Optimal Projections in the Distance-Based Statistical Methods

TL;DR

This paper proposes an optimized projection method for distance-based statistics in multivariate data, reducing computational complexity and improving accuracy over random projection approaches.

Contribution

It introduces a novel approach to pre-calculate optimal projection directions, enabling faster computation and better approximation of distances in multivariate statistical tests.

Findings

Significant computational speed-up from $O(m^2)$ to $O(n m \, \log m)$.

Exact solutions for optimal projections in 2D and specific cases.

Simulation results show improved accuracy over random projections.

Abstract

This paper introduces a new way to calculate distance-based statistics, particularly when the data are multivariate. The main idea is to pre-calculate the optimal projection directions given the variable dimension, and to project multidimensional variables onto these pre-specified projection directions; by subsequently utilizing the fast algorithm that is developed in Huo and Sz\'ekely [2016] for the univariate variables, the computational complexity can be improved from $O(m^2)$ to $O(n m \cdot \mbox{log}(m))$, where $n$ is the number of projection directions and $m$ is the sample size. When $n \ll m/\log(m)$, computational savings can be achieved. The key challenge is how to find the optimal pre-specified projection directions. This can be obtained by minimizing the worse-case difference between the true distance and the approximated distance, which can be formulated as a nonconvex optimization problem in a general setting. In this paper, we show that the exact solution of the nonconvex optimization problem can be derived in two special cases: the dimension of the data is equal to either $2$ or the number of projection directions. In the generic settings, we propose an algorithm to find some approximate solutions. Simulations confirm the advantage of our method, in comparison with the pure Monte Carlo approach, in which the directions are randomly selected rather than pre-calculated.