betaFeedback

WorksMathematics

Bayesian nonparametric mixtures of asymmetric Archimedean copulas

Abstract

Archimedean copulas are exchangeable: a single generator parameter controls every pairwise association, and the copula is invariant under permutations of its arguments. Even when mixtures of Archimedean copulas are formulated, the exchangeability feature is inherited. Non exchangeable copulas can be defined through the Khoudraji transformation of an Archimedean copula. This relies in replacing the arguments by a power transformation and adding a factor based on the independence copula. The construction produces pairwise Kendall tau coefficients that depend on the power parameters and recovers the original Archimedean copula as a particular case. We propose a Bayesian nonparametric mixture model by mixing over the copula generator parameter and the parameters that characterize the asymmetry. We use the Poisson–Dirichlet process as a mixing distribution, which groups observations by both strength and direction of dependence. We derive the copula density, the pairwise association measures, the tail coefficients, and the full posterior conditional distributions to perform a Bayesian analysis. We carry out simulation studies and real data analyses. The results show that the Khoudraji mixture improves greatly on, or agrees with, the exchangeable fit.

Introduction

Let X=(X1,…,Xd)\mathbf{X} = (X_1, \ldots, X_d) be a continuous dd-dimensional random vector with joint cumulative distribution function (henceforth, CDF) FF and marginal CDFs FjF_j. By Sklar [20] there is a unique copula C:[0,1]d→[0,1]C : [0, 1]^d \to[0, 1] such that F(x1,…,xd)=C(F1(x1),…,Fd(xd))F(x_1, \ldots, x_d) = C(F_1(x_1), \ldots, F_d(x_d)). Dependence measures such as Kendall’s tau are functionals of CC alone [14].

A leading parametric class is the Archimedean family

C(u1,…,ud)=ψ−1{ψ(u1)+⋯+ψ(ud)},(1)C(u_1,\ldots,u_d)=\psi^{-1}\{\psi(u_1)+\cdots+\psi(u_d)\}, \tag*{(1)}

generated by a continuous, decreasing, and convex function ψ:[0,1]→[0,∞]\psi:[0,1]\to[0,\infty] with ψ(1)=0\psi(1)=0 and ψ(0)=∞\psi(0)=\infty ([14, 12]). The copula CψC_\psi is exchangeable in the sense that Cψ(uσ(1),…,uσ(d))=Cψ(u1,…,ud)C_\psi(u_{\sigma(1)},\ldots,u_{\sigma(d)})=C_\psi(u_1,\ldots,u_d) for every permutation σ\sigma, and every pairwise margin is the same bivariate Archimedean copula. Consequently, every pair (Xj,Xk)(X_j,X_k) shares the same Kendall tau τθ\tau_\theta, a single number determined by the generator parameter θ\theta.

Pan et al. [16] mixed Archimedean copulas with a Poisson–Dirichlet process ([18]) on θ\theta. The Bayesian nonparametric (BNP) mixture is more flexible than a single Archimedean copula because it can be multimodal and it clusters the sample in subgroups with different association parameters. However, every component remains exchangeable, so the model cannot distinguish τ12\tau_{12} from τ13\tau_{13}. This work is motivated by the need to extend Pan et al. [16] to include non-exchangeable copula structures within a BNP mixture approach. Importantly, the new construction generalizes and contains the exchangeable case.

A standard device for breaking exchangeability, due to Khoudraji [9] and studied by Genest et al. [7], Liebscher [11] and Durante and Salvadori [5], is to combine an Archimedean copula with the independence copula after coordinate-wise power transforms. The resulting family is no longer symmetric in its arguments, pairwise Kendall τ\tau’s may differ, and one coordinate can be closer to independence than another, while the original Archimedean copula and the independence copula remain nested.

This paper extends the Bayesian nonparametric mixture of symmetric Archimedean copulas ([16]) to the assymmetric Khoudraji case. This involves placing a Poisson–Dirichlet process prior on the pair (θ,α)(\theta,\boldsymbol{\alpha}), where θ\theta is the generator parameter and α=(α1,…,αd)∈[0,1]d\boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_d)\in[0,1]^d is the Khoudraji asymmetry vector.

The rest of the paper is organised as follows. Section 2 reviews the transform, derives the copula density, and records association and tail formulae. Section 3 defines the Bayesian nonparametric mixture and the posterior sampler. Section 4 describes numerical experiments that parallel those of Pan et al. [16], including refits of their occupancy and red-wine analyses. Section 5 concludes.

Khoudraji-transformed Archimedean copulas

The construction

Let CθC_\theta be a dd-dimensional Archimedean copula with generator ψθ\psi_\theta and association parameter θ∈Θ\theta\in\Theta, and let Π(u)=∏j=1duj\Pi(\mathbf{u})=\prod_{j=1}^{d}u_j denote the independence copula. For α=(α1,…,αd)∈[0,1]d\boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_d)\in[0,1]^d, the Khoudraji transform of CθC_\theta is

Cθ,α(u)=Cθ(u1α1,…,udαd)Π(u11−α1,…,ud1−αd)=ψθ−1{ψθ(u1α1)+⋯+ψθ(udαd)}∏j=1duj1−αj.(2)\begin{aligned} C_{\theta,\boldsymbol{\alpha}}(\mathbf{u})=C_\theta(u_1^{\alpha_1},\ldots,u_d^{\alpha_d})\Pi(u_1^{1-\alpha_1},\ldots,u_d^{1-\alpha_d}) \\ &=\psi_\theta^{-1}\{\psi_\theta(u_1^{\alpha_1})+\cdots+\psi_\theta(u_d^{\alpha_d})\}\prod_{j=1}^{d}u_j^{1-\alpha_j}. \tag*{(2)} \end{aligned}

The product of two copulas evaluated in coordinate powers is itself a copula for every α∈[0,1]d\boldsymbol{\alpha} \in[0,1]^d ([9, 11]). We denote by cθ,αc_{\theta,\boldsymbol{\alpha}} the density of Cθ,αC_{\theta,\boldsymbol{\alpha}} when it exists.

Proposition 1 (Special exchangeable cases) Let Cθ,αC_{\theta,\boldsymbol{\alpha}} be given by (2).

(a) If α=1\boldsymbol{\alpha} = 1, then Cθ,α=CθC_{\theta,\boldsymbol{\alpha}} = C_\theta.

(b) If α=0\boldsymbol{\alpha} = 0, then Cθ,α=ΠC_{\theta,\boldsymbol{\alpha}} = \Pi.

(c) If αj=0\alpha_j = 0 for a single coordinate jj, then UjU_j is independent of (Uk)k≠j(U_k)_{k\ne j} under Cθ,αC_{\theta,\boldsymbol{\alpha}}.

(d) The bivariate margin of (Uj,Uk)(U_j,U_k) is the bivariate Khoudraji transform of the bivariate Archimedean copula generated by ψθ\psi_\theta, with asymmetry parameters (αj,αk)(\alpha_j,\alpha_k). In particular Cθ,α(u,v)≠Cθ,α(v,u)C_{\theta,\boldsymbol{\alpha}}(u,v) \ne C_{\theta,\boldsymbol{\alpha}}(v,u) whenever αj≠αk\alpha_j \ne\alpha_k, so the copula is not exchangeable.

Proof. (a) and (b) are immediate from Cθ(u)Π(1)=Cθ(u)C_\theta(\mathbf{u})\Pi(\mathbf{1}) = C_\theta(\mathbf{u}) and Cθ(1)Π(u)=Π(u)C_\theta(\mathbf{1})\Pi(\mathbf{u}) = \Pi(\mathbf{u}). (c) Substituting uju_j and uk=1u_k = 1 for k≠jk \ne j gives Cθ,α(1,…,uj,…,1)=ujC_{\theta,\boldsymbol{\alpha}}(1,\ldots,u_j,\ldots,1) = u_j, and the joint CDF factors are given by uju_j times a copula on the remaining coordinates because Cθ(…,uj,…)=Cθ(…,1,…)C_\theta(\ldots,u_j,\ldots) = C_\theta(\ldots,1,\ldots) does not depend on uju_j except through explicit power uj1−αj=uju_j^{1-\alpha_j} = u_j. (d) The (j,k)(j,k) margin is

Cθ,α(u,v,1,…,1)=Cθ(uαj,vαk,1,…,1)u1−αjv1−αk.C_{\theta,\boldsymbol{\alpha}}(u,v,1,\ldots,1) = C_\theta(u^{\alpha_j},v^{\alpha_k},1,\ldots,1)u^{1-\alpha_j}v^{1-\alpha_k}.

