A vector of Dirichlet processes

: Random probability vectors are of great interest especially in view of their application to statistical inference. Indeed, they can be used for identifying the de Finetti mixing measure in the representation of the law of a partially exchangeable array of random elements taking values in a separable and complete metric space. In this paper we describe the construction of a vector of Dirichlet processes based on the normalization of an exchangeable vector of completely random measures that are jointly inﬁnitely divisible. After deducing the form of the multivariate Laplace exponent associated to the vector of the gamma completely random mea- sures, we analyze some of their distributional properties. Our attention particularly focuses on the dependence structure and the speciﬁc partition probability function induced by the proposed vector.


Introduction
Random probability measures represent the key ingredient for the actual implementation of Bayesian nonparametric procedures and the Dirichlet process is the most celebrated example. Indeed, their probability distribution plays the role of a prior: this is combined with the data to yield a posterior, or a predictive, distribution that is used to determine either exact or approximate Bayesian inferences on quantities of interest. The most notable advances in the field have been achieved under the assumption that the observations X 1 , . . . , X N are the initial segment of an (ideally) infinite exchangeable sequence of random elements (X i ) i≥1 taking values in some complete and separable metric space X endowed with the Borel σ-algebra X . If P X stands for the set of all probability measures on (X, X ) and P X is the σ-algebra induced by the topology of weak convergence on P X , the well-known de Finetti representation theorem leads to a mixture representation of the finite-dimensional distributions of (X i ) i≥1 of the form P[(X 1 , . . . , X N ) ∈ A] = P X p N (A) Q(dp) ∀A ∈ X N , ∀N ≥ 1 (1) where p N = N 1 p and Q is a probability measure on (P X , P X ) that takes on the name of de Finetti measure of the sequence (X i ) i≥1 . This is equivalent to stating that the X i 's are conditionally independent, namely P[X 1 ∈ A 1 , . . . , X N ∈ One of the most popular applications of nonparametric priors to statistical inference concerns density estimation and is based on hierarchical mixture models where the X i 's are latent variables and Q is the law of a Dirichlet process. See [21]. The dramatic advances in the implementation of Markov Chain Monte Carlo simulation algorithms in the last two decades have, then, made Bayesian nonparametric methods directly applicable to a wide range of real world problems. Moreover, a considerable body of work has been devoted to the proposal of alternatives to the Dirichlet process, i.e. different Q's in (1), for modelling exchangeable random elements. See [20] for a recent review.

Partial exchangeability
The notion of exchangeability characterized through (1) corresponds to a more intuitive idea of homogeneity among the observations. While this seems to be a natural assumption if one aims at prediction, it does not typically correspond to many situations of practical interest where data are originated under different experimental conditions or are, more generally, indexed by a collection of covariates that are relevant for statistical inference. This is the case of data related to different studies for which it is reasonable to assume exchangeability within and not between studies. For example, when comparing the survival time of two groups of patients displaying the same pathology and treated with two different therapies, the data can be thought of as labeled by a binary covariate identifying the specific treatment a patient is subject to. In this case, it is not realistic to assume that any two observations labeled by distinct covariate values, i.e. originating from different groups, are "homogeneous". At the same time, one might be interested in clustering patients subject to a therapy and borrow information from the outcomes on patients in the other group. Similar issues arise in more general regression problems where data are labeled by a collection of covariates taking values in a general space, and exchangeability can be assumed only for observations associated to the same covariate realization.
