Minimax rates clarify error in conditional generative learning models
Minimax Rates for Conditional Generative Learning under Wasserstein Loss
Machine Learning
Summary
This paper studies how accurately computers can learn to generate new data based on given conditions, measuring errors with a mathematical tool called Wasserstein loss. The authors figure out the best possible error rates (called minimax rates) when the underlying data changes smoothly and the data conditions aren't restricted. They also address how the error changes when focusing on the worst-case parts of the data and how this affects learning. Finally, they provide both theoretical limits and matching algorithms that achieve these error rates under various technical assumptions.
What this means in practice
- •For machine learning engineers: Set realistic expectations for accuracy and sample size when building conditional generative models with Wasserstein-based error metrics.
- •For data scientists: Use derived minimax rates to assess risks in conditional density estimation when data coverage is uneven or has irregular smoothness.
A theory result. No direct application yet.
Authors
Sungwon Park, Kyoung-Sook Moon, Hongjoong Kim
Abstract
We study minimax rates for conditional generative learning under integrated Wasserstein loss, with Hölder continuous conditional laws and an unrestricted covariate distribution. We evaluate unrestricted generators through their conditional output distributions. For expected Wasserstein loss, the minimax rate agrees with the rate under uniform covariates and follows from balancing localization bias against unconditional distribution estimation error. Higher moments of the loss can instead be dominated by unobserved covariate regions, which force an additional risk contribution of order $n^{-1}$, where $n$ is the sample size. An adaptive localization argument gives matching upper bounds. The results cover unrestricted response laws and possibly vanishing Besov densities for $W_1$, and extend to higher Wasserstein orders under uniform response density bounds and integrability conditions. A multiscale lower bound closes the logarithmic gap in unconditional $W_1$ minimax rates in dimension two for both unrestricted distributions and Besov density classes. We also establish minimax rates for conditional Brenier maps under density, integrability, and uniform curvature conditions.