Archimedean copulas satisfy Cθ(u,v,1,…,1)=ψθ−1{ψθ(u)+ψθ(v)}C_\theta(u,v,1,\ldots,1) = \psi_\theta^{-1}\{\psi_\theta(u)+\psi_\theta(v)\}, the corresponding bivariate Archimedean copula. The resulting function of (u,v)(u,v) is therefore (2) in dimension two with parameters (αj,αk)(\alpha_j,\alpha_k). It is symmetric in (u,v)(u,v) if and only if αj=αk\alpha_j = \alpha_k. ◇

Thus, the vector α\boldsymbol{\alpha} interpolates between a fully exchangeable Archimedean copula and independence, coordinate by coordinate. The copula of Pan et al. [16] is the corner α=1\boldsymbol{\alpha} = 1 of the present family.

Bivariate density

For posterior inference, we need the density of the copula. In dimension two write α,β∈[0,1]\alpha,\beta\in[0,1] for the two asymmetry parameters and Cθ(u,v)C_\theta(u,v) for the bivariate Archimedean copula, with first partials

Cθ(u)(u,v)=∂∂uCθ(u,v)=ψθ(1)(u)ψθ(1){Cθ(u,v)},Cθ(v)(u,v)=∂∂vCθ(u,v)=ψθ(1)(v)ψθ(1){Cθ(u,v)},(3)C_\theta^{(u)}(u,v) = \frac{\partial}{\partial u}C_\theta(u,v) = \frac{\psi_\theta^{(1)}(u)}{\psi_\theta^{(1)}\{C_\theta(u,v)\}}, \qquad C_\theta^{(v)}(u,v) = \frac{\partial}{\partial v}C_\theta(u,v) = \frac{\psi_\theta^{(1)}(v)}{\psi_\theta^{(1)}\{C_\theta(u,v)\}}, \tag*{(3)}

and cross second derivative

cθ(u,v)=∂2∂u∂vCθ(u,v)=−ψθ(1)(u)ψθ(1)(v)ψθ(2){Cθ(u,v)}[ψθ(1){Cθ(u,v)}]3,(4)c_\theta(u,v) = \frac{\partial^2}{\partial u\partial v}C_\theta(u,v) = -\frac{\psi_\theta^{(1)}(u)\psi_\theta^{(1)}(v)\psi_\theta^{(2)}\{C_\theta(u,v)\}}{[\psi_\theta^{(1)}\{C_\theta(u,v)\}]^3}, \tag*{(4)}

where ψθ(1)(t)\psi_\theta^{(1)}(t) and ψθ(2)(t)\psi_\theta^{(2)}(t) are the first and second derivatives of the generator ψθ(t)\psi_\theta(t) with respect to tt, respectively. Table 1 shows five generators ψθ\psi_\theta for the most common Archimedean families, together with their first two derivatives.

Copulaψθ(t)\psi_\theta(t)ψθ(1)(t)\psi_\theta^{(1)}(t)ψθ(2)(t)\psi_\theta^{(2)}(t)
AMHlog⁡{1−θ(1−t)t}\log\left\{\frac{1-\theta(1-t)}{t}\right\}θ−1t{1−θ(1−t)}\frac{\theta-1}{t\{1-\theta(1-t)\}}(1−θ)(1−θ+2θt)t{1−θ(1−t)}2\frac{(1-\theta)(1-\theta+2\theta t)}{t\{1-\theta(1-t)\}^{2}}
CLAt−θ−1t^{-\theta}-1−θt−(θ+1)-\theta t^{-(\theta+1)}θ(θ+1)t−(θ+2)\theta(\theta+1)t^{-(\theta+2)}
FRA−log⁡{exp⁡(−θt)−1exp⁡(−θ)−1}-\log\left\{\frac{\exp(-\theta t)-1}{\exp(-\theta)-1}\right\}θexp⁡(−θt)exp⁡(−θt)−1\frac{\theta\exp(-\theta t)}{\exp(-\theta t)-1}θ2exp⁡(−θt){exp⁡(−θt)−1}2\frac{\theta^2\exp(-\theta t)}{\{\exp(-\theta t)-1\}^{2}}
GUM(−log⁡t)θ(-\log t)^\theta−θt(−log⁡t)θ−1-\frac{\theta}{t}(-\log t)^{\theta-1}θt2(−log⁡t)θ−1+θ(θ−1)t2(−log⁡t)θ−2\frac{\theta}{t^2}(-\log t)^{\theta-1}+\frac{\theta(\theta-1)}{t^2}(-\log t)^{\theta-2}
JOE−log⁡{1−(1−t)θ}-\log\{1-(1-t)^\theta\}−θ(1−t)θ−11−(1−t)θ-\frac{\theta(1-t)^{\theta-1}}{1-(1-t)^\theta}θ(θ−1)(1−t)θ−2+θ(1−t)2θ−2{1−(1−t)θ}2\frac{\theta(\theta-1)(1-t)^{\theta-2}+\theta(1-t)^{2\theta-2}}{\{1-(1-t)^\theta\}^{2}}

Table 1. Common Archimedean families used as the inner copula CθC_\theta in (2). Generator ψθ(t)\psi_\theta(t), together with its first ψθ(1)(t)\psi_\theta^{(1)}(t) and second ψθ(2)(t)\psi_\theta^{(2)}(t) derivatives.

Proposition 2 (Bivariate Khoudraji density). The copula (2) in dimension two admits the density

cθ,α,β(u,v)=αβcθ(uα,vβ)+α(1−β)v−βCθ(u)(uα,vβ)+β(1−α)u−αCθ(v)(uα,vβ)+(1−α)(1−β)u−αv−βCθ(uα,vβ),(5)c_{\theta,\alpha,\beta}(u,v)=\alpha\beta c_\theta(u^\alpha,v^\beta)+\alpha(1-\beta)v^{-\beta}C_\theta^{(u)}(u^\alpha,v^\beta) +\beta(1-\alpha)u^{-\alpha}C_\theta^{(v)}(u^\alpha,v^\beta)+(1-\alpha)(1-\beta)u^{-\alpha}v^{-\beta}C_\theta(u^\alpha,v^\beta), \tag*{(5)}

for (u,v)∈[0,1]2(u,v)\in[0,1]^2, provided cθc_\theta exists. If α=β=1\alpha=\beta=1 then cθ,α,β=cθc_{\theta,\alpha,\beta}=c_\theta; if α=β=0\alpha=\beta=0 then cθ,α,β=1c_{\theta,\alpha,\beta}=1.

Proof. Write C(u,v)=A(u,v)B(u)D(v)C(u,v)=A(u,v)B(u)D(v) with A=Cθ(uα,vβ)A=C_\theta(u^\alpha,v^\beta), B=u1−αB=u^{1-\alpha} and D=v1−βD=v^{1-\beta}. Differentiating the product twice and using Au=Cθ(u)(uα,vβ)αuα−1A_u=C_\theta^{(u)}(u^\alpha,v^\beta)\alpha u^{\alpha-1}, Bu=(1−α)u−αB_u=(1-\alpha)u^{-\alpha} together with the analogous formulae in vv produces four terms, which rearrange into (5). The two exchangeable cases are Proposition 1(a)–(b). ◇

The four terms in (5) correspond to the four ways of assigning each coordinate to the Archimedean factor or to the independence factor. Substituting (3) and (4) expresses the density entirely in terms of ψθ\psi_\theta and its first two derivatives, so the implementation of Pan et al. [16] extends by evaluating CθC_\theta, Cθ(u)C_\theta^{(u)}, Cθ(v)C_\theta^{(v)} and cθc_\theta at (uα,vβ)(u^\alpha,v^\beta) and forming the linear combination (5).

In dimension d>2d>2 the density exists whenever CθC_\theta has a density (McNeil and Nešlehová [12]). It is the mixed partial of (2) and can be evaluated by automatic differentiation or by the recursive expansion of Liebscher [11]; the bivariate formula is the one required for the pairwise association calculations below and for the numerical experiments of § 4.

Pairwise Kendall tau

Kendall’s tau of a pair (U,V)(U,V) is

τθ,α,β=4∫01∫01Cθ,α,β(u,v)cθ,α,β(u,v) du dv−1,(6)\tau_{\theta,\alpha,\beta}=4\int_0^1\int_0^1 C_{\theta,\alpha,\beta}(u,v)c_{\theta,\alpha,\beta}(u,v)\,\mathrm{d}u\,\mathrm{d}v-1, \tag*{(6)}

where Cθ,α,βC_{\theta,\alpha,\beta} is any bivariate margin of Proposition 1(d). Unlike the Archimedean case, τθ,α,β\tau_{\theta,\alpha,\beta} depends on the pair through (α,β)(\alpha,\beta) as well as through θ\theta.

