The Setting
For , clipping a stable random gradient allows its coordinates to be bounded while controlling the error in their th moments with prescribed edge weights. The dimension question concerns realizations of finite subsets of real spaces in real coordinate spaces.
For , the coordinate distance on is given by (1): (1)✦ An admissible dimension for points and distortion is a nonnegative integer such that every injective family in every real admits and a scale satisfying (2): (2)✦ The universal dimension for points and distortion is the infimum in the nonnegative integers of the admissible dimensions for points and distortion .
Normalized Gradients and Clipping
For the gradient estimates, let be a finite vertex set with , let be a nonempty finite edge set, and let be endpoint maps. Assume that every vertex is incident to an edge and every pair of distinct vertices is joined by an edge in one of the two orientations. Fix positive edge scales and positive weights with .
The incidence matrix and normalized gradient matrix are defined in (3); the normalized-gradient space is the image of : (3)
For a scalar probability measure on , write for its characteristic function. For a scalar probability measure on and , the capped th moment of is the integral in (4): (4)
Definition 1. For real , the clipping function is defined by (5): (5)✦ For real and , the coordinatewise clipping of is the vector .
The Claim
Theorem 1. Let be a probability measure on . Suppose , and , and assume that for every . The logarithmic moment lower bound is (6): (6)✦
Suppose now that and , with for every real . Let be a probability space and a measurable map whose every coordinate has law . For every , there is a measurable map satisfying the retraction moment bound (7): (7)✦
Assume instead that for every real . Let be a finite index set and let , for , satisfy for every . For each , there is a measurable map satisfying the weighted stable gradient bound (8), where is the product probability measure: (8)✦
For every and , there is a real constant such that the universal dimension satisfies (9): (9)✦
In the paper, the result reads: for an error parameter and constants , and , with , let be the controlled retraction and let be the paper’s weights; for let , take and ; then , , and , while choosing gives and -weighted error at most , with and .
What Is Checked
The moment statements concern scalar probability laws and finite normalized-gradient spaces with positive edge scales and positive weights of total mass one, while the universal dimension statement concerns every injective finite family in every real space.