There clearly is a wide range of applied problems which motivate the great interest that has recently emerged for the definition of nonparametric priors accommodating for forms of dependence more general than exchangeability and able to capture the "heterogeneity" featured by the data. Here we shall focus on a proposal that is suited for handling inferential problems involving partially exchangeable data. In this setting, if z denotes a set of covariates taking values in Z, the observations are elements from a collection of sequences {(X i (z)) i≥1 : z ∈ Z} such that for any n ≥ 1, positive integers q 1 , . . . , q n P[(X q1 (z 1 ), . . . , X qn (z n )) ∈ A] = P |q| X (p q1 z1 × · · · × p qn zn )(A) Q Z (dp z1 , . . . , dp zn ) for any A ∈ X |q| , where X qj (z j ) = (X 1 (z j ), . . . , X qj (z j )), q = (q 1 , . . . , q n ) and |q| = q 1 + · · · + q n . Here Q Z is the probability distribution of a collection of possibly dependent random probability measuresp := {p z : z ∈ Z} and plays the role of prior distribution in a similar fashion as Q does in the exchangeable case (1). Indeed, note that the mixture representation in (2) is due to de Finetti in [5] and it amounts to assuming exchangeability for the sequence of observations (X i (z)) i≥1 corresponding to the same covariate value z in Z and conditional independence among observations corresponding to different covariate values. Hence, data corresponding to different covariate realizations are not conditionally identically distributed. The model can be equivalently summarized in a hierarchical form as follows When Q Z is degenerate on P X so thatp z =p z ′ (almost surely) for any z = z ′ , there is no heterogeneity among the data and exchangeability is recovered. To our knowledge, the first proposal of a nonparametric prior in this framework has been provided in [2], where Z is a finite set and Q Z is the distribution of a mixture of products of Dirichlet processes. However, the most recent literature has been spurred by S.N. MacEachern's seminal papers [22,23] where a dependent Dirichlet process is defined by means of a series representation. The main idea is to rely on the stick-breaking construction of the Dirichlet process and define, for each z in Z,p where are independent and identically distributed (i.i.d.) from some probability distribution P 0,z . For any z,p z is a Dirichlet process with base measure α z = θ z P 0,z and dependence between any two random probabilitiesp z and p z ′ , for z = z ′ , is induced by possible dependence between V i,z and V i,z ′ and/or between Y i,z and Y i,z ′ . Later developments that build on MacEachern's idea can be found, e.g., in [6,12,28,7,8]. A more complete and stimulating picture of the recent state of the art in the field, including a variety of interesting applications, is provided in [13]. The practical use of these models has been eased by the developments of suitable MCMC sampling algorithms, that make use of the stick-breaking representation ofp z in (3) and lead to an approximate evaluation of Bayesian inferences. As a matter of fact, the availability of an R package, named DPpackage, that allows for an automatic implementation of MCMC samplers with dependent processes, has made these models accessible to interested practitioners as well. See [15] for details.
The present paper relies on a different approach and defines each dependent random probability measurep z by normalizing a completely random measure (CRM) defined byμ where (J i,z ) i≥1 is a sequence of non-negative independent random variables such In view of this definition, dependence betweenp z andp z ′ will be induced by the dependence betweenμ z andμ z ′ and this is summarized through a multivariate Lévy intensity. In our proposal (Y i,z ) i≥1 = (Y i ) i≥1 and more importantlyμ z is, for any z ∈ Z, a gamma CRM which yields, through normalization, a Dirichlet process. Hence, the model discussed in this paper can be seen as a dependent Dirichlet process as in (3), with Z being a finite set. However, we are not able to establish which form of dependence among the stick-breaking weights V i,z leads to the dependent Dirichlet process prior that will be discussed in the next sections.
The use of dependent CRMs in the construction of dependent random probability measures has been considered in other work, as well. For example, in [9] and [17] dependence is induced through Lévy copulas that allow to define multivariate Lévy intensities with fixed marginals, thus operating in a similar fashion as traditional copulas do for probability distributions. More specifically, in [9] a class of dependent neutral to the right processes for the analysis of partially exchangeable survival data is constructed, whereas in [17] a bivariate vector of two-parameter Poisson-Dirichlet processes is introduced. Note that the latter is closer in spirit to the present work, the main distinctive feature being that it relies on the normalization of a random measure which is not completely random. As to our knowledge, before these two contributions the use of Lévy copulas has been mostly confined to applications in Finance as well detailed, e.g., in [3]. Furthermore, a construction that does not rely on Lévy copulas can be found in [19]: here the authors define dependent CRMs by means of the superposition of independent completely random measures some of which are shared. See also [26] for a similar approach aimed at applications to spatial statistics. An allied contribution that, similarly to ours, applies to the case where Z is a finite set is in [24] where the authors model eachp z as a convex linear combination of Dirichlet processes one of which is shared through different z's thus inducing dependence.

