Optimal Stopping and Constraints for Diffusion Models of Signals with Discontinuities
Gaussian process regression models can be utilized in recovery of discontinuous signals. Their computational complexity is linear in the number of observations if applied with the covariance functions of nonlinear diffusion. However, such processes often result in hard-to-control jumps of the signal value. Synthetic examples presented in this work indicate that Bayesian evidence-maximizing stopping and knowledge whether signal values are discrete help to outperform the steady state solutions of nonlinear diffusion filtering.