✨ TL;DR
This paper proposes Vine Denoising Copula (VDC), an amortized approach to vine copulas that trains a single bivariate denoising model and reuses it across all vine edges, enabling fast and tractable high-dimensional density and information estimation. The method combines neural flexibility with classical copula interpretability while achieving significant computational speedups.
Modeling high-dimensional dependencies with tractable likelihoods is challenging. Classical vine-copula methods are interpretable and maintain exact likelihoods but are computationally expensive due to per-edge optimization. Conversely, neural density estimators are flexible but lack the structured interpretability of copulas and can be difficult to apply for information-theoretic quantities like mutual information and total correlation.
VDC trains a single bivariate denoising model that is amortized and reused across all edges in a vine structure. For each edge, given pseudo-observations, the model predicts a density grid. An IPFP/Sinkhorn projection is then applied to enforce non-negativity, unit mass, and uniform marginals, ensuring the result is a valid copula density. This replaces expensive per-edge optimization with efficient GPU inference while preserving the exact vine likelihood and copula interpretation.
What the paper shows.
VDC achieves strong bivariate density accuracy and competitive mutual information and total correlation estimation across synthetic and real-data benchmarks. The method delivers substantial speedups for high-dimensional vine fitting, making explicit information estimation and dependence decomposition feasible at scales where repeated vine fitting would be costly. However, conditional downstream inference performance is mixed.
The paper acknowledges that conditional downstream inference remains mixed, suggesting the method may not be equally effective for all downstream tasks. The approach is limited to the vine-copula framework and may not capture all forms of high-dimensional dependencies. Computational gains are primarily in the fitting phase; the method's performance on tasks requiring conditional inference or complex downstream applications is not fully characterized.
✨ Generated by Claude · Apr 25, 2026 · Read the PDF for authoritative content.