On Constant Factor Approximation for Earth Mover Distance over Doubling Metrics

Shi Li

DOI: https://doi.org/10.48550/arxiv.1002.4034

2010-01-01

Abstract:Given a metric space $(X,d_X)$, the earth mover distance between two distributions over $X$ is defined as the minimum cost of a bipartite matching between the two distributions. The doubling dimension of a metric $(X, d_X)$ is the smallest value $\alpha$ such that every ball in $X$ can be covered by $2^\alpha$ ball of half the radius. We study efficient algorithms for approximating earth mover distance over metrics with bounded doubling dimension. Given a metric $(X, d_X)$, with $|X| = n$, we can use $\tilde O(n^2)$ preprocessing time to create a data structure of size $\tilde O(n^{1 + \e})$, such that subsequently queried EMDs can be $O(\alpha_X/\e)$-approximated in $\tilde O(n)$ time. We also show a weaker form of sketching scheme, which we call "encoding scheme". Given $(X, d_X)$, by using $\tilde O(n^2)$ preprocessing time, every subsequent distribution $\mu$ over $X$ can be encoded into $F(\mu)$ in $\tilde O(n^{1 + \e})$ time. Given $F(\mu)$ and $F(\nu)$, the EMD between $\mu$ and $\nu$ can be $O(\alpha_X/\e)$-approximated in $\tilde O(n^\e)$ time.