Main results and outline of the paper
As already anticipated, the present paper introduces a new vector of gamma CRMsμ = (μ 1 , . . . ,μ n ), which is characterized in terms of a Lévy intensity ν on (R + ) n × X. Eachμ i has the same marginal distribution and is such that for probability measure P 0 on X, for any set A in X such that P 0 (A) > 0 and for some constant c > 0. This construction has two merits. On one side it ensures mutual independence of vectors (μ 1 (A i ), . . . ,μ n (A i )), as i = 1, . . . , k, for any collection of pairwise disjoint sets A 1 , . . . , A k in X and for any k ≥ 1.
On the other, it allows one to evaluate exactly the Laplace transform of the vector (μ 1 , . . . ,μ n ) through where λ = (λ 1 , . . . , λ n ) ∈ (0, ∞) n , A is any set in X ,μ(A) = (μ 1 (A), . . . ,μ n (A)) and s, t = n i=1 s i t i for any s and t in R n . Moreover, ν is such that with dy −i denoting integration with respect to all y coordinates but the i-th.
In the special case where λ is such that λ 1 = · · · = λ n we show that A more elaborated expression, still available in closed form, is recovered if any two of the λ i 's coincide. This will be detailed in one of the next sections. In particular, if λ j = 0 for any j = i and λ i > 0, one obtains (5) so thatμ i is, indeed, a gamma completely random measure whose intensity is displayed on the right hand side of (7). Notice that the expression (8) is invariant with respect to permutations of the λ i 's, a property which remains valid even for the general case allowing for ties in (λ 1 , . . . , λ n ) as it can be noted from (25). Therefore the vectorμ is exchangeable. This is per se an interesting distinctive feature since the model still preserves analytical tractability despite a non-Markovian form of dependence. Both the availability of the Laplace transform and the independence of the increments ofμ are key tools for determining, e.g., the mixed moments ofμ. If we confine here to the simple case where n = 2, we show that for any set A in X where ( * ) denotes the set where the vectors λ = (λ 1 , . . . , λ j ) and s i = (s 1,i , s 2,i ), for i = 1, . . . , j, vary in the sum. This result, that can be extended to the case of gamma CRM vectors of dimension n > 2, is relevant for analyzing some distributional properties useful for Bayesian nonparametric inference. Indeed, if one is willing to analyze data for which partial exchangeability is a suitable form of dependence, one can set Q Z as the probability distribution of (p 1 , . . . ,p n ) This induces a vector of dependent Dirichlet processes, i.e. a vector such that eachp i is a Dirichlet process with base measure cP 0 . Moreover, Z = {1, . . . , n}. Making use of the same technique that leads to proving (9) one can evaluate the correlation coefficient betweeñ p i (A) andp j (B), for any i = j and A, B ∈ X . In particular, it is shown that for any A in X for some constant K c that depends on the total mass c > 0, for some function ξ c that can be represented in terms of hypergeometric functions 1 F 1 and 2 F 1 . The integral in (10) can be evaluated numerically and, more importantly, the right-hand side is independent from set A: this fact typically motivates the use of (10) as a measure of overall dependence betweenp 1 andp 2 . The illustration of the properties of (p 1 , . . . ,p n ) is completed by the determination of the associated partition probability function. Indeed, notice that eachp i is discrete with probability 1 so that ties may occur among the data: this induces a partition with any two observations belonging to the same cluster if and only if they coincide. The analysis of the random partition associated to a random discrete distribution is important for computational purposes. For example, when the observations are exchangeable as in (1), the so-called exchangeable partition probability function (EPPF) is a key tool for devising a MCMC sampler for density estimation in mixture models. See, e.g., [20]. This also motivates a similar investigation in the partially exchangeable case as emphasized in [19]. We shall be able to determine a closed form expression of the partially exchangeable partition probability function (pEPPF) associated to (p 1 ,p 2 ). The extension to the case n > 2, though feasible analytically, leads to complicated expressions that are not displayed here.
