--- title: Empirical Discrete Copulas url: https://www.emergentmind.com/topics/empirical-discrete-copula type: topic --- # Empirical Discrete Copulas An empirical discrete copula is a nonparametric representation of dependence for data observed on a finite rank grid or with genuinely discrete margins. In the literature, the term covers several related constructions: the rank-based empirical copula, which is itself a discrete step function on a finite grid; the multilinear or checkerboard copula, which continuously interpolates that grid object; and, more recently, the empirical discrete copula array defined as the Csiszár \(I\)-projection of empirical frequencies onto the polytope of uniform-margin probability arrays. These constructions address a common difficulty: for discrete margins, Sklar’s theorem no longer yields a unique copula on \([0,1]^d\), so dependence must be represented either on the marginal ranges, by an interpolated extension, or by a support-aware projection principle [1407.1200], [2506.12316], [1901.08741]. ## 1. Conceptual basis and the discrete complication For continuous margins, a copula \(C\) is the joint distribution of the transformed uniforms \(U_j=F_j(X_j)\), and the empirical copula is naturally rank-based. If \(X_1,\dots,X_n\) are i.i.d. with continuous margins, the empirical copula can be written as \[ C_n(u) \;=\; \frac{1}{n}\sum_{i=1}^n \mathbf{1}\bigl\{\hat U_{i1}\le u_1,\ldots,\hat U_{id}\le u_d\bigr\}, \] with pseudo-observations \(\hat U_{ij}=R_{ij}/(n+1)\) or \(R_{ij}/n\). This object is already discrete: it is a step function supported on a finite rank grid and changes only at grid points determined by the ranks [1012.2133]. The discrete-margin case is structurally different. Sklar’s theorem identifies a copula uniquely only on \(\mathrm{Ran}(F_1)\times\cdots\times\mathrm{Ran}(F_d)\), or, equivalently, on the sample space induced by the marginal jumps. Multiple copulas on \([0,1]^d\) can generate the same joint discrete law through different allocations of mass to rectangles in the unit cube. This non-uniqueness is not merely formal; it affects likelihood construction, interpretation of dependence, and the meaning of “the” copula for discrete data [1304.0905], [1901.08741]. A central consequence is that empirical dependence for discrete data cannot be reduced to a single universally accepted object. The main responses in the literature are: retaining the rank-grid empirical copula as a discrete object; extending it by multilinear interpolation; or defining a canonical uniform-margin representative by KL projection on finite supports. ## 2. Principal empirical constructions The main empirical constructions differ in both target object and inferential intent. | Construction | Core object | Representative source | |---|---|---| | Rank-based empirical copula | Step function on the finite rank grid | [1012.2133] | | Multilinear or checkerboard empirical copula | Continuous interpolation of a discrete empirical copula | [1407.1200] | | Empirical beta copula | Bernstein smoothing with degree equal to sample size | [1607.04430] | | Empirical discrete copula array | \(I\)-projection of empirical frequencies onto uniform margins | [2506.12316] | | Copula-like discrete pmf on rectangular supports | \(I\)-projection with IPFP/Sinkhorn computation | [2307.04225] | The rank-based empirical copula is the direct empirical cdf of pseudo-observations. Its strengths are exact rank equivariance and immediate computability, but it remains a step function and does not resolve the non-uniqueness created by ties or genuinely discrete margins. The checkerboard construction replaces the step function by multilinear interpolation inside each cell of the rank grid. In one sample-based form, \[ C_n^{\#}(u) := \frac{1}{n} \sum_{i=1}^n \prod_{j=1}^d \min\{ \max(nu_j-R_{ij}^{(n)}+1,0),1\}, \] which yields a genuine copula in finite samples and coincides with multilinear extension on the empirical grid [1607.04430]. The empirical beta copula goes one step further by replacing indicator functions with Beta cdfs: \[ C_n^{\beta}(u_1,\dots,u_d)=\frac{1}{n}\sum_{i=1}^n\prod_{j=1}^d B(u_j;R_{ij}^{(n)},n+1-R_{ij}^{(n)}). \] It is exactly the empirical Bernstein copula with all polynomial