Proposition 3 (Pairwise Kendall tau). Write τθ\tau_\theta for Kendall’s tau of CθC_\theta as in Table 1. For the Khoudraji copula (2),

(a) τθ,α,β=τθ\tau_{\theta,\alpha,\beta}=\tau_\theta if α=β=1\alpha=\beta=1;

(b) τθ,α,β=0\tau_{\theta,\alpha,\beta}=0 if α=0\alpha=0 or β=0\beta=0;

(c) if α=β∈[0,1]\alpha=\beta\in[0,1], then τθ,α,α\tau_{\theta,\alpha,\alpha} interpolates continuously between 00 (at α=0\alpha=0) and τθ\tau_\theta (at α=1\alpha=1);

(d) in general τθ,α,β\tau_{\theta,\alpha,\beta} is obtained from (6) with Cθ,α,βC_{\theta,\alpha,\beta} and cθ,α,βc_{\theta,\alpha,\beta} given by (2) and (5), and the map (α,β)↦τθ,α,β(\alpha,\beta)\mapsto\tau_{\theta,\alpha,\beta} is continuous on [0,1]2[0,1]^2 for any valid θ∈Θ\theta\in\Theta.

Proof. (a) and (b) follow from Proposition 1. For (c) and (d), Cθ,α,β(u,v)C_{\theta,\alpha,\beta}(u,v) is jointly continuous in (α,β,u,v)(\alpha,\beta,u,v) on [0,1]2×(0,1)2[0,1]^2\times(0,1)^2 and bounded, so (6) is continuous in (α,β)(\alpha,\beta) for any valid θ∈Θ\theta\in\Theta. The endpoint values are (a) and (b). ◇

There is, in general, no closed form for τθ,α,β\tau_{\theta,\alpha,\beta} except at the corners of the unit square in (α,β)(\alpha,\beta). Numerical integration of (6), or a Monte Carlo estimate from a sample of the copula, is therefore the operational formula.

The essential modelling gain is that the Khoudraji copula for d>2d>2 can now carry a vector of pairwise tau’s. For d=3d=3 and α=(α1,α2,α3)\boldsymbol{\alpha}=(\alpha_1,\alpha_2,\alpha_3), the three pairs (τθ,α1,α2,τθ,α1,α3,τθ,α2,α3)(\tau_{\theta,\alpha_1,\alpha_2},\tau_{\theta,\alpha_1,\alpha_3},\tau_{\theta,\alpha_2,\alpha_3}) are functions of the same θ\theta but of three different pairs of asymmetry parameters, and they coincide if and only if α1=α2=α3=1\alpha_1=\alpha_2=\alpha_3=1 (the Archimedean corner) or the corresponding α\alpha’s vanish.

Tail coefficients

Write λL=lim⁡u↓0C(u,u)/u\lambda_L=\lim_{u\downarrow0}C(u,u)/u and λU=lim⁡u↑1{1−2u+C(u,u)}/(1−u)\lambda_U=\lim_{u\uparrow1}\{1-2u+C(u,u)\}/(1-u) for the lower and upper tail dependence coefficients of a bivariate copula C(u,v)C(u,v). For the five Archimedean families of Table 1 one has λL=2−1/θ\lambda_L=2^{-1/\theta} and λU=0\lambda_U=0 for Clayton with θ>0\theta>0; λL=0\lambda_L=0 and λU=2−21/θ\lambda_U=2-2^{1/\theta} for Gumbel and for Joe; and λL=λU=0\lambda_L=\lambda_U=0 for Frank and AMH.

Proposition 4 (Tail coefficients of the Khoudraji transform). Let Cθ,α,βC_{\theta,\alpha,\beta} denote (2) in dimension two with λL(θ,α,β)\lambda_L(\theta,\alpha,\beta) and λU(θ,α,β)\lambda_U(\theta,\alpha,\beta) its corresponding lower and upper tail coefficients, respectively.

(a) λL(θ,α,β)=0\lambda_L(\theta,\alpha,\beta)=0 unless α=β=1\alpha=\beta=1, in which case the Archimedean coefficient is recovered.

(b) If the inner copula CθC_\theta is Gumbel, the Khoudraji transform is an extreme-value copula with

λU(θ,α,β)=α+β−(αθ+βθ)1/θ,\lambda_U(\theta,\alpha,\beta)=\alpha+\beta-(\alpha^\theta+\beta^\theta)^{1/\theta},

which is strictly positive whenever α,β>0\alpha,\beta>0 and θ>1\theta>1, and equals $2-2^{1/\theta}atat\alpha=\beta=1.

(c) If the inner copula CθC_\theta is Clayton, Frank or AMH, then λU(θ,α,β)=0\lambda_U(\theta,\alpha,\beta)=0. If it is Joe, upper-tail dependence is inherited in diluted form for α,β>0\alpha,\beta>0.

Thus lower-tail dependence (Clayton) requires both coordinates to be fully assigned to the Archimedean factor, whereas upper-tail dependence (Gumbel, Joe) survives as soon as both asymmetry parameters are positive.

Proof Any copula satisfies C(s,t)≤min⁡(s,t)C(s,t)\leq\min(s,t), so

Cθ,α,β(u,u)u≤umax⁡(α,β)+1−α−β=u1−min⁡(α,β).\frac{C_{\theta,\alpha,\beta}(u,u)}{u}\leq u^{\max(\alpha,\beta)+1-\alpha-\beta}=u^{1-\min(\alpha,\beta)}.

The right-hand side tends to 0 unless min⁡(α,β)=1\min(\alpha,\beta)=1, i.e. unless α=β=1\alpha=\beta=1. This is (a). For (b), Gumbel is extreme-value with stable tail dependence function ℓ(x,y)=(xθ+yθ)1/θ\ell(x,y)=(x^\theta+y^\theta)^{1/\theta}. The Khoudraji transform of an extreme-value copula is extreme-value (Genest et al. [7]; Durante and Salvadori [5]), with Pickands function

Aα,β(w)=[(αw)θ+{β(1−w)}θ]1/θ+(1−α)w+(1−β)(1−w).A_{\alpha,\beta}(w)=\left[(\alpha w)^\theta+\{\beta(1-w)\}^\theta\right]^{1/\theta}+(1-\alpha)w+(1-\beta)(1-w).

The identity λU=2{1−A(1/2)}\lambda_U=2\{1-A(1/2)\} yields the displayed formula. For θ>1\theta>1 the map x↦xθx\mapsto x^\theta is strictly convex, (αθ+βθ)1/θ<α+β(\alpha^\theta+\beta^\theta)^{1/\theta}<\alpha+\beta whenever α,β>0\alpha,\beta>0, and therefore λU>0\lambda_U>0. (c) follows from the corresponding Archimedean coefficients and Proposition 1(a). ◇

Asymmetry of the density survives for α≠β\alpha\ne\beta even when both tail coefficients vanish: cθ,α,β(u,v)≠cθ,α,β(v,u)c_{\theta,\alpha,\beta}(u,v)\ne c_{\theta,\alpha,\beta}(v,u), so probability mass concentrates towards one axis. In the following § 3 we propose a mixture that can describe populations in which some clusters are exchangeable and heavy-tailed (Archimedean, α=1\alpha=1) while others are asymmetric, with tail behaviour governed by Proposition 4.