The outline of the paper is as follows. In Section 2 we shall introduce some basic facts on CRMs that are necessary for an understanding of the main results we have obtained. In Section 3 some known examples of multivariate vectors of CRMs are provided and a new vector of dependent gamma CRMs is introduced. Section 4 is devoted to a detailed description of Lévy copulas and of their connection with our vector of dependent gamma CRMs. As already pointed out, attention will be focused in Section 5 on the determination of the Laplace transform of the vector since it is a key quantity for the evaluation of posterior inferences in a Bayesian nonparametric framework. Finally, a normalization procedure is adopted to yield a collection of dependent Dirichlet processes which correspond to a covariate space Z = {1, . . . , n}, for some positive integer n: we shall examine their dependence structure and state a result concerning the partially exchangeable partition probability function in Section 6.

Normalized completely random measures
Among the different possibilities emerged in the literature for defining a probability distribution Q on (P X , P X ) in (1), we shall focus on a strategy that makes use of completely random measures. For this reason we devote the present section to the introduction of some preliminary material, which also serves for clarifying the main notation that will be used henceforth.
Denote by M X that space of boundedly finite measures on (X, X ), namely each element µ in M X is a measure on (X, X ) such that µ(B) < ∞ for any bounded set B ∈ X . The space M X is endowed with the so-called weak ♯ topology (see [4] for details) and M X stands for the corresponding Borel σ-algebra. A completely random measure (CRM) is a random elementμ defined on some probability space (Ω, F , P) and taking values in (M X , M X ) such that, for any k-tuple A 1 , . . . , A k of pairwise disjoint sets in X and for any k ≥ 2, the random variablesμ(A 1 ), . . . ,μ(A k ) are mutually independent. If it is further assumed thatμ does not have random masses at fixed locations, theñ where 1 A is the indicator function of set A. Moreover, the measure ν is such that R + ×B min{s, 1} ν(ds, dx) < ∞, for any B in X , and is also referred to as the intensity, or the Lévy intensity, of the CRMμ. It is important to emphasize that it characterizesμ through its Laplace functional transform representation.
Indeed, if f : X → R is any measurable function such that |f | dμ < ∞ almost surely, then Given its relevance, in the sequel we shall refer to as the Laplace exponent ofμ. A popular example is the so-called gamma CRM with base measure cP 0 , where c > 0 and P 0 is a probability measure on (X, X ). This is characterized by the Lévy intensity ν(ds, dx) = s −1 e −s ds c P 0 (dx) which leads to the following form of the Laplace functional transform for any measurable function f : X → R such that log(1 + |f |) dP 0 < ∞. If one sets f = λ1 A for any set A in X and λ > 0, from (13) one deduces thatμ(A) has Laplace transform as in (5) and is, thus, gamma distributed with shape and scale parameters equal to cP 0 (A) and 1, respectively. Another noteworthy example corresponds to ν(ds, dx) = σs −1−σ dσ P 0 (dx)/Γ(1 − σ), where σ ∈ (0, 1), and identifies the so-called σ-stable CRM. In this case, for any measurable function A choice of f = λ1 A shows thatμ(A) has a positive stable distribution with parameter σ ∈ (0, 1). As shown in [20], CRMs are the basic building blocks for the definition of nonparametric priors in an exchangeable framework. Indeed, suitable transformations ofμ lead to define random probability measures whose distribution can be used as the mixing Q in (1). The transformation of interest for the present paper is the normalization. Indeed, if the intensity ν ofμ is such that ν(R + × X) = ∞, it is possible to set Q as the probability distribution of which defines a normalized CRM or a normalized random measure with independent increments. See [27] and [14]. For example, the Dirichlet process with base measure cP 0 can be obtained through (15) ifμ is a gamma CRM whose Laplace functional transform is as in (13). Analogously, a shown in [16] the normalized σ-stable CRM with base measure P 0 coincides with the normalization of a CRM characterized through (14). The construction displayed in (15) is also relevant for our proposal. Indeed, we shall consider a vector (μ 1 , . . . ,μ n ) of dependent CRMs such that marginally eachμ i is a gamma CRM. A vector of dependent random probabilities will, then, be obtained by normalizingμ i , for i = 1, . . . , n. The resulting dependent Dirichlet processes will represent a candidate for modelling partially exchangeable data as in (2) when the cardinality of Z is n.

