Compensated Convex Transforms and Geometric Singularity Extraction from Semiconvex Functions

TL;DR

This paper introduces compensated convex transforms to extract geometric singularities from semiconvex functions, providing asymptotic analysis and new landscape functions for understanding the structure of singular sets.

Contribution

It develops a novel framework using compensated convex transforms for analyzing singularities in semiconvex and DC-functions, including asymptotic expansions and limit behaviors.

Findings

Limit of scaled valley transform equals squared radius of minimal bounding sphere.

Gradient of upper transform converges to center of minimal bounding sphere.

Scale 1-edge transform provides lower bounds related to subdifferential radii.

Abstract

We apply upper and lower compensated convex transforms, which are `tight' one-sided approximations of a given function, to the extraction of fine geometric singularities from semiconvex/semiconcave functions and DC-functions in $\mathbb{R}^n$ (difference of convex functions). Well-known examples of (locally) semiconcave functions include the Euclidean distance and squared distance functions. For a locally semiconvex function $f$ with general modulus, we show that `locally' a point is a singular (non-differentiable) point if and only if it is a scale $1$-valley point, and if $x$ is a singular point, then locally the limit of the scaled valley transform exists at every point $x$ and $ \lim_{\lambda\to \infty}\lambda V_\lambda (f)(x)=r_x^2/4$, where $r_x$ is the radius of the minimal bounding sphere of the (Fr\'echet) subdifferential $\partial_- f(x)$ and $V_\lambda (f)(x)$ is the valley transform at $x$. Thus the limit function $\mathcal{V}_\infty(f)(x):=\lim_{\lambda\to+\infty}\lambda V_\lambda (f)(x)=r_x^2/4$ gives a `scale $1$-valley landscape function' of the singular set for a locally semiconvex function $f$, and also provides an asymptotic expansion of the upper transform $C^u_\lambda(f)(x)$ when $\lambda \to \infty$. For a locally semiconvex function $f$ with linear modulus we show that the limit of the gradient of the upper compensated convex transform $\lim_{\lambda\to+\infty}\nabla C^u_\lambda(f)(x)$ exists and equals the centre of the minimal bounding sphere of $\partial_- f(x$, and that for a DC-function $f=g-h$, the scale $1$-edge transform satisfies $\liminf_{\lambda\to+\infty}\lambda E_\lambda (f)(x)\geq (r_{g,x}-r_{h,x})^2/4$, where $r_{g,x}$ and $r_{h,x}$ are the radii of the minimal bounding spheres of the subdifferentials $\partial_- g$ and $\partial_- h$ of the convex functions $g$ and $h$ at $x$ respectively.