BNP mixtures

Model

Let ξ=(θ,α)∈Ξ:=Θ×[0,1]d\xi=(\theta,\boldsymbol{\alpha})\in\Xi:=\Theta\times[0,1]^d collect the Archimedean parameter and the Khoudraji vector. We mix the family {Cθ,α(⋅):ξ∈Ξ}\{C_{\theta,\boldsymbol{\alpha}}(\cdot):\xi\in\Xi\} with a Poisson–Dirichlet process on Ξ\Xi, exactly as Pan et al. [16] mixed {Cθ(⋅):θ∈Θ}\{C_\theta(\cdot):\theta\in\Theta\} on Θ\Theta. A probability measure GG has a Poisson–Dirichlet prior G∼PD⁡(a,b,G0)G\sim\operatorname{PD}(a,b,G_0) with a∈[0,1)a\in[0,1), b>−ab>-a and centering measure G0G_0 on Ξ\Xi when

G(⋅)=∑k=1∞ωkδξk(⋅).(7)G(\cdot)=\sum_{k=1}^{\infty}\omega_k\delta_{\xi_k}(\cdot). \tag*{(7)}

with stick-breaking weights ωk\omega_k as in Pitman and Yor [18] and atoms ξk∼iidG0\xi_k\mathrel{\overset{\mathrm{iid}}{\sim}}G_0. The mixture copula is

C(u)=∫Cξ(u) G(dξ)=∑k=1∞ωkCθk,αk(u).(8)C(\mathbf{u})=\int C_\xi(\mathbf{u})\,G(\mathrm{d}\xi)=\sum_{k=1}^{\infty}\omega_k C_{\theta_k,\boldsymbol{\alpha}_k}(\mathbf{u}). \tag*{(8)}

Hierarchically, for observations i=1,…,ni = 1, \ldots, n,

(U1i,…,Udi)∣θi,αi∼indcθi,αi(ui),(θi,αi)∣G∼iidG,G∼PD⁡(a,b,G0),(9)\begin{aligned} (U_{1i}, \ldots, U_{di}) \mid\theta_i, \boldsymbol{\alpha}_i &\mathrel{\mathop{\sim}^{\mathrm{ind}}} c_{\theta_i,\boldsymbol{\alpha}_i}(\mathbf{u}_i),\\ (\theta_i,\boldsymbol{\alpha}_i) \mid G &\mathrel{\mathop{\sim}^{\mathrm{iid}}} G,\\ G &\sim\operatorname{PD}(a,b,G_0), \tag*{(9)} \end{aligned}

where cθ,αc_{\theta,\boldsymbol{\alpha}} is the copula density associated to (2). Because GG is almost surely discrete, P(ξi=ξi′)>0P(\boldsymbol{\xi}_i = \boldsymbol{\xi}_{i'}) > 0, and the distinct values of ξi\boldsymbol{\xi}_i define clusters that share both a dependence strength θ\theta and an asymmetry pattern α\boldsymbol{\alpha}. Smaller aa and bb produce more ties.

The centering measure G0G_0 is defined through its density g0g_0 given by,

g0(θ,α)=h(θ)∏j=1dBe⁡(αj∣cα,dα),(10)g_0(\theta,\boldsymbol{\alpha}) = h(\theta) \prod_{j=1}^{d} \operatorname{Be}(\alpha_j \mid c_\alpha,d_\alpha), \tag*{(10)}

where h(θ)h(\theta) is the centering density of the Archimedean parameter that varies according to its parameter space Θ\Theta. Specifically, we take Un⁡(θ∣−1,1)\operatorname{Un}(\theta\mid-1,1) for AMH, Ga⁡(θ∣cθ,dθ)\operatorname{Ga}(\theta\mid c_\theta,d_\theta) for Clayton, Ga⁡(θ−1∣cθ,dθ)\operatorname{Ga}(\theta- 1 \mid c_\theta,d_\theta) for Gumbel and Joe, and N⁡(θ∣μθ,λθ)\operatorname{N}(\theta\mid\mu_\theta,\lambda_\theta) for Frank. We restrict Clayton to θ>0\theta> 0: negative bivariate Clayton copulas have a singular boundary component and therefore are not represented by the absolutely continuous likelihood used here. Independent beta margins on the αj\alpha_j cover the whole unit interval and include the Archimedean corner in the support (the prior mass near α=1\alpha= 1 is controlled by cα,dαc_\alpha,d_\alpha). A more concentrated choice, such as Be⁡(αj∣8,1)\operatorname{Be}(\alpha_j \mid8,1), encodes the belief that variables in the data are close to exchangeable; Be⁡(αj∣1,1)\operatorname{Be}(\alpha_j \mid1,1) is a vaguer prior belief.

Marginalising GG yields the same generalised Pólya urn as in Pitman [17], now on Ξ\Xi:

f(ξi∣ξ−i)=b+amib+n−1g0(ξi)+∑j=1mini,j∗−ab+n−1δξi,j∗(ξi),(11)f(\boldsymbol{\xi}_i \mid\boldsymbol{\xi}_{-i}) = \frac{b + a m_i}{b + n - 1} g_0(\boldsymbol{\xi}_i) + \sum_{j=1}^{m_i} \frac{n^*_{i,j} - a}{b + n - 1} \delta_{\boldsymbol{\xi}^*_{i,j}}(\boldsymbol{\xi}_i), \tag*{(11)}

with the obvious notation for the mim_i distinct values ξi,j∗\boldsymbol{\xi}^*_{i,j} and frequencies ni,j∗n^*_{i,j} in ξ−i\boldsymbol{\xi}_{-i}.

Proposition 5 (Mixture Kendall tau) Let C(U,V)C(U,V) the marginal bivariate copula corresponding to mixture model (8) with ξ=(θ,α,β)\boldsymbol{\xi} = (\theta,\boldsymbol{\alpha},\boldsymbol{\beta}). Kendall’s tau for the pair (U,V)(U,V) is

τ=4∑k=1∞∑j=1∞ωkωj∬[0,1]2Cξk(u,v)cξj(u,v) dudv−1.(12)\tau= 4 \sum_{k=1}^{\infty} \sum_{j=1}^{\infty} \omega_k \omega_j \iint_{[0,1]^2} C_{\boldsymbol{\xi}_k}(u,v)c_{\boldsymbol{\xi}_j}(u,v)\,\mathrm{d}u\mathrm{d}v - 1. \tag*{(12)}

In general, the vector (τjl)j<l(\tau_{jl})_{j<l} for the pairs (Uj,Ul)(U_j,U_l) is not constrained to have equal entries.

Proof Substituting both the mixture copula (8) and its mixture density into (6), and then expanding their product, gives (12). The cross-component terms with k≠jk \ne j show, in particular, that Kendall’s tau is not obtained by simply averaging the component-specific tau’s.

The result in Proposition 5 is a correction of equation (9) in Pan et al. [16]. In § 3.3 we will provide a way of approximating (12) using the predictive mixture.

Posterior distributions

The full posterior conditional distribution of ξi=(θi,αi)\boldsymbol{\xi}_i = (\theta_i, \boldsymbol{\alpha}_i) is

f(ξi∣u,ξ−i)∝(b+ami)cξi(ui)g0(ξi)+∑j=1mi(ni,j∗−a)cξi(ui)δξi,j∗(ξi).f(\boldsymbol{\xi}_i \mid\mathbf{u}, \boldsymbol{\xi}_{-i}) \propto(b + a m_i)c_{\boldsymbol{\xi}_i}(\mathbf{u}_i)g_0(\boldsymbol{\xi}_i) + \sum_{j=1}^{m_i}(n^*_{i,j} - a)c_{\boldsymbol{\xi}_i}(\mathbf{u}_i)\delta_{\boldsymbol{\xi}_{i,j}^*}(\boldsymbol{\xi}_i).

The likelihood is not conjugate for g0g_0, so we sample with a Gibbs sampler ([21]) using Algorithm 8 of Neal [13], as in Pan et al. [16], with the atom now equal to ξ\boldsymbol{\xi} rather than θ\theta. Specifically,

  1. For each i=1,…,ni = 1,\ldots,n, draw rr auxiliary values ξ∗={ξmi+1∗,…,ξmi+r∗}\boldsymbol{\xi}^* = \{\boldsymbol{\xi}_{m_i+1}^*, \ldots, \boldsymbol{\xi}_{m_i+r}^*\} from g0g_0.

  1. Draw ξi\boldsymbol{\xi}_i from the discrete distribution on {ξi,1∗,…,ξi,mi∗}∪ξ∗\{\boldsymbol{\xi}_{i,1}^*,\ldots,\boldsymbol{\xi}_{i,m_i}^*\}\cup\boldsymbol{\xi}^* with masses proportional to (ni,j∗−a)cξi,j∗(ui)(n^*_{i,j}-a)c_{\boldsymbol{\xi}_{i,j}^*}(\mathbf{u}_i) on the existing atoms and to {(b+ami)/r}cξj∗(ui)\{(b+a m_i)/r\}c_{\boldsymbol{\xi}_j^*}(\mathbf{u}_i) on the auxiliary atoms.