Gamma CRM and Dirichlet process vectors
The definition outlined in Section 2 can be extended to the case of CRM vectors. These are random elementsμ = (μ 1 , . . . ,μ n ) defined on (Ω, F , P) and taking values in (M n X , M n X ) such that, for any k ≥ 2 and for any choice of pairwise disjoint sets A 1 , . . . , A k in X , the vectors (μ 1 (A i ), . . . ,μ n (A i )) are, for i = 1, . . . , k, mutually independent. In a similar fashion as for the one-dimensional case, a Poisson process type representation holds true and (μ 1 , . . . ,μ n ) = r (J r,1 , . . . , J r,n ) δ Xr .
In the above representation, {(J r,1 , . . . , J r,n , X r ) : r = 1, 2, . . .} are points from a Poisson process on ((R + ) n , X) with intensity measure ν such that n−i and x is the usual notation for the Euclidean norm of vector x ∈ R n . Similarly to the one-dimensional case, ν characterizesμ through its Laplace functional transform representation. Indeed, for any collection of measurable functions f i : X → R such that |f | dμ i < ∞, almost surely, i = 1, . . . , n, one has where f = (f 1 , . . . , f n ), y = (y 1 , . . . , y n ) and y, f (x) = n i=1 y i f i (x). In order to simplify the notation, we shall henceforth identify the Laplace exponent ofμ as ψ * ν,n (f ) = Example 1. Instances of multivariate CRMs with a view to applications in Bayesian nonparametric inference can be found, e.g., in [9] and in [17]. For example, a vector of CRMs (μ 1 ,μ 2 ) with σ-stable margins having Laplace transform (14) has Lévy intensity of the form for any θ > 0. The corresponding Laplace exponent coincides with For further details, see [9] where this vector of dependent CRMs has been used to define neutral to the right priors for pairs of survival functions with two-sample data. An obvious simplification occurs when θ = 1/σ. In this case, indeed, one has This simplifed structure has allowed the definition of a bivariate version of the two-parameter Poisson-Dirichlet process in [17].
In contrast to this example, we now wish to introduce a CRM vectorμ whose components marginally are gamma CRMs. This is accomplished by introducing the following multivariate Lévy intensity on (R + ) n × X with |y| = n =1 y i . It is easy to check that for any A ∈ B(R + ), which implies thatμ 1 , . . . ,μ n marginally are identically distributed gamma CRMs. It is worth noting that (18) corresponds to the superposition of n vector of CRMs with the intensity of the ith summand (µ * i,1 , . . . , µ * i,n ) being Moreover, one has ν * 1 ((R + ) n × B) = ∞, whereas ν * i ((R + ) n × B) < ∞ for any i = 2, . . . , n and for any B ∈ X . Hence, for any i ≥ 2 the vector (µ * i,n , . . . , µ * n,n ) has a finite number of jumps and acts as a multivariate Poisson compound process. This structure is reminiscent of a completely random measure whose normalization has been shown in [29] to be dense in the class of homogeneous normalized completely random measures.
The components of the gamma CRM vector will be normalized, thus yielding a random probability vector (p 1 , . . . ,p n ) = μ 1 µ 1 (X) , . . . ,μ ñ µ n (X) that we shall refer to as n-variate Dirichlet process. Indeed, since eachμ i is a gamma process with base measure cP 0 , thenp i is a Dirichlet process with base measure cP 0 , for any i = 1, . . . , n. Before discussing the uses of (19) for drawing inferences with partially exchangeable observations, we shall linger on some important distributional properties of the un-normalized gamma CRM vectorμ.

