A Lusin theorem for nonlocal gradients

What the math gives to ML

The paper defines a fractional analogue of the gradient in which each location compares its value with all other locations through a distance-singular, direction-aware kernel. This offers a principled long-range interaction operator for spatial feature maps, with the fractional order controlling the balance between local and distant differences. A practical transfer is a residual nonlocal-gradient block or regularizer for vision models, implemented by truncating the integral to a multiscale offset stencil. The extracted material does not include the quantitative Lusin theorem, so the strongest directly supported transfer is the constructive fractional operator itself.