Differentiable D-vine Copulas for Localized Anomaly Detection

A new estimation framework combines gradient-based maximum likelihood estimation with beam search to optimize D-vine copula configurations. This approach avoids the sub-optimal local minima common in sequential greedy selection during multivariate distribution modeling.