Connection with Lévy copulas
A popular tool for the definition of multivariate distributions on R d , with fixed margins, is represented by copulas. See [25] for an extensive account on the topic. A similar approach has been recently introduced in an infinite-dimensional setting for the definition of multidimensional CRM vectors with fixed marginal CRMs. Lévy copulas have been originally employed for applications in mathematical finance, with the goal of describing the dynamics of portfolios including multiple dependent assets whose evolution in time is modeled through a Lévy process. See [3]. Dependence among the components of a Lévy process vector is, then, induced by means of Lévy copulas which act at the level of Lévy intensities. Since increasing Lévy processes are special case of CRMs, one can easily extend the use of Lévy copulas for defining CRM vectors, as illustrated in [9] and [17]. In view of the definition provided in the previous section, one might wonder whether it is possible to identify the Lévy copula that induces (18) starting from gamma univariate margins. Note that, according to Theorem 5.4 in [3] such a copula is unique. The interest in such a result is motivated by the fact that a possible representation of dependence through a Lévy copula allows one to simulate approximate realizations of the trajectories of the CRM vector. Indeed, this can be achieved by relying on an extension of the well-known Ferguson and Klass algorithm. For details, see Algorithm 6.5 in [3]. Besides this, it can provide further insight into the dependence structure featured by the CRM vector we have introduced. Let us focus on the two-dimensional case, i.e. n = 2, and suppose each marginal CRM is homogeneous with intensity that can be represented as follows Note that the our proposal clearly fits this setting. A Lévy copula is, then, a function C : [0, ∞] 2 → [0, ∞] such that (i) C(y 1 , 0) = C(0, y 2 ) = 0 for any positive y 1 and y 2 , (ii) C has uniform margins, i.e. C(y 1 , ∞) = y 1 and C(∞, y 2 ) = y 2 , (iii) for all y 1 < z 1 and y 2 < z 2 , C(y 1 , y 2 )+C(z 1 , z 2 )−C(y 1 , z 2 )−C(y 2 , z 1 ) ≥ 0.
For example the independence case, i.e. A×B ρ(s 1 , s 2 ) ds 1 ds 2 = A ρ(s 1 ) ds 1 + B ρ 2 (s 2 ) ds 2 for any A and B in B(R + ), corresponds to the Lévy copula On the other hand, the case of completely dependent CRMs corresponds to C (y 1 , y 2 ) = min{y 1 , y 2 } which yields a vector (μ 1 ,μ 2 ) such that for any x and y in X eitherμ i ({x}) < µ i ({y}) orμ i ({x}) >μ i ({y}), for i = 1, 2, almost surely. Intermediate cases, between these two extremes, can be detected, for example, by relying on the Lévy-Clayton copula defined by with the parameter θ regulating the degree of dependence. It can be seen that lim θ→0 C θ = C ⊥ and lim θ→∞ C θ = C .
We shall now focus on the determination of the Lévy copula that identifies the specific gamma CRM vector characterized by the Lévy measure (18). We shall use the notation Γ(a, x) = ∞ x s a−1 e −s ds for the incomplete gamma function, whereas Γ −1 (a, x) is the inverse function of x → Γ(a, x), for any a ∈ R. Proposition 1. The measure ν defined in (18) with n = 2 can be recovered by applying the Lévy Copula to a pair of gamma CRMs.
Proof. The tail integral of each marginal gamma CRM is

F. Leisen et al.
On the other hand, the tail integral associated to (18) with n = 2 is After the change of variable s = y 1 + y 2 and t = y 1 we obtain From Theorem 5.3 in [3], the copula C for this process is characterized by which in this case reduces to Setting y i = Γ(0, x i ), for i = 1, 2, completes the proof.