  1. Let (ξ1∗,…,ξm∗)(\boldsymbol{\xi}_1^*,\ldots,\boldsymbol{\xi}_m^*) be the distinct values in {ξi}i=1n\{\boldsymbol{\xi}_i\}_{i=1}^n. Resample each ξℓ∗=(θℓ∗,αℓ∗)\boldsymbol{\xi}_\ell^*=(\theta_\ell^*,\boldsymbol{\alpha}_\ell^*) from

f(ξℓ∣c.c.)∝g0(ξℓ)∏{i:ξi=ξℓ∗}cξℓ(ui)f(\boldsymbol{\xi}_\ell\mid c.c.) \propto g_0(\boldsymbol{\xi}_\ell)\prod_{\{i:\boldsymbol{\xi}_i=\boldsymbol{\xi}_\ell^*\}}c_{\boldsymbol{\xi}_\ell}(\mathbf{u}_i)

by two Metropolis–Hastings random-walk steps, constrained to Ξ\Xi, where c.c.c.c. stands for clustering configuration. First, θℓ\theta_\ell is updated with a uniform proposal on (θℓ∗−κθ,θℓ∗+κθ)∩Θ(\theta_\ell^*-\kappa_\theta,\theta_\ell^*+\kappa_\theta)\cap\Theta, whose acceptance probability includes the ratio of the (truncated) proposal widths. Second, with θℓ\theta_\ell held fixed, each αℓ,j\alpha_{\ell,j} is updated on the logit scale,

logit⁡(αℓ,j′)=logit⁡(αℓ,j∗)+εj,εj∼N(0,κα2),\operatorname{logit}(\alpha_{\ell,j}')=\operatorname{logit}(\alpha_{\ell,j}^*)+\varepsilon_j,\qquad\varepsilon_j\sim\mathcal{N}(0,\kappa_\alpha^2),

where κα\kappa_\alpha is a standard deviation. The Gaussian increment is symmetric on the logit scale, so the proposal density on α\boldsymbol{\alpha} carries the logistic Jacobian and the vector αℓ′\boldsymbol{\alpha}_\ell' is accepted with probability

min⁡{1,f(θℓ∗,αℓ′∣c.c.)f(θℓ∗,αℓ∗∣c.c.)∏j=1dαℓ,j′(1−αℓ,j′)αℓ,j∗(1−αℓ,j∗)}.\min\left\{1,\frac{f(\theta_\ell^*,\boldsymbol{\alpha}_\ell'\mid c.c.)}{f(\theta_\ell^*,\boldsymbol{\alpha}_\ell^*\mid c.c.)}\prod_{j=1}^{d}\frac{\alpha_{\ell,j}'(1-\alpha_{\ell,j}')}{\alpha_{\ell,j}^*(1-\alpha_{\ell,j}^*)}\right\}.

The hyperparameters (a,b)(a,b) are assigned a beta prior for aa and a shifted-gamma prior for bb given aa of the form

f(a,b)=Ga⁡(b+a∣cb,db)Be⁡(a∣ca,da).f(a,b)=\operatorname{Ga}(b+a\mid c_b,d_b)\operatorname{Be}(a\mid c_a,d_a).

This is updated by the exchangeable partition probability function of Pitman [17]. The corresponding posterior conditional distributions are sampled using MH steps. For aa we use the proposal a′∼Un⁡(max⁡{0,a−κa},min⁡{a+κa,1})a'\sim\operatorname{Un}(\max\{0,a-\kappa_a\},\min\{a+\kappa_a,1\}), with the ratio of proposal widths in the acceptance probability, and for bb a random walk on the log scale, log⁡(b′+a)=log⁡(b+a)+ε\log(b'+a)=\log(b+a)+\varepsilon with ε∼N(0,κb2)\varepsilon\sim\mathcal{N}(0,\kappa_b^2), so that b′+a>0b'+a>0 automatically.

The tuning parameters (κθ,κα,κa,κb)(\kappa_\theta, \kappa_\alpha, \kappa_a, \kappa_b), started at (0.1,1,1,1)(0.1, 1, 1, 1), are adapted only during the burn-in period, in batches of 50 iterations, towards the acceptance window [0.3,0.4][0.3, 0.4] [19]. After batch j=1,2,…j = 1, 2, \ldots each tuning parameter κ\kappa is multiplied by exp⁡(±δj)\exp(\pm\delta_j) with δj=min⁡{0.05,j−0.6}\delta_j = \min\{0.05, j^{-0.6}\}, where the sign is positive if the acceptance rate AR(j)AR^{(j)} of the corresponding update in that batch exceeds the target 0.35 (the centre of the window) and negative otherwise; the acceptance counters are then reset. The adaptation step δj\delta_j vanishes as jj grows, and it stops altogether at the end of burn-in. Hence the retained draws come from a Markov chain with fixed transition kernel (finite adaptation), rather than from a chain that keeps adapting until the end of the run, and no additional ergodicity conditions for adaptive MCMC are required.

Fitting measure, pairwise tau, and clustering

The log pseudo-marginal likelihood LPML=∑i=1nlog⁡(CPOi)\mathrm{LPML} = \sum_{i=1}^{n} \log(\mathrm{CPO}_i) is computed using a Monte Carlo approximation of the conditional predictive ordinate CPOi=f(ui∣u−i)\mathrm{CPO}_i = f(\mathbf{u}_i \mid\mathbf{u}_{-i}) [6, 15] as

CPO^i=[1L∑l=1L{ξi(l)(ui)}−1]−1\widehat{\mathrm{CPO}}_i = \left[\frac{1}{L}\sum_{l=1}^{L}\left\{\xi_i^{(l)}(\mathbf{u}_i)\right\}^{-1}\right]^{-1}

where ξi(l)=(θi(l),αi(l))\boldsymbol{\xi}_i^{(l)} = (\theta_i^{(l)}, \boldsymbol{\alpha}_i^{(l)}) is the atom to which observation ii is allocated at retained draw ll. The same estimator is used for all models and data sets in § 4, so that our LPML values are directly comparable with those reported by Pan et al. [16]. As a check, we also computed the harmonic mean of the Poisson–Dirichlet predictive densities, which uses the whole mixture; for the 102 simulation fits it differs from the estimator we use by less than 3 LPML units in 99 fits (see the Github repository).

Pairwise Kendall tau’s are estimated from simulations from the complete posterior predictive mixture. This retains the cross-component terms in (12). A specific clustering is obtained from the posterior similarity matrix by the SALSO search of Dahl et al. [4] with Binder loss; given the chosen partition we report for each cluster ℓ\ell, the posterior of (θℓ∗,αℓ∗)(\theta_\ell^*, \boldsymbol{\alpha}_\ell^*) together with its specific weight (nℓ∗−a)/(b+n)(n_\ell^* - a)/(b + n).

The cluster-specific vector αℓ∗\boldsymbol{\alpha}_\ell^* is the new inferential object. Coordinates with αℓ,j∗\alpha_{\ell,j}^* near one are those through which cluster ℓ\ell transmits Archimedean dependence; coordinates with αℓ,j∗\alpha_{\ell,j}^* near zero are those that cluster ℓ\ell treats as independent of the rest.

Numerical experiments

All programs used in this section, together with the configuration files, seeds and scripts, are available at https://github.com/rcraiu/NPB_Asymmetric_Archimedean. We translated, with the help of Cursor ([1]), some of our subroutines written in R into C++ via the Rcpp framework.

The experiments below extend those of Pan et al. [16]. Throughout, the centering measure (10) is defined by cθ=4c_\theta= 4, dθ=1d_\theta= 1 for Clayton, Gumbel and Joe, and μθ=0\mu_\theta= 0 and λθ=4\lambda_\theta= 4 for Frank. Additionally, cα=dα=1c_\alpha= d_\alpha= 1 unless otherwise stated, and for the prior on (a,b)(a,b) we take ca=cb=1c_a = c_b = 1 together with da=db=20d_a = d_b = 20 that favours a small number of components. MCMC is run for 15,000 iterations with a burn-in of 5,000 and a thinning of 5. Simulations 1, 3 and 4 each use one data set generated from the seed recorded in the reproducibility files, whereas Simulation 2 uses 20 seeded data sets; in all cases the competing kernels and the exchangeable mixture of Pan et al. [16] (the current model with α=1\alpha= 1) are fitted to identical observations.

Simulation 1: a single asymmetric component

We generated n∈{200,500}n \in\{200,500\} observations from a bivariate Khoudraji–Clayton copula with θ=5\theta= 5 and (α,β)∈{(1,1),(0.8,0.3),(0.4,0.4),(0,0)}(\alpha,\beta) \in\{(1,1),(0.8,0.3),(0.4,0.4),(0,0)\}. The four choices are the exchangeable Clayton copula, an asymmetric copula, a symmetric but diluted copula, and independence. We fitted (9) and the exchangeable mixture with Clayton, Frank and Gumbel kernels.

Table 2 gives the results. When the truth is exchangeable, the extended model sends both Khoudraji coordinates towards one: at n=500n = 500 the posterior means range from (0.96,0.96)(0.96,0.96) to (0.99,0.99)(0.99,0.99). The exchangeable model has the larger LPML for all three kernels, as expected when the extra parameters are unnecessary. For the asymmetric truth, the distinction becomes clear at n=500n = 500. The Clayton fit recovers (α^,β^)=(0.86,0.35)(\hat{\alpha},\hat{\beta}) = (0.86,0.35) and raises LPML from 27.15 to 44.32; the Frank and Gumbel gains are 10.66 and 5.22, respectively. The corresponding posterior-predictive tau under Clayton is 0.230, close to the empirical 0.239. At n=200n = 200 the sample tau is only 0.093, and consequently the direction is weakly identified.

