An improvement on the Rado bound for the centerline depth

TL;DR

This paper improves the lower bounds on the half-space depth for certain flats in high-dimensional spaces, surpassing the classical Rado bound for most cases.

Contribution

The authors provide new lower bounds for the half-space depth of k-flats in ^d, improving upon the Rado bound for all pairs (d,k) except the known optimal cases.

Findings

For k=1 and d
, a line with depth at least 1/d + 1/(3d^3) exists.

For all 0<k<d-1, there exists a k-flat with depth at least 1/(d+1-k) + 1/(3(d+1-k)^3).

The bounds are strictly better than the classical Rado bound in these cases.

Abstract

Let $\mu$ be a Borel probability measure in $\mathbb R^d$. For a $k$-flat $\alpha$ consider the value $\inf \mu(H)$, where $H$ runs through all half-spaces containing $\alpha$. This infimum is called the half-space depth of $\alpha$. Bukh, Matou\v{s}ek and Nivasch conjectured that for every $\mu$ and every $0 \leq k < d$ there exists a $k$-flat with the depth at least $\tfrac{k + 1}{k + d + 1}$. The Rado Centerpoint Theorem implies a lower bound of $\tfrac{1}{d + 1 - k}$ (the Rado bound), which is, in general, much weaker. Whenever the Rado bound coincides with the bound conjectured by Bukh, Matou\v{s}ek and Nivasch, i.e., for $k = 0$ and $k = d - 1$, it is known to be optimal. In this paper we show that for all other pairs $(d, k)$ one can improve on the Rado bound. If $k = 1$ and $d \geq 3$ we show that there is a 1-dimensional line with the depth at least $\tfrac{1}{d} + \tfrac{1}{3d^3}$. As a corollary, for all $(d, k)$ satisfying $0 < k < d - 1$ there exists a $k$-flat with depth at least $\tfrac{1}{d + 1 - k} + \tfrac{1}{3(d + 1 - k)^3}$.