.**

We assume $m>1$, and let $Y$ be a symmetric random variable such that the distribution of $Y^{2}$ is the free $F$-distribution $\mathrm{f}F(1,m)$. Then $G_{Y}(z)$, the Cauchy transform of the distribution of $Y$, is given by

[TABLE]

Proof.

Let $M_{Y}(z)$ and $M_{Y^{2}}(z)$ be the moment generating functions of the random variables $Y$ and $Y^{2}$, respectively. Since $Y$ is symmetric, these generating functions satisfy the relation

[TABLE]

which implies the relation between the Cauchy transforms that

[TABLE]

By the assumption the Cauchy transform $G_{Y^{2}}(z)$ can be obtained by one for the scaled free beta prime distributed random variable $mX(1,m)$

[TABLE]

where $G_{X\!(1,m)}$ is the Cauchy transform of the free beta prime distribution $\mathrm{f}{\bm{\beta}}^{\prime}(1,m)$ given by (1) in Proposition 3.2. Hence we have

[TABLE]

∎

Applying Stieltjes inversion formula to $G_{Y}(z)$, we can easily obtain the probability measure of the random variable $Y$, which is absolutely continuous with respect to Lebesgue measure. This measure is our desired free $T$-distribution.

Definition 3.7**.**

For $m>1$, the free $T$-distribution $\mathrm{f}T(m)$ of parameter $m$ is the compactly supported probability measure on \big{[}-\mbox{\scriptsize\dfrac{2m}{m-1}},\mbox{\scriptsize\dfrac{2m}{m-1}}\big{]} with the density

[TABLE]

Remark 3.8**.**

The density function of the free $T$-distribution $\mathrm{f}T(m)$ has the following limits:

[TABLE]

In classical probability, it is known that the $T$-distribution $T(m)$ of the parameter $m$ becomes the standard Gaussian in the limit $m\to\infty$, and the Cauchy distribution can be obtained as the special case of $m=1$.

In the limit $m\to\infty$, the density of the free $T$-distribution $\mathrm{f}T(m)$ tends to the standard semicircle law, the free counterpart of the classical standard Gaussian. Since it is known from the theory of free stable laws [4] that the free counterpart of the classical Cauchy distribution is given by the Cauchy distribution itself, hence, we can state that the density of the free $T$-distribution $\mathrm{f}T(m)$ becomes the free Cauchy distribution when $m$ goes to $1$. Thus the above limits are compatible with those in classical probability. **

3.4 The free beta distribution

In classical probability, it is well known that if $Y$ is distributed according to the beta distribution ${\bm{\beta}}(a,b)$ then $\dfrac{Y}{1-Y}$ is the beta prime ${\bm{\beta}}^{\prime}(a,b)$-distributed random variable, or equivalently that if $X$ is distributed according to the beta prime distribution ${\bm{\beta}}^{\prime}(a,b)$ then $\dfrac{X}{1+X}$ is the beta ${\bm{\beta}}(a,b)$-distributed random variable.

Based on this fact, it is natural to introduce the free beta distribution as follows: Let $X$ be a self-adjoint random variable in a $C^{*}$-probability space distributed according to the free beta prime distribution $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$. Then we will regard the distribution of the reciprocal B=\big{(}\bm{1}+X^{-1}\big{)}^{-1}, as the free beta distribution.

For simplicity, we first deal with the case of $a>1$ and $b>1$ so that no atomic parts appear in $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$ or $\mathrm{f}{\bm{\beta}}^{\prime}(b,a)$.

Proposition 3.9**.**

We assume $a>1$ and $b>1$, and let $X$ be a free beta prime $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$-distributed self-adjoint random variable in a $C^{*}$-probability space. Then the Cauchy transform of the distribution of B=\big{(}\bm{1}+X^{-1}\big{)}^{-1} is given by

[TABLE]

where the branch of the square root is chosen so that $\displaystyle{\lim_{z\to\infty}G_{B}(z)=0}$.

Proof.

We note that $X^{-1}$ is a free beta prime $\mathrm{f}{\bm{\beta}}^{\prime}(b,a)$-distributed random variable. Hence the Cauchy transform $G_{X^{-1}}(z)$ of the distribution of ${X^{-1}}$ is given by the formula (1) in Proposition 3.2 with exchanging $a$ with $b$, that is,

[TABLE]

On the other hand, we easily find (see, for instance, [2]) that if $W$ is a strictly positive random variable in a $C^{*}$-probability space with compact support, then $W$ is invertible and the Cauchy transform of $W^{-1}$ is given by the formula

[TABLE]

Now we apply this formula to the random variable $W=\bm{1}+X^{-1}$ with $G_{W}(z)=G_{X(b,a)}(z-1)$ since $W$ is the shift of $X^{-1}$ by $\bm{1}$.

Combining the formulas above, we can have

[TABLE]

which yields our desired formula. ∎

We introduce the free beta distribution by the Cauchy transform $G_{B}(z)$. Although we have derived $G_{B}(z)$ under the condition $a>1$ and $b>1$, it can be found that $G_{B}(z)$ is still valid for \big{\{}(a,b)\,|\,a>0,\,b>0,\,\mbox{and}\,a+b>1\big{\}}.

Applying Stieltjes inversion formula, one can recover the probability measure as follows: Concerning the point mass of the measure, the function $G_{B}(z)$ has the simple poles at $z=0$ and $z=1$, and the residues are calculated as

[TABLE]

respectively. Thus the measure has a point masses $(1-a)$ at [math] if $0<a<1$ and $(1-b)$ at $1$ if $0<b<1$.

