[Paper Review] An essay on copula modelling for discrete random vectors; or how to pour new wine into old bottles

Gery Geenens|arXiv (Cornell University)|Jan 25, 2019

Financial Risk and Volatility ModelingEconomics, Econometrics and Finance39 references3 citations

TL;DR

This paper proposes a reimagined copula framework for discrete random vectors by reviving century-old ideas from Udny Yule, replacing the classical copula definition based on probability integral transforms with a 'nucleus' concept that captures dependence structure independently of margins. The method enables unique, margin-free dependence modeling in discrete data, restoring the core benefits of copulas—flexibility and identifiability—while resolving the long-standing identifiability problem of classical copulas in the discrete case.

ABSTRACT

Copulas have now become ubiquitous statistical tools for describing, analysing and modelling dependence between random variables. Sklar's theorem, "the fundamental theorem of copulas", makes a clear distinction between the continuous case and the discrete case, though. In particular, the copula of a discrete random vector is not identifiable, which causes serious inconsistencies. In spite of this, downplaying statements are widespread in the related literature, and copula methods are used for modelling dependence between discrete variables. This paper calls to reconsidering the soundness of copula modelling for discrete data. It suggests a more fundamental construction which allows copula ideas to smoothly carry over to the discrete case. Actually it is an attempt at rejuvenating some century-old ideas of Udny Yule, who mentioned a similar construction a long time before copulas got in fashion.

Motivation & Objective

• To address the fundamental flaw in classical copula theory when applied to discrete random vectors: the non-identifiability of the copula due to discontinuous CDFs.
• To revive and formalize historical ideas from Udny Yule (1912) and others, who proposed standardizing contingency tables to uniform margins to isolate dependence.
• To construct a new class of discrete copulas—'nuclei'—that are uniquely defined and independent of marginal distributions, restoring the core promise of copulas: separating dependence from margins.
• To demonstrate that classical continuous copulas can still describe dependence in discrete distributions (e.g., Poisson, geometric) when properly reinterpreted via the nucleus framework.
• To introduce new discrete copula families (e.g., Bernoulli, Binomial, Geometric, Poisson copulas) that are consistent with the new framework and maintain all desirable properties of classical copulas.

Proposed method

• Define a 'nucleus' as an equivalence class of bivariate discrete distributions sharing the same dependence structure, independent of marginal distributions.
• Replace the classical copula definition based on the probability integral transform with a construction that standardizes margins to uniform distributions via frequency scaling, as proposed by Yule (1912).
• Use iterative proportional fitting (IPF) to couple arbitrary discrete marginals with a given dependence structure, ensuring the resulting joint distribution respects the target copula structure.
• Introduce the concept of a 'copula pmf'—the discrete analog of the copula density—that uniquely characterizes the dependence structure in the nucleus framework.
• Generalize the framework from Bernoulli to finite-support and countably infinite supports (e.g., ℕ×ℕ), showing that classical copulas can still describe dependence in such settings when interpreted via the nucleus.
• Construct new copula families such as the Geometric copula (for Marshall-Olkin bivariate geometric) and Poisson copula (for bivariate Poisson), derived from the nucleus-based approach.

Experimental results

Research questions

• RQ1 Can a discrete copula framework be constructed that restores the identifiability and margin-freeness of classical copulas in the discrete case?
• RQ2 How can historical ideas from Yule (1912) and Mosteller (1968) be formalized into a modern statistical framework for discrete dependence modeling?
• RQ3 Is it possible to define a unique 'copula pmf' for discrete distributions that captures dependence independently of the marginal distributions?
• RQ4 Can classical continuous copulas (e.g., Clayton, Gaussian) be meaningfully applied to discrete data when interpreted through the nucleus framework?
• RQ5 What new discrete copula families (e.g., Bernoulli, Poisson, Geometric) emerge from this framework, and how do they relate to known multivariate discrete distributions?

Key findings

• The classical copula definition fails for discrete random vectors because the probability integral transform does not yield uniform marginals, leading to non-identifiable copulas.
• By defining a 'nucleus' as an equivalence class of distributions with the same dependence structure, the proposed framework restores identifiability and margin-freeness in discrete settings.
• The method allows coupling any discrete marginals (e.g., Poisson(2)) with any classical copula (e.g., Clayton with θ = -0.2, Gaussian with ρ = -0.8) via IPF, producing valid bivariate distributions with the desired dependence.
• The framework supports the existence and uniqueness of a 'copula pmf' for discrete distributions, analogous to the copula density in the continuous case.
• New discrete copula families are introduced—such as the Geometric copula and Poisson copula—that describe the dependence in well-known discrete multivariate distributions.
• The approach reconciles classical copula theory with discrete data by showing that while Sklar’s theorem cannot be applied directly, the core idea of dependence separation remains valid through the nucleus concept.