Flexible Multivariate Density Estimation with Marginal Adaptation

Our article addresses the problem of flexibly estimating a multivariate density while also attempting to estimate its marginals correctly. We do so by proposing two new estimators that try to capture the best features of mixture of normals and copula estimators while avoiding some of their weaknesses. The first estimator we propose is a mixture of normals copula model that is a flexible alternative to parametric copula models such as the normal and t copula. The second is a marginally adapted mixture of normals estimator that improves on the standard mixture of normals by using information contained in univariate estimates of the marginal densities. We show empirically that copula based approaches can behave much better or much worse than estimators based on mixture of normals depending on the properties of the data. We provide fast and reliable implementations of the estimators and illustrate the methodology on simulated and real data.

💡 Research Summary

The paper tackles the longstanding challenge of estimating a multivariate probability density while simultaneously recovering accurate marginal distributions for each variable. Traditional approaches fall into two camps. Parametric copula models (e.g., Gaussian, t, Clayton, Gumbel) explicitly model dependence but force the marginals into a limited set of predefined families, leading to severe bias when the true marginals are skewed, multimodal, or otherwise non‑Gaussian. On the other hand, mixture‑of‑normals (MoN) models are highly flexible in capturing the overall shape of a multivariate density, yet they do not guarantee that the implied one‑dimensional margins match the data; this mismatch can be especially problematic when the marginal shapes deviate strongly from normality.

To bridge this gap, the authors propose two novel estimators that blend the strengths of copula and mixture approaches while mitigating their weaknesses. The first estimator, called the Mixture of Normals Copula (MNC), proceeds in two stages. First, each variable’s marginal density is estimated non‑parametrically (e.g., kernel density estimation or local polynomial regression), yielding (\hat g_j) for the (j)‑th coordinate. These estimated marginals are then transformed to “pseudo‑normal’’ margins by standardizing with respect to the corresponding normal distribution. On the transformed data, a finite mixture of multivariate normal components is fitted, but the mixture is interpreted as a copula: the component weights, means, and covariances describe the dependence structure while the previously estimated marginals guarantee that the final joint density respects the observed univariate shapes. By allowing an unrestricted number of mixture components and flexible covariance structures, MNC can capture highly non‑linear, asymmetric dependence that ordinary Gaussian or t‑copulas cannot.

The second estimator, Marginally Adapted Mixture of Normals (MAMN), augments the standard MoN likelihood with a correction term that explicitly incorporates the univariate marginal estimates. The modified log‑likelihood is

\