degrees equal to \(n\), hence a genuine copula without an external smoothing parameter [1607.04430]. The projection-based construction is different in kind. Instead of starting from ranks, it starts from an empirical joint probability array \(\hat Q_n\) on a finite support and defines the empirical discrete copula as \[ \hat C_n \;=\; \operatorname*{arg\,min}_{P\in\mathcal U} D_{\mathrm{KL}}(P\Vert \hat Q_n), \] where \(\mathcal U\) is the polytope of arrays with uniform margins. Here the copula is not an interpolated cdf on \([0,1]^d\) but a uniform-margin pmf array that preserves the dependence structure encoded by the empirical support and cell frequencies [2506.12316]. ## 3. Multilinear and checkerboard formulations for count data For count data, the multilinear or checkerboard copula plays a central role because it converts a discontinuous cdf into a continuous function by interpolation within each marginal jump cell. If \(X=(X_1,\dots,X_d)\) has discrete margins \(F_j\), and \(u_j^-,u_j^+\) are the adjacent points of the marginal range bracketing \(u_j\), the interpolation weights are \[ \lambda_{F_j}(u_j)= \begin{cases} \dfrac{u_j-u_j^-}{u_j^+-u_j^-}, & u_j^-\neq u_j^+,\\[1ex] 1,&\text{otherwise,} \end{cases} \] and the multilinear copula is \[ C(u_1,\dots,u_d)= \sum_{S\subseteq\{1,\dots,d\}} \lambda_{H,S}(u)\, H\{F_1^{-1}(u_{S_1}),\dots,F_d^{-1}(u_{S_d})\}, \] with \(\lambda_{H,S}(u)\) the product of the marginal interpolation weights [1407.1200]. The empirical version replaces \(H\) and \(F_j\) by their empirical counterparts. This yields an empirical multilinear copula process \[ \widehat{\mathbb C}_n=\sqrt{n}(\widehat C_n-C), \] whose asymptotics are more delicate than in the continuous case. In general, \(\widehat{\mathbb C}_n\) does not converge in law on \(\mathcal C([0,1]^d)\) with the uniform norm, because the partial derivatives of \(C\) are discontinuous along the Cartesian product of the marginal ranges. The process does converge in \(\mathcal C(K)\) for compact \(K\subset\mathcal O\), where \(\mathcal O\) is the dense open set obtained by removing the Cartesian product of the marginal ranges [1407.1200]. That restricted weak convergence is still strong enough to recover asymptotic distributions for classical dependence functionals. The same framework yields limits for discrete-data versions of Kendall’s tau and Spearman’s rho and produces a Cramér–von Mises-type test of independence, \[ S_n=n\int_{[0,1]^d}\{\widehat C_n(u)-\Pi(u)\}^2\,d\Pi(u), \] which is consistent even for sparse contingency tables and can be calibrated by multiplier bootstrap [1407.1200]. This construction also underlies later smoothing schemes. In particular, the empirical checkerboard copula is used as the finite-sample genuine copula that feeds hierarchical Bernstein-type smoothers in arbitrary dimension, including empirical Bayes degree selection for multivariate Bernstein copulas [2112.10351]. ## 4. I-projection, odds ratios, and copula-like decomposition The \(I\)-projection approach defines a discrete copula directly at the pmf level. Let \(Q\) be a \(d\)-way probability array on a finite grid. The discrete copula \(C\) is the KL projection of \(Q\) onto the set \(\mathcal U\) of arrays with uniform margins: \[ C=\operatorname*{arg\,min}_{P\in\mathcal U} \sum_{i_1,\dots,i_d} P_{i_1,\dots,i_d} \log\!\Big(\frac{P_{i_1,\dots,i_d}}{Q_{i_1,\dots,i_d}}\Big). \] The empirical discrete copula is obtained by applying the same map to the empirical frequency array [2506.12316]. The Lagrangian first-order conditions imply a multiplicative scaling form, \[ P_{i_1,\dots,i_d} = Q_{i_1,\dots,i_d}\, \prod_{\ell=1}^d \beta^{(\ell)}_{i_\ell}, \] which makes the connection to iterative proportional fitting immediate. In the bivariate case this reduces to row and column scaling; in higher dimension it becomes multiway IPF or Sinkhorn scaling [2506.12316]. On rectangular supports, this yields a copula-like decomposition of a pmf \(p\) into its margins and a uniform-margin array \(u=U(p)\), with \[ p = I_{p^{[1]},p^{[2]}}(u), \] provided the support condition ensuring existence with the same support is satisfied. In that setting, \(p\) and \(u\) are diagonally equivalent and share the same odds-ratio matrix [2307.04225]. This