The Laplace exponent
As already mentioned, the Laplace exponent (12) of a CRMμ is an important tool for its possible uses in Bayesian nonparametric inference as pointed out, e.g., in [20]. For example, it may help to determine, via differentiation, a formula for the so-called exchangeable partition probability function (EPPF) corresponding to the distribution of a sample of exchangeable observations (X n ) n≥1 as in (1) with Q being the probability distribution of a normalized CRMp :=μ/μ(X). The EPPF is also a key quantity needed for the posterior calculus ofp, given the data. Similarly, the availability of an exact expression of ψ * ν is decisive for evaluating posterior inferences, either exact or approximated, with exchangeable survival data when X = R + andμ is both taken to define a neutral to the right prior for the survival function as and as the mixing measure in a prior representation for the hazard rate function throughS for some kernel function k( · ; · ) on R + × X. All these arguments are still relevant when working in a multivariate framework, where a multidimensional analogue of ψ * ν can be similarly applied to a Bayesian nonparametric model with partially exchangeable observations: this motivates the focus we are reserving in this Section to the determination of the Laplace exponent induced by (18).
Before getting started, it is worth noting that, since (μ 1 , . . . ,μ n ) has independent increments, its distribution is characterized by a choice of f 1 , . . . , f n in (16) such that f i = λ i 1 A for any set A in X , λ i ∈ R + and i = 1, . . . , n. In this case ψ * ν,n (f ) = H(A) ψ ν,n (λ) where λ = (λ 1 , . . . , λ n ) and Our goal is to prove the following Proposition 2. Let λ ∈ (R + ) n be such that it consists of l ≤ n distinct values denoted asλ 1 , . . . ,λ l with respective multiplicities (n 1 , . . . , n l ). Then where To prove Proposition 2, we first show in detail how to deal with the case n = 2 and, then, deduce the result for an arbitrary n by induction. A recursive formula provided in Lemma 1 will be the engine for the induction. For general n a delicate aspect we will encounter is related to deriving a form for ψ ν,n (λ) when not all the coordinates in λ are distinct. For n > 2 for the sake of clarity it will be convenient to deal separately with a derivation when there are no ties in λ (Proposition 4) and then eventually to extend the proof to the case of general λ ∈ R n . For the case n = 2, the joint Laplace functional transform is given as follows.
Proposition 3. Let ν be the Lévy intensity introduced in (18) with n = 2. The corresponding Laplace exponent has the following form: F. Leisen et al. Proof. Suppose λ 1 = λ 2 . Correspondingly one has ψ ν,2 (λ 1 , λ 2 ) = I 1 (λ 1 , λ 2 ) + I 2 (λ 1 , λ 2 ) where The change of variable y 1 + y 2 = w and y 1 /(y 1 + y 2 ) = z leads to and, similarly and combining these two expression one obtains ψ ν . Proceeding in a similar fashion, and with some useful simplifications, one also obtains ψ ν,2 (λ 1 , λ 1 ) when The statement of Proposition 2 points out that one needs to take into account possible ties in the vector λ when determining an expression, in closed form, of ψ ν,n . Hence, when dealing with the case n > 2 we shall first assume that λ has no ties and, then, move on to the case where any two λ i and λ j , with i = j, may coincide.
We are now in a position to state and prove the following representation of the Laplace exponent of our multivariate gamma CRM vector, when there are no ties in its argument λ. where φ n is as in (26).
Proof. Suppose (30) holds true for n and we shall show that this implies the validity of (30) for n + 1. By virtue of Proposition 3 the proof is thus completed by induction. Since (30) holds true for n, for any λ ∈ (R + ) n+1 one has If these two expressions are plugged in the recursive relation (27) one has .
After some algebra, one shows that ψ ν,n+1 satisfies (30) and the proof is completed.
To achieve a complete proof of Proposition 2, it only remains to extend the technique used for the proof of Proposition 4, by taking this time into account possible ties in λ ∈ (R + ) n .
Proof of Proposition 2. If B j = {i : λ i =λ j }, for any j = 1, . . . , l, and |y| j = i∈Bj y i for any y ∈ (R + ) n , one has, similarly to (28), This implies that Leisen et al. where the last equality follows from (29). Now note that and from this deduce which, from (28) and by virtue of the definition of the function φ l in (30), completes the proof of (25).

