Inapproximability for metric embeddings into R^d

TL;DR

This paper investigates the computational difficulty of embedding n-point metric spaces into low-dimensional Euclidean spaces with minimal distortion, establishing inapproximability bounds and exploring geometric topology and probabilistic methods.

Contribution

It introduces new inapproximability results for metric embeddings into R^d, using simple reductions and geometric topology, and compares these bounds with known embedding algorithms.

Findings

Inapproximability factor roughly n^(1/(22d-10)) for fixed d≥2

No polynomial-time algorithm can distinguish embeddable spaces with constant distortion from those requiring n^(c/d) distortion for d≥3

Constructs a metric space with high embedding distortion in R^2, yet with well-embeddable subspaces

Abstract

We consider the problem of computing the smallest possible distortion for embedding of a given n-point metric space into R^d, where d is fixed (and small). For d=1, it was known that approximating the minimum distortion with a factor better than roughly n^(1/12) is NP-hard. From this result we derive inapproximability with factor roughly n^(1/(22d-10)) for every fixed d\ge 2, by a conceptually very simple reduction. However, the proof of correctness involves a nontrivial result in geometric topology (whose current proof is based on ideas due to Jussi Vaisala). For d\ge 3, we obtain a stronger inapproximability result by a different reduction: assuming P \ne NP, no polynomial-time algorithm can distinguish between spaces embeddable in R^d with constant distortion from spaces requiring distortion at least n^(c/d), for a constant c>0. The exponent c/d has the correct order of magnitude, since every n-point metric space can be embedded in R^d with distortion O(n^{2/d}\log^{3/2}n) and such an embedding can be constructed in polynomial time by random projection. For d=2, we give an example of a metric space that requires a large distortion for embedding in R^2, while all not too large subspaces of it embed almost isometrically.