Fast Learning Rates for Physics-Informed Kernel Methods
This paper provides a theoretical analysis of how differential information can improve the learning rate of kernel methods in physics-informed machine learning, which is relevant for researchers working on model optimization.