Investigating the dependence structure
The determination of the Laplace exponent discussed in the previous section is preliminary to an investigation of the dependence structure among the random probabilities of the vectorp = (p 1 , . . . ,p n ) of dependent Dirichlet processes defined by (19). Givenp is an infinite-dimensional object, some hints on the dependence between any pairp i andp j , for i = j, are conveyed by the correlation coefficient betweenp i (A) andp j (A), for any A ∈ X . For this reason, we shall first provide an expression for such a coefficient and, then, rely on the same technique for determining the partially exchangeable partition probability function of the sample X q1 (z 1 ), . . . , X qn (z n ) characterized through (2). Indeed, almost sure discretenessp i implies that P[X i (z κ ) = X j (z ℓ )] > 0 for any i, j, κ and ℓ.
In other terms, ties may appear both within each sample and between different samples X (qκ) (z κ ) and X (q ℓ ) (z ℓ ). Hence, the q 1 + · · · + q n data consist of k distinct values forming clusters of sizes N 1 , . . . , N k Moreover, N j = n i=1 q j,i ≥ 1 with q j,i denoting the number of observations from X qi (z i ) coinciding with the j-th distinct value in the sample. In the following, we confine our treatment to the case n = 2.

Mixed moments and correlations
We rely on an approach used in [17] for defining a bivariate two-parameter Poisson-Dirichlet process: this arises as the normalization of a random measure that does not necessarily satisfy the property of independence when evaluated on pairwise disjoint sets as a CRM does. Accordingly, a key quantity is going to be where the notation corresponds to the one set forth in (20) and in (21): this is legitimate since the case we are examining is homogeneous. See (18). Note that one can also write g ρ (q 1 , q 2 ; s, t) = I 1 (q 1 , q 2 ; s, t) + I 2 (q 1 , q 2 ; s, t) where y q1 1 y q2 2 e −sy1−ty2 e −y1−y2 y 1 + y 2 dy 1 dy 2 when q 1 + q 2 ≥ 1. Moreover, g ρ (0, 0; s, t) ≡ 1. a simple change of variable into polar coordinates yields In a similar fashion one determines I 2 (q 1 , q 2 ; s, t) thus yielding The availability of the g ρ function allows us to determine an expression of the mixed moments of the un-normalized vector (μ 1 (A),μ 2 (A)), with A ∈ X . This is a preliminary step towards the determination of the mixed moments of the normalized vector (p 1 (A),p 2 (A)). In the sequel, for any two vectors x = (x 1 , . . . , x d ) and y = (y 1 , . . . , y d ) in N d 0 , one writes x ≺ y if either |x| < |y| or |x| = |y| and x 1 < y 1 or if |x| = |y| with x i = y i for i = 1, . . . , j and x j+1 < y j+1 for some j in {1, . . . , d}.
The procedure described in the previous proof can be used to determine the correlation coefficient betweenp 1 (A) andp 2 (B), for any A and B in X . In particular, when A = B, it will be seen that such a correlation does not depend on the specific setp 1 andp 2 are evaluated at: this finding is typically used to motivate the use of corr(p 1 (A),p 2 (A)) as a measure of the dependence betweeñ p 1 andp 2 . The main difference with respect to the previous proof is the use of a gamma integral representation of the normalizing total random massesμ 1 (X) andμ 2 (X). In order to make notation simpler, we set where 1 F 1 is the confluent hypergeometric function and 2 F 1 is the Gauss hypergeometric function.
Proposition 6. Let A and B be any two sets in X and suppose that (p 1 ,p 2 ) is a Dirichlet vector defined as in (19), with n = 2. Then In order to exploit the independence of the increments of (μ 1 ,μ 2 ) on can consider the (measurable) partition of X generated by {A, B}. Hence, one can rewrite the right-hand-side of (35) as follows × E e −sμ1(X * )−tμ2(X * ) ds dt where X * = X \ (A∆B) and ∆ stands for the symmetric difference between sets. Each factor in the integrand can be easily evaluated. As for the first one, for example, one has