support-aware view is closely related to the “nucleus” perspective for discrete dependence. Positive row and column scalings preserve odds ratios, so the canonical uniform-margin representative can be interpreted as the margin-free dependence core of a contingency table. This suggests that, for finite discrete supports, the most natural analogue of a copula is not an arbitrary extension of a subcopula to \([0,1]^2\), but a uniform-margin pmf preserving the relevant odds-ratio structure [1901.08741]. Structural zeros are part of this formulation rather than nuisances. Existence and uniqueness depend on the support, and sampling zeros can be stabilized by a smoothed empirical array \[ \widehat{\pp}_n=\frac{n}{n+1}\pp_n+\frac{1}{n+1}\qq, \] where \(\qq\) places equal mass on the non-zero support. This guarantees existence of the empirical projection while respecting the support geometry [2506.12316]. ## 5. Asymptotics, estimation, and testing The projection-based empirical discrete copula admits a full large-sample theory on finite supports. Under iid sampling, \[ \sqrt{n}\,(\widehat{\bm \gamma}_n-\bm \gamma_{\pp}) \Rightarrow \mathcal N(0,\Sigma_\gamma), \qquad \Sigma_\gamma=J_\gamma^*(\bm p)\,V\,J_\gamma^*(\bm p)^\top, \] with \(V=\mathrm{Diag}(\bm p)-\bm p\bm p^\top\). The covariance has an explicit sandwich form determined by the derivative of the projection map and the multinomial covariance of the empirical frequency array [2506.12316]. One direct consequence is asymptotic normality for Yule’s concordance coefficient, the discrete analogue of Spearman’s rho in that framework: \[ \sqrt{n}\,(\widehat\Upsilon_n-\Upsilon(\gamma)) \Rightarrow \mathcal N(0,\sigma_\Upsilon^2), \qquad \sigma_\Upsilon^2=\kappa_{r_1,r_2}^2\,\bm v^\top\Sigma_\gamma\bm v. \] The same derivative theory yields Wald tests for quasi-independence in multivariate contingency tables [2506.12316]. For rectangular supports, the differentiability of the \(I\)-projection also underpins nonparametric and parametric estimation of discrete copulas, including method-of-moments inversion based on \(\Upsilon\), Goodman–Kruskal’s gamma, and Kendall’s tau \(b\), as well as maximum pseudo-likelihood based on a parametric discrete copula family \(u^{[\theta]}\). Goodness-of-fit can then be formulated through chi-square-type statistics comparing the nonparametric \(u^{[N]}\) with the parametric \(u^{[\hat\theta^{[N]}]}\), with asymptotic null distributions computed from eigenvalues or approximated by semi-parametric bootstrap [2307.04225]. A recurring caution in the broader discrete-copula literature is that computational surrogates for discrete likelihoods are not automatically valid substitutes for exact support-aware inference. In multivariate normal copula regression with binary, Poisson, or negative binomial margins, continuous-extension or jittering estimators were shown to bias latent correlations downward and to understate standard errors, whereas maximum simulated likelihood based on rectangle probabilities remained nearly as efficient as exact maximum likelihood [1304.0905]. ## 6. Computation, scalability, and extensions Computation depends strongly on the chosen empirical object. For rank-based empirical copulas, one concern is memory rather than likelihood evaluation. In the bivariate streaming setting, a Greenwald–Khanna summary can be adapted to maintain a compact approximation \(\hat C_S(u_1,u_2)\) to the empirical copula with the uniform error guarantee \[ \bigl|\hat C_S(u_1,u_2)-\hat C(u_1,u_2)\bigr|\le 5\epsilon, \] using worst-case space \[ O\!\left(\frac{1}{\epsilon^2}\log^2(\epsilon n)\right). \] This makes online approximation feasible and also supplies bivariate building blocks for higher-dimensional pair-copula decompositions [1805.05168]. For high-dimensional discrete or mixed margins, direct likelihoods based on \(2^T\) rectangle evaluations are often intractable. One response is to bypass rectangle probabilities by augmenting with latent uniforms and approximating the resulting posterior variationally. In drawable vine copulas for ordinal and mixed time series, the variational Bayes estimator operates on the augmented posterior \[ p(\theta,u\mid y)\propto c(u\mid \theta)\,p(\theta)\prod_t I(a_t\le u_t