The density $f_{\mbox{\scriptsize f}{\bm{\beta}}}(x;a,b)$ of the Lebesgue absolutely continuous part is given by

[TABLE]

which is supported on the interval $[\kappa_{-},\kappa_{+}]$, where \displaystyle{\kappa_{\pm}=\bigg{(}\mbox{\small\dfrac{\sqrt{a(a+b-1)}\pm\sqrt{b}}{a+b}}\bigg{)}^{\!2}} are two real roots of the quadratic equation $(a+b)^{2}x^{2}-2(a^{2}+ab-a+b)x+(a-1)^{2}=0$ and satisfy $[\kappa_{-},\kappa_{+}]\subseteq[0,1]$.

Definition 3.10**.**

Let the parameters $a$ and $b$ satisfy $a>0$, $b>0$, and $a+b>1$. The free beta distribution ${\mathrm{f}{\bm{\beta}}}(a,b)$ is the probability measure

[TABLE]

where \displaystyle{\kappa_{\pm}=\bigg{(}\mbox{\small\dfrac{\sqrt{a(a+b-1)}\pm\sqrt{b}}{a+b}}\bigg{)}^{\!2}}. **

4 The rationales of the free analogue

4.1 The first rationale of the free analogue

The standard semi-circular distribution can be regarded as the free analogue of the standard normal distribution, and the square of a standard semi-circular element has the distribution of the free Poisson of parameter $1$. Hence it is natural to regard the probability distribution of the freely independent sum of $m$-many squares of standard semi-circular elements as the free $\chi^{2}$-distribution of degree of freedom $m$, which is given by the free Poisson distribution of parameter $m$.

In classical probability, the $\chi^{2}$-distribution is in a class of gamma distributions, and the ratio of independent gamma distributed random variables gives the beta prime distribution. Indeed, if $X_{1}$ and $X_{2}$ are independent and distributed according to the gamma distributions $\Gamma(a,\theta)$ and $\Gamma(b,\theta)$, respectively, where both have the same scaling parameter $\theta$, then the random variable $X_{1}/X_{2}$ has the beta prime distribution ${\bm{\beta}}^{\prime}(a,b)$.

From this standpoint, the free beta prime distribution introduced in the previous section can be regarded as the distribution of a ratio of freely independent free gamma distributed random variables because the free Poisson random variables have the free gamma distributions in the sense of free $\chi^{2}$-distributions. This is the first naive rationale of our free analogue of the beta prime distribution.

4.2 The score functions and the potentials

The second rationale is related to the score functions for the Fisher informations, in other words, the potentials of the diffusion processes. Here we will briefly recall the score functions both in the classical and the free cases.

In classical probability, for a probability measure $\mu$ with the differentiable density $f$, the function

[TABLE]

is called the score function of $\mu$. Then the classical Fisher information $I(\mu)$ of $\mu$ (with respect to the location parameter) can be given by the square of the $L^{2}$-norm of the score function $||\rho_{\mu}||^{2}$ in $L^{2}(d\mu)$, that is,

[TABLE]

One of the basic properties of the score function is the following Stein’s relation: For a smooth function $p$, applying integration by parts, the relation

[TABLE]

holds.

On the other hand, the free analogues of entropy and Fisher information for self-adjoint non-commutative random variables were introduced and begun to study in Voiculescu’s paper [41] (see, for survey, [42]). In the univariate case, the free Fisher information $\Phi(\mu)$ of a compactly supported probability measure $\mu$ is given by

[TABLE]

where \big{(}{\mathcal{H}}f\big{)}(x) is the ($\pi$-multiplicated) Hilbert transform of $f$ defined by the principal value integral

[TABLE]

By the expression of the free Fisher information, it is natural to understand that the double of the Hilbert transform 2\big{(}{\mathcal{H}}f\big{)}(x) corresponds to a free analogue of the classical score function, because the square of the $L^{2}$-norm ||2\big{(}{\mathcal{H}}f\big{)}||^{2} in $L^{2}(d\mu)$ is the free Fisher information of $\mu$.

Indeed the function 2\big{(}{\mathcal{H}}f\big{)}(x) satisfies the following identity (see, for instance, [27]): Let $\mu$ be a compactly supported probability measure on ${\mathbb{R}}$ with continuous density $f$, and assume that $\mu$ has finite free Fisher information. Then, for a continuously differentiable function $\eta$ on ${\mathbb{R}}$, we have

[TABLE]

where ${\mathcal{S}}={\mathrm{Supp}}(\mu)$, which can be regarded as the free counterpart of the classical Stein’s relation because the difference quotient D\eta=\mbox{\small\dfrac{\eta(x)-\eta(y)}{x-y}} works as non-commutative derivative.

Comparing the free Stein’s relation with the classical one, one can find that the sign of the free score function is opposite to the classical one, which is, however, compatible from the viewpoint of the potentials in diffusion processes.

For the function $V\in C^{1}({\mathbb{R}})$,

[TABLE]

is called the Gibbs distribution of the potential $V$, which is obtained as the long-time asymptotically stationary distribution for the diffusion process on ${\mathbb{R}}$ with the drift potential $V$. It is clear that the classical score function of the Gibbs distribution $g$ is given by $\dfrac{g^{\prime}(x)}{g(x)}=-V^{\prime}(x)$.

