A spectral-compensated scheme for space-parameter Poisson noise functionals: error bounds and complexity estimates

What the math gives to ML

The paper gives a constructive approximation for infinite-activity positive-jump noise by removing jumps below an amplitude cutoff and replacing their aggregate with a variance-matched Gaussian term. Its transferable asset is the explicit Wasserstein error improvement from O(\varepsilon^{1-\alpha/2}) to O(\varepsilon), together with the resulting reduction in jump evaluations from O(\tau^{-2\alpha/(2-\alpha)}) to O(\tau^{-\alpha}) for stable-type intensities. A promising neural-network use is a Levy-noise module for stochastic neural ODEs, diffusion samplers, or heavy-tailed optimizer perturbations, where exact simulation of infinitely many small jumps is otherwise expensive. The compensation should be tested as a drop-in simulator acceleration while preserving the large-jump law and matching the small-jump drift and variance.