On Divergence-based Distance Functions for Multiply-connected Domains

TL;DR

This paper introduces a divergence distance function based on $f$-divergence for multiply-connected domains, which aligns with the Green's function gradient and can be used for efficient robotic path planning.

Contribution

It establishes a divergence distance based on $f$-divergence that shares key properties with the Green's function, enabling gradient-based path planning in complex domains.

Findings

The divergence distance's gradient is opposite to that of the Green's function.

The divergence distance has a single extremum at the target point.

It offers computational advantages for path planning.

Abstract

Given a finitely-connected bounded planar domain $\Omega$, it is possible to define a {\it divergence distance} $D(x,y)$ from $x\in\Omega$ to $y\in\Omega$, which takes into account the complex geometry of the domain. This distance function is based on the concept of $f$-divergence, a distance measure traditionally used to measure the difference between two probability distributions. The relevant probability distributions in our case are the Poisson kernels of the domain at $x$ and at $y$. We prove that for the $\chi^2$-divergence distance, the gradient by $x$ of $D$ is opposite in direction to the gradient by $x$ of $G(x,y)$, the Green's function with pole $y$. Since $G$ is harmonic, this implies that $D$, like $G$, has a single extremum in $\Omega$, namely at $y$ where $D$ vanishes. Thus $D$ can be used to trace a gradient-descent path within~$\Omega$ from $x$ to $y$ by following $\nabla_x D(x,y)$, which has significant computational advantages over tracing the gradient of $G$. This result can be used for robotic path-planning in complex geometric environments.