On the other hand, in the free context, Biane and Speicher (see [5] and [6]) investigated the free analogue of diffusion process via random matrix models. They derived that the long-time asymptotically stationary measure \nu_{\mbox{\tinyV}} for the free diffusion process with the potential $V\in C^{1}({\mathbb{R}})$ is characterized by the Euler-Lagrange equation

[TABLE]

where $g$ is the compactly supported density g(x)dx=d\mbox{\nu_{\mbox{\tiny$V$}}}(x).

This stationary measure \nu_{\mbox{\tinyV}} is called the equilibrium measure for the free diffusion process with the potential $V$. It is obvious that the free score function of the equilibrium measure \nu_{\mbox{\tinyV}} is given by 2\big{(}{\mathcal{H}}g\big{)}(x)=V^{\prime}(x) just like with the Gibbs distribution.

The classical and the free score functions mentioned above, namely, the derivative of the potential $V^{\prime}(x)$ gives the second rationale of our free analogues.

4.3 The second rationale of the free analogue

We shall see that the free beta prime distribution and the free analogues of the classical $T$ and beta distributions derived in the previous section can be characterized by exactly the same (up to constant) potentials for the Gibbs forms of corresponding classical distributions.

The following formula is helpful for us to find the ($\pi$-multiplicated) Hilbert transform of the probability measure on ${\mathbb{R}}$ (see, for instance, Chapter 3 in [22]): For a compactly supported probability measure $\mu$ on ${\mathbb{R}}$, the ($\pi$-multiplicated) Hilbert transform \big{(}{\mathcal{H}}\mu\big{)} can be obtained by the formula,

[TABLE]

where $G_{\mu}$ is the Cauchy transform of $\mu$.

We shall list the corresponding potentials of our free analogues below.

(i) The free beta prime distribution:

• $\bullet$

The density of the classical beta prime distribution ${\bm{\beta}}^{\prime}(a,b)$ and its potential:

[TABLE]

for $x>0$, where ${B}$ is the beta function.

• $\bullet$

The derivative of the potential:

[TABLE]

• $\bullet$

The free score function of the free beta prime distribution ${\mathrm{f}{\bm{\beta}}^{\prime}(a,b)}$:

The Cauchy transform of ${\mathrm{f}{\bm{\beta}}^{\prime}(a,b)}$ is given by (1) in Proposition 3.2 and its free score function is

[TABLE]

(ii) The free $T$-distribution:

• $\bullet$

The density of the classical $T$-distribution $T(m)$ and its potential:

[TABLE]

• $\bullet$

The derivative of the potential:

[TABLE]

• $\bullet$

The free score function of the free $T$-distribution ${\mathrm{f}T(m)}$:

The Cauchy transform of ${\mathrm{f}T(m)}$ is given by (3) in Proposition 3.6 and its free score function is

[TABLE]

(iii) The free beta distribution:

• $\bullet$

The density of the classical beta distribution ${\bm{\beta}}(a,b)$ and its potential:

[TABLE]

for $0<x<1$, where ${B}$ is the beta function.

• $\bullet$

The derivative of the potential:

[TABLE]

• $\bullet$

The free score function of the free beta distribution ${\mathrm{f}{\bm{\beta}}(a,b)}$:

The Cauchy transform of ${\mathrm{f}{\bm{\beta}}(a,b)}$ is given by (4) in Proposition 3.9 and its free score function is

[TABLE]

Remark 4.1**.**

Concerning the second rationale, it should be noted that Hasebe and Szpojankowski pointed out such a correspondence between the classical and the free distributions in [21] based on the maximization problem of the entropy functionals with an external potential $V$.

In particular, they showed the correspondence between the classical and the free generalized inverse Gaussian distributions. They also mentioned the maps from the classical Gaussian to the semicircle and from the classical gamma to the free Poisson. Hence the free analogue of the classical distributions derived in the previous section can be appended as new examples. **

5 The combinatorial representation of the moments

5.1 Non-crossing linked partitions and the Motzkin paths

In order to describe the moments in combinatorial way, we shall use the set partitions. For the set $[n]=\{1,2,\ldots,n\}$, a partition of $[n]$ is a collection $\pi=$ $\{B_{1},B_{2},$ $\ldots,$ $B_{k}\}$ of non-empty disjoint subsets of $[n]$ which are called blocks and whose union is $[n]$. For a block $B$, we denote by $|B|$ the size of the block $B$, that is, the number of the elements in the block $B$. A block $B$ will be called singleton if $|B|=1$.

We say two blocks $B_{i}$ and $B_{j}$ in $\pi$ are crossing if there exist elements $b_{1},b_{2}\in B_{i}$, $c_{1},c_{2}\in B_{j}$ such that $b_{1}<c_{1}<b_{2}<c_{2}$. The blocks $B_{i}$ and $B_{j}$ are said to be non-crossing if they are not crossing. A partition $\pi$ is called non-crossing if its blocks are pairwise non-crossing. We denote the set of all non-crossing partitions of the set $[n]$ by $\mathcal{NC}(n)$. The notion of non-crossing partition was first introduced in [24]. For more about non-crossing partitions, see the survey of Simion [32].

