Into the Danger Zone: Stable Extrapolation in High-Dimensional Function and Operator Learning

License: See LICENSE file · Updated 2026-09-29
My rating

This research paper investigates when stable extrapolation remains possible when test distributions differ from training distributions in scientific machine learning. Its focus is high-dimensional function and operator learning, with particular attention to assumptions on target functions or operators and the extent of extrapolation beyond the training domain. The work also examines existing theoretical error controls based on additive penalties that measure training-test distribution discrepancies.

Tags: