Computational Aspects of the Colorful Carath\'eodory Theorem

TL;DR

This paper investigates the computational complexity of the colorful Carathéodory problem, introduces approximation algorithms, and explores related problems like the nearest colorful polytope, revealing complexity classifications.

Contribution

It presents a polynomial-time approximation algorithm for the colorful Carathéodory problem and analyzes the complexity of related optimization problems, including local and global optima.

Findings

Polynomial-time algorithm for approximate colorful choices

NP-hardness of computing global optima for nearest colorful polytope

PLS-completeness of finding local optima in NCP

Abstract

Let $C_1,\dots,C_{d+1}\subset \mathbb{R}^d$ be $d+1$ point sets, each containing the origin in its convex hull. We call these sets color classes, and we call a sequence $p_1, \dots, p_{d+1}$ with $p_i \in C_i$, for $i = 1, \dots, d+1$, a colorful choice. The colorful Carath\'eodory theorem guarantees the existence of a colorful choice that also contains the origin in its convex hull. The computational complexity of finding such a colorful choice (CCP) is unknown. This is particularly interesting in the light of polynomial-time reductions from several related problems, such as computing centerpoints, to CCP. We define a novel notion of approximation that is compatible with the polynomial-time reductions to CCP: a sequence that contains at most $k$ points from each color class is called a $k$-colorful choice. We present an algorithm that for any fixed $\varepsilon > 0$, outputs an $\lceil \epsilon d\rceil$-colorful choice containing the origin in its convex hull in polynomial time. Furthermore, we consider a related problem of CCP: in the nearest colorful polytope problem (NCP), we are given sets $C_1,\dots,C_n\subset\mathbb{R}^d$ that do not necessarily contain the origin in their convex hulls. The goal is to find a colorful choice whose convex hull minimizes the distance to the origin. We show that computing a local optimum for NCP is PLS-complete, while computing a global optimum is NP-hard.