In the context of free probability, Dykema introduced a new structure of partitions, the non-crossing linked partitions in [12] (see also [29]), which can be regarded as a non-crossing partition having some links between blocks with certain restrictions. The restricted link between blocks introduced by Dykema in [12] is as follows:

• Let $E$ and $F$ be subsets of $[n]$, We say that $E$ and $F$ are nearly disjoint if for every $i\in E\cap F$, one of the following holds:

[TABLE]

He derived the structure of non-crossing linked partitions in his study on the multiplicative free convolution and the $T$-transform.

Definition 5.1**.**

A non-crossing linked partition of $[n]=\{1,2,\ldots,n\}$ is a collection $\pi$ of non-empty subsets of $[n]$ whose union is $[n]$, and any two distinct elements of $\pi$ are non-crossing and nearly disjoint. We denote by $\mathcal{NCL}(n)$ the set of all non-crossing linked partitions of $[n]$.

For $\pi\in\mathcal{NCL}(n)$, although elements of $\pi$ are not disjoint in general, we refer to an element of $\pi$ as a block of $\pi$. **

Here, we recall some terminologies and the basic properties of non-crossing linked partitions observed in [12].

Remark 5.2**.**

• • (1)

Any given $k\in[n]$ belongs to either exactly one block or exactly two blocks; we will say $k$ is singly or doubly covered by $\pi$, accordingly.

• (2)

The elements $1$ and $n$ are singly covered by $\pi$.

• (3)

Any two distinct elements $E$ and $F$ of $\pi$ have at most one element in common. Moreover, if $E\cap F\neq\phi$ (that is, $|E\cap F|=1$), then both $E$ and $F$ have at least two elements.

One of graphical representations of non-crossing linked partitions has been described in [12], which is a modification of the usual pictures of non-crossing partitions in the following way: The non-crossing partitions $\pi\in\mathcal{NC}(n)\subseteq\mathcal{NCL}(n)$ are drawn in the usual way with all angles being right angles. Suppose $\pi\in\mathcal{NCL}(n)\!\setminus\!\mathcal{NC}(n)$. If $E,F\in\pi$ with $E\neq F$, and if $E\cap F={\min(E)}$, then the line connecting $\min(E)$ to the next element in $E$ is started with diagonal line, that is, the diagonally started lines indicate the links.

Example 5.3**.**

The non-crossing linked partition

[TABLE]

has the graphical representation,

[TABLE]

There are three doubly covered elements, $2$, $7$, and $9$. **

In [11], Chen, Wu, and Yan proposed another graphical representation of non-crossing linked partitions, called the linear representation, which is defined as follows: Given a non-crossing linked partition $\pi$ of $[n]$, list $n$ vertices in a horizontal line with labels $1$, $2$, $\ldots$, $n$. For each block $E=\{i_{1},i_{2},\ldots,i_{\ell}\}$ of $\pi$ with $i_{1}=\min(E)$ and $\ell\geq 2$, draw an arc between $i_{1}$ and $i_{j}$ for each $j=2,\ldots,\ell$, where we should always put the arc $(i,j)$ above the arc $(i,k)$ if $j>k$.

Using the linear representation, they constructed a bijection between non-crossing linked partitions and Schröder paths and derived various enumerative results on non-crossing linked partitions (see, for details, [11]).

Here we shall introduce another graphical representation of non-crossing linked partitions, namely, the card arrangements which is essentially the same as the above graphical representation of Dykema, but we should much more consider the heights of lines in order to reveal the relation between non-crossing linked partitions and the Motzkin paths.

Definition 5.4**.**

A Motzkin path of length $n$ is a non-negative lattice path from $(0,0)$ to $(n,0)$ in the integer lattice ${\mathbb{Z}}_{\geq 0}\times{\mathbb{Z}}_{\geq 0}$ consisting of three types of steps:

$u=(1,1)$: up step,   $d=(1,-1)$: down step,   $t=(1,0)$: transit step. **

The set partitions are closely related to the Motzkin paths. Indeed Flajolet investigated the correspondence between all partitions of $n$ elements and colored (integer labeled) Motzkin paths of length $n$ in [18].

In the paper [44], the representation of non-crossing partitions by cards arrangements associated with the Motzkin paths was shown, which was the similar technique to the juggling patterns in [13] but they were required to prepare different kinds of cards. Namely, they used the opening, the closing, the intermediate, and the singleton cards in order to represent non-crossing partitions. We list these cards below and illustrate the representation of non-crossing partitions by cards again for our convenience.

The opening cards:

The opening card $O_{i}\;(i=0,1,2,\ldots)$ has $i$ inflow lines from the left and $(i\!+\!1)$ outflow lines to the right, where one new line starts from the middle point on the ground level. For each $j\geq 1$, the inflow line of the $j$th level goes through out to the $(j\!+\!1)$st level without any crossing. The card $O_{i}$ is called the opening card of level $i$.

[TABLE]

The opening card represents the minimal element of a block of non-singleton.

The closing cards :

The closing card $C_{i}\;(i=1,2,3,\ldots)$ has $i$ inflow lines from the left and $(i\!-\!1)$ outflow lines to the right. On the card $C_{i}$, only the line of the lowest level goes down to the middle point on the ground level and ends. For each $j\geq 2$, the inflow line of the $j$th level goes through out to the $(j\!-\!1)$st level without any crossing. The card $C_{i}$ is called the closing card of level $i$.

[TABLE]

The closing card represents the maximal element in a block of non-singleton.