True α\alphannKernelLPML ExchLPML Khouθ^\hat{\theta}α^\hat{\alpha}τ~\tilde{\tau}τ^\hat{\tau}
(1, 1)200CLA208.15206.385.90(0.98, 0.99)0.7580.725
(1, 1)200FRA167.23166.2811.89(0.99, 0.97)0.7580.684
(1, 1)200GUM158.22142.585.07(0.92, 0.91)0.7580.660
(0.8, 0.3)200CLA-0.210.583.93(0.41, 0.34)0.0930.085
(0.8, 0.3)200FRA1.190.641.42(0.55, 0.54)0.0930.060
(0.8, 0.3)200GUM0.36-0.874.05(0.21, 0.22)0.0930.060
(0.4, 0.4)200CLA1.053.454.47(0.48, 0.37)0.1420.138
(0.4, 0.4)200FRA3.752.922.11(0.67, 0.59)0.1420.113
(0.4, 0.4)200GUM1.620.663.76(0.48, 0.18)0.1420.068
(0, 0)200CLA-3.48-0.833.62(0.22, 0.28)-0.0530.010
(0, 0)200FRA-0.450.18-1.18(0.45, 0.45)-0.053-0.022
(0, 0)200GUM-3.52-1.754.75(0.14, 0.15)-0.0530.024
(1, 1)500CLA491.91490.144.78(0.99, 0.99)0.7330.690
(1, 1)500FRA374.46374.1311.40(0.98, 0.98)0.7330.679
(1, 1)500GUM342.88319.764.42(0.96, 0.96)0.7330.612
(0.8, 0.3)500CLA27.1544.325.08(0.86, 0.35)0.2390.230
(0.8, 0.3)500FRA30.2840.945.39(0.94, 0.43)0.2390.230
(0.8, 0.3)500GUM22.0127.232.51(0.84, 0.31)0.2390.190
(0.4, 0.4)500CLA18.9821.255.14(0.52, 0.41)0.1870.190
(0.4, 0.4)500FRA19.3719.913.23(0.74, 0.62)0.1870.167
(0.4, 0.4)500GUM20.8219.201.82(0.55, 0.43)0.1870.161
(0, 0)500CLA-2.60-0.603.56(0.37, 0.18)-0.0120.021
(0, 0)500FRA-0.84-0.45-0.30(0.40, 0.42)-0.012-0.002
(0, 0)500GUM-2.64-1.464.56(0.21, 0.20)-0.0120.019

Table 2. Simulation 1: data generated from a single Clayton-Khoudraji model with θ=5\theta= 5 and the true α\boldsymbol{\alpha} given in the first column. LPML for the exchangeable and Khoudraji mixtures; occupancy-weighted posterior means under the Khoudraji model; sample tau τ~\tilde{\tau}; and posterior-predictive tau τ^\hat{\tau}.

For the diluted symmetric truth, extended Clayton improves LPML by 2.27 at n=500n = 500 and estimates (0.52,0.41)(0.52,0.41); the Frank gain is small and Gumbel does not improve. Under independence, all posterior-predictive tau’s are close to zero and the broad posteriors of both α\alpha and θ\theta reflect their weak identification. Thus, the experiment supports recovery of substantial asymmetry, but also shows that the Khoudraji coordinates should not be interpreted componentwise when dependence is nearly absent.

Simulation 2: a mixture of mirror-asymmetric components

This experiment asks whether the Khoundraji coordinates α\alpha allow the mixture to separate clusters that an exchangeable kernel can not separate. We generated n=500n = 500 observations from the equally weighted two-component Clayton–Khoudraji mixture

c(u,v)=12cθ,0.1,0.65(u,v)+12cθ,0.65,1(u,v),θ=50.c(u,v) = \frac{1}{2}c_{\theta,0.1,0.65}(u,v) + \frac{1}{2}c_{\theta,0.65,1}(u,v), \qquad\theta= 50.

The components are mirror images: they share θ\theta and hence the same Kendall tau (≈0.63\approx0.63) and marginal behaviour, and differ only in the direction of asymmetry. Every exchangeable Archimedean kernel is invariant under the swap (u,v)↦(v,u)(u,v) \mapsto(v,u), so any exchangeable mixture assigns a point and its reflection the same density; the Bayes allocation rule based on

best exchangeable pair of kernels therefore has accuracy 0.50 by symmetry, whereas the allocation rule using the true Khoudraji parameters has accuracy about 0.87 and adjusted Rand index (ARI) about 0.55. The ARI is modest because the two components overlap near the diagonal and in the upper corner.

The prior on θ\theta is Ga(4, 0.08) (mean 50) so that the true value is not in the far tail of the default Ga(4, 1), and the prior on b+ab+a is Ga(2, 10); all other settings are as above. We report the following recovery summaries against the generating labels: the posterior-mean ARI, i.e. the ARI of each retained allocation vector, averaged over draws; the ARI of the SALSO point estimate under Binder loss, restricted to at most two clusters, and under VI loss. The whole experiment is repeated on 20 independent seeded data sets. The oracle fit is that of the Bayes allocation rule using the true parameters.

A single data set. The featured data set (replicate 1) has 252 and 248 observations from the two components and overall sample tau 0.608. Table 3 shows that the extended model raises LPML from 166.58 to 527.10. The exchangeable fit places all observations in one group under Binder and VI loss (ARI 0.00), while its posterior-predictive tau of 0.477 is well below the sample value. The Khoudraji fit gives ARI 0.52 under unrestricted Binder loss, which selects 7 groups (two large groups and small fragments), and 0.54 when Binder loss is restricted to at most two groups, close to the oracle benchmark. These two groups have sizes 246 and 254, and occupancy-weighted posterior means (θ^,α^)=(51.58,0.92,0.69)(\hat{\theta},\hat{\alpha})=(51.58,0.92,0.69) and (54.56,0.68,0.91)(54.56,0.68,0.91) (Table 4): the fit recovers both θ\theta and the mirrored directions of asymmetry. Figure 1 shows that the Khoudraji predictive law reproduces the two ridges on either side of the diagonal, whereas the exchangeable predictive law merges them into a single cloud, and that the PSM has the expected two-block structure. The VI summary, in contrast, merges the Khoudraji fit into a single group (ARI 0.00); this sensitivity to the loss function is reported rather than suppressed, and it reflects genuine posterior uncertainty: the mean posterior co-clustering probability is 0.67 for pairs from the same true component and 0.33 for pairs from different components.

ModelLPMLτ^\hat{\tau}m^\hat{m}ARI^\widehat{\mathrm{ARI}}KBK_BARIB\mathrm{ARI}_BARIB,2\mathrm{ARI}_{B,2}KVK_VARIV\mathrm{ARI}_V
Exch166.580.4771.990.0010.000.0010.00
Khou527.100.5512.600.3470.520.5410.00

Table 3. Simulation 2, replicate 1: data generated from the equally weighted mixture of two Clayton-Khoudraji copulas with θ=50\theta=50 and (α,β)=(1,0.65)(\alpha,\beta)=(1,0.65) and (0.65,1)(0.65,1) (n=500n=500). Posterior summaries: LPML, Kendall’s tau τ^\hat{\tau}, number of groups m^\hat{m}, and ARI; KBK_B and KVK_V are the numbers of groups selected by unrestricted Binder and VI loss with their respective ARIB\mathrm{ARI}_B and ARIV\mathrm{ARI}_V; ARIB,2\mathrm{ARI}_{B,2} is the Binder result constrained to at most two groups.

PartitionClusterSizeθ^\hat{\theta}α^\hat{\boldsymbol{\alpha}}
Binder K≤2K \le2124651.58(0.92, 0.69)
Binder K≤2K \le2225454.56(0.68, 0.91)
VI150053.09(0.80, 0.80)

Table 4. Simulation 2, replicate 1: data generated from the equally weighted mixture of two Clayton-Khoudraji copulas with θ=50\theta= 50 and (α,β)=(1,0.65)(\alpha, \beta) = (1, 0.65) and (0.65,1)(0.65, 1) (n=500n = 500). Cluster sizes and occupancy-weighted parameter summaries for the Khoudraji Clayton fit under the two-cluster Binder and unrestricted VI summaries.

