Work

Limit laws for shallow loops in planar Gaussian kernel density estimates

Year: 2026
Type: preprint
Abstract: A shallow hole in a smoothed density may disappear when the density is estimated from a finite sample. I study this transition through the longest finite bar in the degree-one persistent homology of the closed superlevel sets of a planar Gaussian kernel density estimator with fixed bandwidth. For a Gaussian-difference density family and a circular family with Gaussian observation error, local parameter displacements of order n^(−1/3) produce bar lengths of order n^(−2/3). The rescaled longest bar converges to an explicit law determined by a quartic potential with a Gaussian linear tilt. For a fixed positive local parameter s, the law has an atom at zero, and the probability of a positive bar converges to 1 − exp(−κ_occ s³), with an explicit model-dependent constant κ_occ > 0; this probability tends to zero for s ≤ 0. The proof identifies the saddle–minimum persistence pairing and reduces the full-plane persistence module to a fixed disk with probability tending to one. Within the circular family, central symmetry and agreement with the uniform circle's polynomial moments through degree four yield a common first-order law, covering both discrete and continuous angular arrangements. Numerical studies examine finite-sample curvature effects and compare uniform-circle, regular-hexagon, and regular-octagon samples. MSC Classification: 62G07 , 62G20 , 55N31 , 60F05
Author: Jiaxuan Ren
Language: English
Open Access status: green
OpenAlex ID: W7226025009
Cites: 11
Cited by:
Related to: 10