The intermediate cards :

The intermediate card $I_{i}\;(i=1,2,3,\ldots)$ has $i$ inflow lines and the same number of outflow lines. Only the line of the lowest level goes down to the middle point on the ground and continues its flow as the lowest line again. The rest of inflow lines maintain their levels. We call the card $I_{i}$ the intermediate card of level $i$.

[TABLE]

The intermediate card represents the intermediate (neither the minimal nor the maximal) element of a block of size $\geq 3$.

The singleton cards :

The singleton card $S_{i}\;(i=0,1,2,\ldots)$ has $i$ horizontally parallel lines and the short pole at the middle point on the ground. We call the card $S_{i}$ the singleton card of level $i$, which represents, of course, a singleton.

[TABLE]

Associated with a Motzkin path, we arrange the above cards according to the following rule:

The rule of the arrangements for ${\mathcal{NC}}(n)$

Let \mathbf{p}=\big{(}s_{1},s_{2},...,s_{n}\big{)} be a Motzkin path of length $n$ where $s_{j}\in\{u,d,t\}$, and denote by $y_{j}$ the height of the step $s_{j}$ starting.

• (1)

In case of $s_{j}=u$, up step (resp. $s_{j}=d$, down step), we put the opening (resp. closing) card of level $y_{j}$ at the $j$th site.

• (2)

In case of $s_{j}=t$, transit step, if the height $y_{j}\geq 1$ (not at ground level) then two cards are available, namely, the intermediate card of level $y_{j}$ or the singleton card of level $y_{j}$, but if $y_{j}=0$ (at the ground level) then we have to put $S_{0}$, the singleton card of level [math], at the $j$th site.

The card arrangements constructed by the rule above are called the admissible arrangements. Each Motzkin path yields not only one admissible arrangement in general but each admissible arrangement determines the non-crossing partition of $[n]$ uniquely, the blocks of which are constituted from the connected curves in the pattern on the admissible arrangement (see, for more details, [44]).

Now we shall extend the above representation to the case of non-crossing linked partitions. To this end, we introduce some more cards which will represent the doubly covered elements.

Before making the cards, we classify the doubly covered elements into two types. It is from the definition that a doubly covered element is contained in two blocks and the minimal element of one or the other. Let $k$ be a doubly covered element such that $k\in E\cap F$ with $k=\min(F)$.

• (i)

If $k=\max(E)$, then we call $k$ the doubly covered element of type I.

• (ii)

Otherwise, in the case where $k\neq\max(E)$, that is, $k$ is an intermediate element of $E$ ($k\neq\min(E)$ follows by definition) then $k$ is said to be the doubly covered element of type II.

Here we will construct the cards for the doubly covered elements according to these types.

The cards for doubly covered elements of type I :

We shall make the cards $T_{i}\;(i=1,2,3,\ldots)$ for doubly covered elements of type I. The card $T_{i}$ has $i$ inflow lines and the same number of outflow lines. Only the inflow line of the lowest level goes down to the middle point on the ground and ends, which indicates the end of a block. But immediately a new line starts with $\pi/4$ - slope from the same middle point on the ground level and will be the outflow line of the lowest level. The rest of inflow lines maintain their levels.

[TABLE]

On the card $T_{i}$, the $\pi/4$ - slope indicates the beginning of a new block, which means a doubly covered element of type I.

The cards for doubly covered elements of type II :

We prepare the cards $U_{i}\;(i=1,2,3,\ldots)$ for doubly covered elements of type II. The card $U_{i}$ has $i$ inflow lines from the left and $(i\!+\!1)$ outflow lines to the right, where only the inflow line of the lowest level goes down to the middle point on the ground and continues its flow as the second (not the lowest) level again. This connected flow indicates an intermediate element. Immediately a new line starts with $\pi/4$ - slope from the same middle point on the ground level and will be the outflow line of the lowest level, which indicates the beginning of a new block and, hence, the card $U_{i}$ represents a double covered element of type II.

[TABLE]

We shall give the representation of non-crossing linked partitions by the arrangements of cards, which is almost the same as for non-crossing partitions but more cards are available at some steps.

The rule of the arrangements for ${\mathcal{NCL}}(n)$

Let \mathbf{p}=\big{(}s_{1},s_{2},...,s_{n}\big{)}\; $s_{j}\in\{u,d,t\}$ be a Motzkin path of length $n$ and denote by $y_{j}$ the height of the step $s_{j}$ starting. Associated with the Motzkin path, we shall arrange the cards along with the following rule:

• (1)

In case of $s_{j}=u$ (up step), if the height $y_{j}=0$ (at ground level) then we put the opening card of level [math], and if $y_{j}\geq 1$ (not at ground level) then we put the card for doubly covered elements of type II or the opening card of level $y_{j}$ at the $j$th site, that is, two cards are available.

• (2)

In case of $s_{j}=t$ (transit step), if the height $y_{j}\geq 1$ (not at ground level) then three cards are available, namely, doubly covered elements of type I of level $y_{j}$, or the intermediate card of level $y_{j}$, or the singleton card of level $y_{j}$, but if $y_{j}=0$ (at ground level) then we have to put the singleton card of level [math] at the $j$th site, that is, there is no possibility other than $S_{0}$.

• (3)

If $s_{j}=d$ (down step), then we put the closing card of level $y_{j}$ at the $j$th site, which is unique possibility.

