Copula Transformations for Data-Consistent Inversion
arXiv:2609.02832
2026
Architecture
1 ideas extracted · analyzed Sep 3, 2026
What the math gives to ML
The paper identifies a precise failure mode of sequential distribution matching: satisfying each marginal push-forward constraint does not generally recover the desired joint law, because the observed and predicted copulas may differ. This suggests a reusable dependence-correction layer for neural networks that separately calibrates one-dimensional marginals and then transports the predicted copula into the observed copula. The most practical first target is multi-output probabilistic prediction or inverse-problem amortizers, where marginal calibration is possible but correlations and higher-order dependence are wrong.
Ideas from this paper
Unverified
2026
Add a post-processing or differentiable output layer that first matches each predicted marginal distribution to the data marginal and then corrects the remaining joint dependence by transporting the predicted copula to the observed copula. This directly targets the discrepancy that sequential marginal or quantity-of-interest constraints cannot remove.
Useful6/10
Difficulty6/10
Novelty6/10