[Paper Review] A note on dichotomies for metric transforms

Manor Mendel, Assaf Naor|arXiv (Cornell University)|Feb 9, 2011

Advanced Harmonic Analysis Research17 references3 citations

TL;DR

This paper establishes a dichotomy for metric transforms induced by nondecreasing concave functions $ω$ with $\omega(0)=0$: either all finite metric spaces embed with distortion arbitrarily close to 1 into $\omega$-transforms, or the path metric $P_n$ requires distortion at least $n^{\alpha}$ for some $\alpha>0$. The result resolves a key case in metric cotype theory and shows that $\theta$-snowflake transforms of $\mathbb{R}$ achieve distortion $n^{1-\theta}$ for $P_n$, highlighting that expanders are not always the worst-case spaces.

ABSTRACT

We show that for every nondecreasing concave function w:R+ --> R+ with w(0)=0, either every finite metric space embeds with distortion arbitrarily close to 1 into a metric space of the form (X,w o d) for some metric d on X, or there exists a=a(w)>0 and n_0=n_0(w)\in N such that for all n>n_0, any embedding of {0,...,n} into a metric space of the form (X,w o d) incurs distortion at least n^a.

Motivation & Objective

• To establish a dichotomy for metric transforms $\omega \circ d$ induced by concave, nondecreasing $\omega$ with $\omega(0)=0$.
• To resolve the metric cotype dichotomy problem for the class of $\omega$-metric transforms $\omega(\text{MET})$.
• To determine whether all finite metric spaces embed with distortion arbitrarily close to 1 into $\omega(\text{MET})$, or whether some spaces (like $P_n$) force polynomial distortion.
• To show that $\theta$-snowflake transforms of $\mathbb{R}$ yield distortion $n^{1-\theta}$ for the path $P_n$, and that this is tight.

Proposed method

• Analyzes the distortion of embeddings of the path space $P_n = \{0,\dots,n\} \subset \mathbb{R}$ into $\omega$-metric transforms $(X, \omega \circ d)$.
• Uses the fact that $\omega$-transforms preserve the triangle inequality due to concavity and monotonicity of $\omega$, ensuring $\omega \circ d$ is a metric.
• Applies a duality argument: if not all finite metric spaces embed with distortion near 1, then $P_n$ must force high distortion.
• Leverages the sharpness of the $\theta$-snowflake case via explicit constructions showing $c_{\omega(\text{MET})}(P_n) = n^{1-\theta}$ for $\omega(t) = t^\theta$, $\theta \in (0,1)$.
• Uses the fact that $P_n$ embeds isometrically into $\mathbb{R}$, and that $\omega$-transforms of $\mathbb{R}$ yield distortion $n^{1-\theta}$ for $P_n$, establishing a lower bound.
• Proves that if $\omega^{-1}(\lambda \omega(t))$ is subadditive for all $\lambda > 0$, then the distortion bound improves, and shows this condition implies $\omega(t) = a t^b$ for $b \in (0,1]$.

Experimental results

Research questions

• RQ1 Does every finite metric space embed with distortion arbitrarily close to 1 into $\omega(\text{MET})$ for a given concave $\omega$ with $\omega(0)=0$?
• RQ2 If not, what is the minimal polynomial distortion that must be incurred by some finite metric space in $\omega(\text{MET})$?
• RQ3 Is the $\theta$-snowflake transform of $\mathbb{R}$ a worst-case example for distortion in $\omega(\text{MET})$?
• RQ4 Can the distortion of $P_n$ in $\omega(\text{MET})$ be bounded by $n^{1-\theta}$ when $\omega(t) = t^\theta$?
• RQ5 Are expanders always the worst-case graphs for metric dichotomy problems, or can other spaces force higher distortion?

Key findings

• For any nondecreasing concave $\omega$ with $\omega(0)=0$, either all finite metric spaces embed into $\omega(\text{MET})$ with distortion arbitrarily close to 1, or $P_n$ requires distortion at least $n^{\alpha}$ for some $\alpha>0$.
• The distortion of $P_n$ in $\omega(\text{MET})$ is exactly $n^{1-\theta}$ when $\omega(t) = t^\theta$ for $\theta \in (0,1)$, proving sharpness of the bound.
• The class $\omega(\text{MET})$ is closed under dilation if and only if $\omega(t) = a t^b$ for $a>0$, $b \in (0,1]$, which characterizes the $\theta$-snowflake case.
• The distortion of any $n$-point subset of $\mathbb{R}$ into the $\theta$-snowflake of $\mathbb{R}$ is at most $(n-1)^{1-\theta}$, and this bound is tight for $P_n$.
• Expanders do not achieve the worst-case distortion in $\omega(\text{MET})$; their distortion in $\theta$-snowflakes grows as $(\log n)^{1-\theta}$, not $n^{1-\theta}$.
• If $\omega^{-1}(\lambda \omega(t))$ is subadditive for all $\lambda>0$, then the distortion bound improves, and such $\omega$ must be a power function $t^b$.