Asymptotic confidence regions for density ridges (Q2040044)

Ridges of a given distribution on \(\mathbb{R}^d\) are a generalization of a mode: lower dimensional geometric features defined as maxima of the underlying density constrained to be within a given subspace. The present paper details the construction of data-driven asymptotic confidence regions for ridges, under mild assumptions (for example, to ensure differentiability and avoid degeneracy). Beginning with ridge estimates based on smoothed kernel density functions, the author uses Gaussian approximation for suprema of empirical processes and extreme value theory for suprema of \(\chi\)-fields indexed by manifolds to construct confidence regions. Bias correction methods are then used to extend these results to the ridge of primary interest.