On Measures of Dissimilarity on the Majorization Simplex

Abstract

The majorization simplex $\mathcal{P}_n$ is the set of probability vectors of length $n$ whose entries are arranged in non-increasing order. Under the majorization order, $\mathcal{P}_n$ is a lattice, and it is the natural space for comparing distributions when the labels of the outcomes are not relevant. In this work, we study measures of dissimilarity (and, similarity) on $\mathcal{P}_n$. Our first set of results concerns the Rényi-entropy based distance $d_α(p,q)=H_α(p)+H_α(q)-2H_α(p\vee q)$ proposed by Yadav and Shkel (2026), which generalizes a metric of Cicalese, Gargano and Vaccaro (2013). We show that $d_α$ is a metric on $\mathcal{P}_n$ for $α\in[1,\infty)$, while for $α\in[0,1)\cup\{\infty\}$, $d_α$ is not a metric in general. Our main result in this part identifies the necessary and sufficient conditions for equality in the triangle inequality of $d_α$ when $α>1$, for which we identify the equality conditions for the supermodularity of the Rényi entropy $H_α$. Our second set of results concerns the total variation distance and the Hellinger affinity (also called the Bhattacharyya coefficient) on $\mathcal{P}_n$. We obtain a number of bounds on these measures, along with some of their equality conditions. The resulting Hellinger affinity based upper bound on the total variation distance can be smaller than the bounds obtained from Pinsker's and the Bretagnolle--Huber inequalities.