Observed, by true component; Exchangeable predictive; Khoudraji predictive; Khoudraji posterior similarity

Figure 1. Simulation 2, replicate 1: data generated from the equally weighted mixture of two Clayton-Khoudraji copulas with θ=50\theta= 50 and (α,β)=(1,0.65)(\alpha, \beta) = (1, 0.65) and (0.65,1)(0.65, 1) (n=500n = 500). Observed pseudo-observations coloured by true component, posterior-predictive contours for the exchangeable and Khoudraji Clayton mixtures, and the Khoudraji posterior similarity matrix with observations ordered by true component.

Twenty data sets. Table 5 and Figure 2 summarize all replicates. The extended model has the larger LPML in every data set, around 348.8 on average (range 296.7 to 416.2). Cluster recovery tells the same story: the exchangeable model has ARI essentially 0 in every data set (it selects a single group in 19 of 20 and two groups once), whereas the extended model attains a mean ARI of 0.55 (standard deviation 0.05) under both unrestricted Binder loss and the PSM cut, matching the oracle benchmark of 0.55 (standard deviation 0.04); the average difference from the oracle is −0.002. The Binder K≤2K \le2 summary returns two groups with opposite orientation (α^>β^\hat{\alpha} > \hat{\beta} in one group and α^<β^\hat{\alpha} < \hat{\beta} in the other) in all 20 Khoudraji fits and in none of the exchangeable fits. The posterior-mean ARI is lower (0.37), reflecting the genuine posterior uncertainty about the allocation of observations near the diagonal and in the upper corner. Unrestricted Binder loss selects between 2 and 11 groups (mean 7.7), but the additional groups are small fragments that do not change the ARI, and VI loss merges the extended fit into a single group in all 20 data sets. We therefore recommend Binder (possibly restricted) or PSM-based summaries for reporting cluster recovery here.

Summary over 20 data setsKhouExch
LPML485.4 (31.5)136.6 (12.7)
m^\hat{m}2.76 (0.13)1.61 (0.37)
KBK_B, Binder7.7 (2.1)1.1 (0.2)
ARI, Binder0.55 (0.05)0.00 (0.00)
ARI, Binder (K≤2K \le2)0.55 (0.05)0.00 (0.00)
ARI, VI0.00 (0.00)0.00 (0.00)
ARI, PSM cut at K=2K = 20.55 (0.04)0.00 (0.00)
ARI0.37 (0.03)0.00 (0.00)
PSM within − between0.37 (0.03)0.00 (0.00)
Oracle ARI0.55 (0.04)
Khou beats Exch in LPML / ARIB,2\mathrm{ARI}_{B,2}20 / 20 of 20
True clusters found, Binder (K≤2K \le2)20 of 200 of 20

Table 5. Simulation 2 over 20 seeded data sets, each generated from the equally weighted mixture of two Clayton-Khoudraji copulas with θ=50\theta= 50 and (α,β)=(1,0.65)(\alpha, \beta) = (1, 0.65) and (0.65,1)(0.65, 1) (n=500n = 500). Mean (standard deviation) of each summary under the Khoudraji and exchangeable models.

Figure 2. Simulation 2 over 20 seeded data sets, each generated from the equally weighted mixture of two Clayton-Khoudraji copulas with θ=50\theta= 50 and (α,β)=(1,0.65)(\alpha, \beta) = (1, 0.65) and (0.65,1)(0.65, 1) (n=500n = 500). Left: ARI of the PSM cut at K=2K = 2 for the exchangeable and Khoudraji fits and for the oracle allocation rule. Centre: posterior-mean ARI. Right: LPML difference, Khoudraji minus exchangeable.

Simulation 3: unequal pairwise tau’s in dimension three

We generate n=1000n = 1000 observations from a three-dimensional Khoudraji–Gumbel copula with θ=4\theta= 4 and α=(1,0.7,0.2)\boldsymbol{\alpha} = (1, 0.7, 0.2). The three sample pairwise tau’s are (0.553,0.188,0.182)(0.553, 0.188, 0.182). The exchangeable mixture compromises at approximately (0.297,0.293,0.294)(0.297, 0.293, 0.294), whereas the extended model gives (0.560, 0.196, 0.184) (Table 6). It also recovers α\alpha almost exactly: its posterior means are (0.995, 0.702, 0.214), with marginal 95% credible intervals (0.985, 1.000), (0.672, 0.732) and (0.200, 0.229), respectively. The resulting LPML gain is 470.34. Clearly, the Khoudraji layer provides a dependence structure that no mixture of exchangeable Archimedean kernels can reproduce.

ModelLPMLα^\hat{\alpha}τ12\tau_{12}τ13\tau_{13}τ23\tau_{23}
Sample——0.5530.1880.182
Exch317.17(1.00,1.00,1.00)0.2970.2930.294
Khou787.51(1.00,0.70,0.21)0.5600.1960.184

Table 6. Simulation 3: n=1000n = 1000 observations generated from a three-dimensional Gumbel-Khoudraji copula with θ=4\theta= 4 and α=(1,0.7,0.2)\boldsymbol{\alpha} = (1,0.7,0.2). LPML, occupancy-weighted posterior mean of α\boldsymbol{\alpha}, and posterior-predictive pairwise Kendall tau’s.

Simulation 4: Gaussian copulas

We revisit Simulation 2 of Pan et al. [16]: a single Gaussian copula with ρ=0.7\rho= 0.7 and an equally weighted two-component Gaussian mixture with ρ∈{−0.7,0.7}\rho\in\{-0.7,0.7\}. Table 7 and Figure 3 show that the additional coordinates do not generally solve this misspecification. For the single Gaussian copula, Khoudraji Clayton gains 16.34 LPML units over exchangeable Clayton, but exchangeable Frank remains overall best (LPML 153.69) and its predictive tau 0.493 closely matches the empirical 0.491. The extended Frank and Gumbel fits place α\alpha close to 1 and lose 1.43 and 2.41 LPML units, respectively.

TruthKernelLPML ExchLPML KhouΔ\Deltaα^\hat{\alpha}τ^\hat{\tau} Exchτ^\hat{\tau} Khou
ρ=0.7\rho= 0.7CLA126.72143.06+16.34(0.85,0.78)0.4280.427
ρ=0.7\rho= 0.7FRA153.69152.26-1.43(0.97,0.96)0.4930.480
ρ=0.7\rho= 0.7GUM141.51139.09-2.41(0.97,0.97)0.4590.438
ρ={−0.7,0.7}\rho= \{-0.7,0.7\}CLA6.9212.89+5.97(0.37,0.33)0.0930.117
ρ={−0.7,0.7}\rho= \{-0.7,0.7\}FRA54.1953.63-0.56(0.93,0.92)0.0050.000
ρ={−0.7,0.7}\rho= \{-0.7,0.7\}GUM15.5210.71-4.80(0.19,0.38)0.1160.098

Table 7. Simulation 4: data generated from a Gaussian copula, either a single component with ρ=0.7\rho= 0.7 or an equally weighted mixture of two Gaussian copulas with ρ=−0.7\rho= -0.7 and ρ=0.7\rho= 0.7 (first column; n=500n = 500). LPML and posterior-predictive tau for exchangeable and Khoudraji models. Here Δ=LPMLKhou−LPMLExch\Delta= \mathrm{LPML}_{\mathrm{Khou}} - \mathrm{LPML}_{\mathrm{Exch}} and α^\hat{\alpha} is reported for the Khoudraji fit.

Simulation 4 posterior-predictive contours and observed samples

Figure 3. Simulation 4: data generated from a Gaussian copula with ρ=0.7\rho= 0.7 (top two rows) or from an equally weighted mixture of Gaussian copulas with ρ=−0.7\rho= -0.7 and ρ=0.7\rho= 0.7 (bottom two rows), n=500n = 500. Observed samples (first column) and posterior-predictive contours; within each pair of rows, exchangeable fits appear above Khoudraji fits.

For the two-component Gaussian mixture the sample tau is only 0.038, despite strong dependence within each component. Frank mixtures capture the opposing signs through two principal clusters and dominate the one-sided Clayton and Gumbel families. Exchangeable Frank attains LPML 54.19, compared with 53.63 for extended Frank; its predictive tau is 0.005. Khoudraji Clayton improves upon exchangeable Clayton but remains more than 40 LPML units behind Frank. Here, the limitation is the generator’s ability to combine positive and negative dependence, not exchangeability, and the Khoudraji coordinates cannot compensate for it.