We again call the card arrangements constructed by the new rule above the admissible arrangements again. Of course, the number of the new admissible arrangements are rather increased compared with the case of non-crossing partitions. Similar to the case of non-crossing partitions, each admissible arrangement determines the non-crossing linked partition uniquely, the blocks of which are constituted from the connected lines in the pattern on the admissible arrangement, where we regard that the lines starting with the $\pi/4$ - slope are not connected to curved lines at doubly covered elements.

Conversely, it can be said that every non-crossing linked partition is represented as above admissible arrangement (recall the graphical representation of Dykema in Example 5.3).

Example 5.5**.**

If the Motzkin path $\mathbf{p}=(u,u,t,d,d)$,

[TABLE]

then we obtain $6$ admissible arrangements because at the second step $u$, we can use one of two cards $O_{1}$, $U_{1}$, and at the third step $t$, three cards $I_{2}$, $S_{2}$, $T_{2}$ are available. We shall list the $6$ admissible arrangements and the corresponding non-crossing liked partitions below:

[TABLE]

[TABLE]

[TABLE]

5.2 Enumeration and the weighted Motzkin paths

We give the generating function of the enumerating polynomials for some set partition statistics in non-crossing linked partitions by using the weighted Motzkin paths.

If each step in a Motzkin path has a weight then it is called the weighted Motzkin path.

Consider the three sequences

[TABLE]

which we use as the weights for the up, the down, and the transit steps in a Motzkin path, respectively.

Let $\mathbf{p}=(s_{1},s_{2},...,s_{n})$ be a Motzkin path of length $n$ where $s_{j}\in\{u,d,t\}$. We make the associated list of weights $\mathbf{w}_{\mathbf{p}}=(w_{1},w_{2},...,w_{n})$ depending on both the type and the height of each step as follows:

[TABLE]

where $y_{j}$ is the height of the step $s_{j}$.

Given a weighted Motzkin path $\mathbf{p}$ of length $n$, the weight of the Motzkin path $wt(\mathbf{p})$ is defined by ${\displaystyle wt(\mathbf{p})=\!\prod_{j=1}^{n}\!w_{j}}$, the product of the weights in $\mathbf{w}_{\mathbf{p}}$.

The following well-known formula on the generating function can be derived with help of the results in [18].

Theorem 5.6**.**

Let $\mathcal{M}_{n}$ be the set of all Motzkin paths of length $n$ and assume each of the paths is weighted in the manner above. We write the sum of all the weights of the Motzkin paths in $\mathcal{M}_{n}$ by

[TABLE]

which is sometimes called the $n$th moment. Then its generating function

[TABLE]

can be expanded into the continued fraction of the form

[TABLE]

This formula enables us to give the generating function of the enumerating polynomials for some set partition statistics in non-crossing linked partitions via the weighted Motzkin paths.

Definition 5.7**.**

For a non-crossing linked partition $\pi\in\mathcal{NCL}(n)$, we shall introduce the following set partition statistics:

• $dc(\pi)$:

the number of doubly covered elements by $\pi$,

• $sc(\pi)$:

the number of singly covered minimal elements by $\pi$, but non-singleton,

• $sg(\pi)$:

the number of singletons in $\pi$.

Example 5.8**.**

For the partition $\pi$ in Example 5.3, each value of the statistics above becomes $dc(\pi)=3$, $sc(\pi)=2$, and $sg(\pi$) = 1. **

Remark 5.9**.**

We should note that the relation

[TABLE]

holds because the minimal element of each block of $\pi\in\mathcal{NCL}(n)$ falls into one of the above three statistics, where $|\pi|$ stands for the number of blocks in $\pi$.

Moreover, only the doubly covered elements are double-counted, thus we have the equality

[TABLE]

We encode the joint statistics $(dc,sc,sg)$ in $\mathcal{NCL}(n)$ by $(\alpha,\beta,\gamma)$, and write the generating function of the enumerating polynomials in $\alpha$, $\beta$, and $\gamma$ as

[TABLE]

Here we assign the weights to the cards and consider the weighted cards arrangements, which yield the weighted Motzkin paths.

How to assign the weight to the cards is simple, that is, we will assign the weights $\alpha$, $\beta$, and $\gamma$ to the cards that correspond to the set partition statistics $dc$, $sc$, and $sg$, respectively. The cards that do not contribute to any statistics should be weighted $1$.

The weight of the cards

• ($\alpha$)

The doubly covered elements correspond to the cards $T_{i}\,(i\geq 1)$ and $U_{i}\,(i\geq 1)$. Thus these cards have the weight $\alpha$.

• ($\beta$)

The singly covered minimal elements of non-singleton correspond to the opening cards $O_{i}\,(i\geq 0)$. Thus these cards have the weight $\beta$.

• ($\gamma$)

The singletons, of course, are represented by the singleton cards $S_{i}\,(i\geq 0)$ which have the weight $\gamma$.

• ($1$)

We assign the weight $1$ to the intermediate cards $I_{i}\,(i\geq 1)$ and the closing cards $C_{i}\,(i\geq 1)$, which do not correspond to any statistics.

Then we define the weight of an admissible arrangement as the product of the weights of the cards used in it.

For instance, the non-crossing linked partition in Example 5.3 can be represented by the admissible cards arrangement

[TABLE]

[TABLE]

Then the weight of the arrangement is given by

[TABLE]

The weights of the steps

Now we can determine the weights of the steps in the weighted Motzkin paths for our statistics according to the rule of the arrangements for ${\mathcal{NCL}}(n)$ as follows:

• ($u$)

For the up step at the ground level, only the opening card $O_{0}$ is available, thus $\mu_{0}=\beta$. While for the up step at the height $i\geq 1$, the opening card $O_{i}\,(i\geq 1)$ and the card for the doubly covered elements of type II $U_{i}\,(i\geq 1)$ are available, thus $\mu_{i}=\alpha+\beta\,(i\geq 1)$.

• ($t$)

For the transit step at the ground level, we can use only the singleton card $S_{0}$, thus $\kappa_{0}=\gamma$. For the up step at the height $i\geq 1$, however, we can use the intermediate card $I_{i}\,(i\geq 1)$, the card for doubly covered elements of type I $T_{i}\,(i\geq 1)$, and the singleton card $S_{i}\,(i\geq 1)$, thus $\kappa_{i}=1+\alpha+\gamma\,(i\geq 1)$.

• ($d$)

For the down step at the height $i\geq 1$, only the closing card $C_{i}\,(i\geq 1)$ is available, thus $\lambda_{i}=1\,(i\geq 1)$.

Consequently, the weights of the steps for the weighted Motzkin path for our joint statistics $(\alpha,\beta,\gamma)$ should be given as follows:

[TABLE]

Example 5.10**.**

Each admissible arrangement for the Motzkin path $\mathbf{p}=(u,u,t,d,d)$ in Example 5.5 has the following weight:

[TABLE]

The sum of the above $6$ weights of the admissible arrangements is

[TABLE]

that is, the weight of the Motzkin path $\mathbf{p}=(u,u,t,d,d)$. **

From Theorem 5.6, it follows that the generating function $\Gamma(z;\alpha,\beta,\gamma)$ can be obtained by the continued fraction of the form

[TABLE]

which is rewritten as

[TABLE]

where the recursive part $h(z;\alpha,\beta,\gamma)$ satisfies the relation

[TABLE]

Eliminating $h(z;\alpha,\beta,\gamma)$, we have the following formula:

Theorem 5.11**.**

The generating function $\Gamma(z)=\Gamma(z;\alpha,\beta,\gamma)$ satisfies the quadratic equation,

[TABLE]

and its closed form is solved as

[TABLE]

5.3 Combinatorial moment formulas

Related to the additive free convolution, the remarkable relation between the moments and the free cumulants (the coefficients of $R$-transform) was discovered by Speicher in [33], which is known as the free moment-cumulant formula.

Theorem 5.12**.**

([33])* Let $\mu$ be a compactly supported probability measure on ${\mathbb{R}}$ and denote its $R$-transform by*

[TABLE]

Then the $n$th moment of $\mu$ can be given by the combinatorial formula

[TABLE]

where $\mathcal{NC}(n)$ is the set of non-crossing partitions of $[n]$.

Like in the additive case, related to the multiplicative free convolution, the combinatorial formula of the moments by the coefficients of $T$-transform was found by Dykema in [12] (see also [29]).

Theorem 5.13**.**

([12])* Let $\mu$ be a compactly supported probability measure on $[0,\infty)$ with non-zero mean and denote its $T$-transform by*

[TABLE]

then the $n$th moment of $\mu$ can be given by the combinatorial formula

[TABLE]

where $\mathcal{NCL}(n)$ is the set of non-crossing linked partitions of $[n]$.

Remark 5.14**.**

Although $\mathcal{NCL}(n)$ can not exactly make a lattice, the moments and the coefficients of $T$-transform determine each other in the same manner as in the additive free moment-cumulant formula. **

At the end of this section we introduce the distinguished operator on the full space $\mathcal{T}(\mathcal{H})$ whose moments are closely related to the non-crossing linked partitions.

Proposition 5.15**.**

Let $\mathcal{H}$ be a Hilbert space and $\mathcal{T}(\mathcal{H})$ be a full Fock space of $\mathcal{H}$. For a unit vector $\xi\in\mathcal{H}$, let $\ell=\ell(\xi)$ and $\ell^{*}=\ell(\xi)^{*}$ be the left creation and the left annihilation operators on a full Fock space $\mathcal{T}(\mathcal{H})$, respectively. We set the operator $X$ in \mathcal{B}\big{(}\mathcal{T}(\mathcal{H})\big{)} as

[TABLE]

Then the $n$th moment of the operator $X$ with respect to the vacuum expectation $\varphi$ is given by

[TABLE]

Hence the moment generating function of the random variable $X$ is given by $\Gamma(z;\alpha,\beta,\gamma)$ in Theorem 5.11.

Proof.

We decompose the operator $X$ and set $u$, $d$, $t$ as follows:

[TABLE]

Then the operators $u$, $d$, and $t$ act on the elementary vectors $\xi^{\otimes n}\in\mathcal{T}(\mathcal{H})$ $(n\geq 0)$ that

[TABLE]

where $\xi^{\otimes 0}$ conventionally means the vacuum vector $\Omega$. By using the weights for the weighted Motzkin paths in (7), we can simply write

[TABLE]

Now we expand $X^{n}=(u+d+t)^{n}$ into $3^{n}$ non-commutative monomials in $s$, $d$, and $t$ as

[TABLE]

then only the monomials of the Motzkin path type could contribute to the vacuum expectation $\varphi(\,\cdot\,)=\langle\,\cdot\,\Omega|\Omega\rangle$. Indeed, it follows that $s_{n}s_{n-1}\cdots s_{1}\,\Omega\in\mathbb{C}\,\Omega$ if and only if $s_{1}s_{2}\cdots s_{n}$ makes a Motzkin path of length $n$. In this case,

[TABLE]

where $wt(\mathbf{p})$ is the weight of the weighted Motzkin path $\mathbf{p}$. Hence it follows that

[TABLE]

where the weight sequences for the weighted Motzkin paths are given as in (7).

Here we should note that the following correspondence between the operators $u$, $d$, $t$ and the cards for $\mathcal{NCL}(n)$ can be found by decomposition of $u$ and $t$:

[TABLE]

where each coefficient corresponds to the weight of the card. We can also find that only the admissible arrangements of $n$ cards could contribute to the vacuum expectation of $X^{n}$. ∎

Remark 5.16**.**

The operator

[TABLE]

has the same distribution as of the operator $X$, and the shifted free Poisson part is contained in ${\widetilde{X}}$ because \displaystyle{\beta{\bm{1}}+\sqrt{\beta}\big{(}\ell+\ell^{*}\big{)}+\ell\ell^{*}} corresponds to the free Poisson random variable of parameter $\beta$. The operators in such a form of $X$ or ${\widetilde{X}}$ were also investigated by Bożejko and Lytvynov in [9]. **

6 The moments of the free beta prime distribution

In this section, we investigate the moments of the free beta prime distribution $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$ and see that $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$ is in the free Meixner family. Moreover we determine the free Meixner class to which the free beta prime distribution should be classified.

Since we know that the $S$-transform of $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$ is given by (2) in the proof of Proposition 3.2, the $T$-transform of $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$ is obtained as $\displaystyle{T_{\mathrm{f}{\bm{\beta}}^{\prime}(a,b)}(z)=\frac{z+a}{b-1-z}}$, which has the following expansion:

[TABLE]

Thus the coefficients of the $T$-transform become

[TABLE]

In order to simplify the expression, we put

[TABLE]

then we can write $\alpha_{0}=s,\;\;\alpha_{k}=t\,u^{k}\;\;(k\geq 1)$ in short.

Applying the moment formula in Theorem 5.13, the $n$th moment of $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$ is given in a combinatorial form by

[TABLE]

Using the relations (5) and (6) among the set partition statistics $sg$, $sc$, and $dc$ in Remark 5.9, the first product in the most right hand side of (8) can be reformulated as

[TABLE]

It is clear that the second product in (8) becomes

[TABLE]

and that

[TABLE]

Consequently, we obtain the following combinatorial formula of the $n$th moment of $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$.

Theorem 6.1**.**

The $n$th moment of the free beta prime distribution $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$ is given by

[TABLE]

where $\displaystyle{s=\frac{a}{b-1}}$, $\displaystyle{t=\frac{a+b-1}{b-1}}$, and $\displaystyle{u=\frac{1}{b-1}}$, equivalently, it is given by

[TABLE]

A model of the $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$-distributed random variable $X(a,b)$ on a full Fock space $\mathcal{T}(\mathcal{H})$ can be given by the form

[TABLE]

where $\ell=\ell(\xi)$ and $\ell^{*}=\ell(\xi)^{*}$ are the left creation and the left annihilation operators for a unit vector $\xi\in\mathcal{H}$, respectively.

The model of the random variable $X(a,b)$ is an immediate consequence of Proposition 5.15.

A family of the free Meixner distributions contains many important laws in free probability, which has been investigated in many literatures, for instance, [1], [8], [10], [14], and [31].

In particular, Bożejko and Bryc in [8] showed that the family of the standard (mean [math] and variance $1$) free Meixner distributions is parameterized by $\theta$ and $\tau$ as

[TABLE]

with the Cauchy transform

[TABLE]

They also showed that the free Meixner distributions can be classified into six classes, which was inspired by the fact that the classical Meixner distributions satisfy Laha-Lukacs properties with similar parameters.

Here we will see that the free beta prime distribution $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$ is in a family of the free Meixner distributions and determine its class in the free Meixner family.

Let $X(a,b)$ be a free beta prime $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$ distributed random variable. By the moment formula, it is easy to find that $X(a,b)$ has mean m=\displaystyle{\mbox{\small\dfrac{a}{b-1}}} and variance v=\displaystyle{\mbox{\small\dfrac{a(a+b-1)}{(b-1)^{3}}}}.

We shall standardize $X(a,b)$ as $\widetilde{X}(a,b)=\dfrac{X(a,b)-m}{\sqrt{v}}$, since we know the Cauchy transform $G_{X(a,b)}$ as in (1) in Proposition 3.2, the Cauchy transform of $\widetilde{X}(a,b)$ can be obtained by direct calculation as follows:

[TABLE]

If we put

[TABLE]

then it follows that

[TABLE]

which means that the distribution of $\widetilde{X}(a,b)$ is in a family of the free Meixner distribution with the above parameters $\theta$ and $\tau$.

According to the classification table by [8], the distribution of the random variable $\widetilde{X}(a,b)$ is classified into the free negative binomial distributions because the classification parameters satisfy the inequality

[TABLE]

Hence we can conclude the following proposition:

Proposition 6.2**.**

The free beta prime distribution $\mathrm{f}{\bm{\beta}}^{\prime}(a,b)$ is in the class of the free negative binomial distributions and, hence, it is freely infinitely divisible.