Occupancy and red-wine data

We refit the two data sets of Pan et al. [16] with the Khoudraji mixture (9), using the same modified rank transform, the same centering measure g0g_0 and the same MCMC length: 15,000 iterations, burn-in 5,000, thinning 5, with the tuning constants adapted during burn-in only (§ 3.2). Copula densities are evaluated in compiled C++, and LPML is computed exactly as in the simulations (§ 3.3), which is the estimator of Pan et al. [16]. The exchangeable model α=1\alpha= 1 is the Poisson–Dirichlet mixture of Pan et al. [16]; comparing the two LPML values isolates the contribution of the Khoudraji device.

Occupancy. We take occupancy measurements of Candanedo and Feldheim [2]: five-minute averages of CO2_2 and humidity ratio on 5 February 2015, first-difference, and rank-transform (n=287n = 287). The sample Kendall tau is 0.481, as seen in Table 8 which reports the LPML for the exchangeable and Khoudraji models for 4 copula families. For Gumbel, the exchangeable refit (115.5) and the Khoudraji model (113.9) are practically indistinguishable: the posterior of (α1,α2)(\alpha_1,\alpha_2) concentrates near one, with means (0.91, 0.92). The star-shaped scatter is therefore exchangeable; asymmetry is not needed, and a single Gumbel component with θ≈2.4\theta\approx2.4 remains the preferred description. In no occupancy model does the Khoudraji layer improve on the exchangeable fit.

ModelExchKhoudraji
AMH76.774.8
CLA92.481.5
FRA91.287.2
GUM115.5113.9

Table 8. Occupancy data (real data: the generating model is unknown). LPML for the exchangeable model (α=1\alpha= 1), and the Khoudraji mixture.

Red wine. The three-dimensional red-wine data of Cortez et al. [3] has size n=1,599n = 1{,}599 and three variables: fixed acidity, citric acid, density. The sample pairwise tau’s are: τ~12=0.484\tilde{\tau}_{12} = 0.484, τ~13=0.457\tilde{\tau}_{13} = 0.457 and τ~23=0.245\tilde{\tau}_{23} = 0.245. An exchangeable Archimedean mixture must use a single τ\tau for all three pairs. Table 9 shows our exchangeable and Khoudraji fits. Allowing α\alpha to vary raises LPML, relative to the exchangeable refits, by 158 for Gumbel, 133 for Frank and 204 for Clayton. The Gumbel Khoudraji mixture is the preferred model. Its posterior mean of α\alpha is (0.93, 0.65, 0.66): fixed acidity stays almost fully in the Archimedean factor, while citric acid and density are diluted, which is exactly the pair with the smaller empirical tau. The posterior mode of the number of components is 4, against 2 in the exchangeable Gumbel fit.

ModelExchKhoudrajiα^\hat{\alpha}
CLA607.9811.9(0.75,0.58,0.60)(0.75, 0.58, 0.60)
FRA679.6812.8(0.94,0.69,0.62)(0.94, 0.69, 0.62)
GUM716.4874.1(0.93,0.65,0.66)(0.93, 0.65, 0.66)

Table 9. Red wine data (real data: the generating model is unknown). LPML for the exchangeable mixture model (α=1\alpha= 1), and the Khoudraji mixture model, together with the posterior mean of (α1,α2,α3)(\alpha_1, \alpha_2, \alpha_3) for fixed acidity, citric acid and density.

Concluding remarks

The Poisson–Dirichlet mixture of Archimedean copulas of Pan et al. [16] inherits the exchangeability of its kernel. Replacing that kernel by the Khoudraji transform (2) keeps the computational structure of their sampler — a Pólya urn on atoms, Neal’s Algorithm 8, and the same updates for (a,b)(a,b) — and adds a vector α∈[0,1]d\boldsymbol{\alpha} \in[0,1]^d that interpolates, coordinate-wise, between Archimedean dependence and independence.

Pairwise Kendall tau’s are then allowed to differ, lower-tail dependence requires αj=αk=1\alpha_j = \alpha_k = 1 while upper-tail dependence (Gumbel, Joe) survives for αj,αk>0\alpha_j, \alpha_k > 0, and clusters may differ in both the strength θ\theta and the direction α\boldsymbol{\alpha} of association. The exchangeable model is obtained when α=1\boldsymbol{\alpha} = 1, so the two specifications are comparable by LPML.

For the occupancy data, the posterior of α\boldsymbol{\alpha} concentrates at 1 and the Khoudraji Gumbel mixture matches the exchangeable fit of Pan et al. [16]. On the red-wine data, where the three pairwise tau’s are not equal, the same Gumbel kernel with an extended α\boldsymbol{\alpha} improves LPML by about 160 and loads Archimedean dependence on fixed acidity while diluting the citric-acid–density pair.

Future work will consider mixtures of vines or hierarchical Archimedean copulas (Hofert and Pham, [8]; Li et al., 2021).

Acknowledgements

The first author was supported by NSERC of Canada RGPIN-2024-04506 Discovery Grant. The second author was supported by Asociación Mexicana de Cultura, A.C..

References

  1. [1]ANYSPHERE, INC. (2026). Cursor: AI code editor. https://cursor.com. Software.
  2. [2]CANDANEDO, L. M. and FELDHEIM, V. (2016). Accurate occupancy detection of an office room from light, temperature, humidity and CO₂ measurements using statistical learning models. Energy and Buildings 112 28–39.DOI
  3. [3]Cortez, P., Cerdeira, A., Almeida, F., Matos, T. and Reis, J. (2009). Modeling wine preferences by data mining from physicochemical properties. Decision Support Systems 47 547–553.DOI
  4. [4]Dahl, D. B., Johnson, D. J. and Müller, P. (2022). Search algorithms and loss functions for Bayesian clustering. Journal of Computational and Graphical Statistics 31 1189–1201.arxiv.org/abs/2105.04451
  5. [5]Durante, F. and Salvadori, G. (2010). On the construction of multivariate extreme value models via copulas. Environmetrics 21 143–161.DOI
  6. [6]Geisser, S. and Eddy, W. F. (1979). A predictive approach to model selection. Journal of the American Statistical Association 74 153–160.DOI
  7. [7]Genest, C., Ghoudi, K. and Rivest, L.-P. (1998). Discussion of “Understanding relationships using copulas”, by E. Frees and E. Valdez. North American Actuarial Journal 2 143–149.
  8. [8]Hofert, M. and Pham, D. (2013). Densities of nested Archimedean copulas. Journal of Multivariate Analysis 118 37–52.arxiv.org/abs/1204.2410
  9. [9]Khoudraji, A. (1995). Contributions à l’étude des copules et à la modélisation des valeurs extrêmes bivariées. Ph.D. thesis, Université Laval, Québec.
  10. [11]Liebscher, E. (2008). Construction of asymmetric multivariate copulas. Journal of Multivariate Analysis 99 2234–2250.DOI
  11. [12]McNeil, A. J. and Nešlehová, J. (2009). Multivariate Archimedean copulas, d-monotone functions and ℓ₁-norm symmetric distributions. The Annals of Statistics 37 3059–3097.
  12. [13]Neal, R. M. (2000). Markov chain sampling methods for Dirichlet process mixture models. Journal of Computational and Graphical Statistics 9 249–265.DOI
  13. [14]Nelsen, R. B. (2006). An Introduction to Copulas. 2nd ed. Springer, New York.
  14. [15]Nieto-Barajas, L. E. and Contreras-Cristán, A. (2014). A Bayesian nonparametric approach for time series clustering. Bayesian Analysis 9 147–170.
  15. [16]Pan, R., Nieto-Barajas, L. E. and Craiu, R. V. (2025). Bayesian nonparametric mixtures of Archimedean copulas. Journal of Agricultural, Biological and Environmental Statistics.arxiv.org/abs/2412.09539
  16. [17]Pitman, J. (1995). Exchangeable and partially exchangeable random partitions. Probability Theory and Related Fields 102 145–158.DOI
  17. [18]Pitman, J. and Yor, M. (1997). The two-parameter Poisson–Dirichlet distribution derived from a stable subordinator. The Annals of Probability 25 855–900.DOI
  18. [19]Roberts, G. O. and Rosenthal, J. S. (2009). Examples of adaptive MCMC. Journal of Computational and Graphical Statistics 18 349–367.DOI
  19. [20]Sklar, A. (1959). Fonctions de répartition à n dimensions et leurs marges. Publ. Inst. Statist. Univ. Paris 8 229–231.
  20. [21]Smith, A. F. M. and Roberts, G. O. (1993). Bayesian computation via the Gibbs sampler and related Markov chain Monte Carlo methods. Journal of the Royal Statistical Society: Series B (Methodological) 55 3–23.DOI

Paper details

vlnc:W_cyywduxj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents