EnergyDistanceMatrix

Introduction

The EnergyDistanceMatrix class implements the energy distance metric, a powerful statistical measure for comparing probability distributions. This metric belongs to the family of distance-based approaches in multivariate statistics and provides a robust method for measuring dissimilarity between distributions without requiring density estimation.

Description

The energy distance derives its name from Newton’s gravitational potential energy, drawing an analogy between statistical distances and physical potential energies. For two distributions P and Q, the energy distance quantifies their dissimilarity by considering all pairwise Euclidean distances between samples.

Mathematical Formulation

For two samples X and Y with sizes n and m respectively, the energy distance is defined as:

\[\mathcal{E}(X,Y) = \left(\frac{2}{nm}\sum_{i=1}^n\sum_{j=1}^m ||x_i - y_j|| - \frac{1}{n^2}\sum_{i=1}^n\sum_{j=1}^n ||x_i - x_j|| - \frac{1}{m^2}\sum_{i=1}^m\sum_{j=1}^m ||y_i - y_j||\right)^{\frac{1}{2}}\]

where: - ||·|| denotes the Euclidean norm - x_i and y_j are sample points from distributions X and Y respectively - n and m are the sample sizes for X and Y

Properties

The energy distance has several important properties:

  1. Non-negativity: E(X,Y) ≥ 0

  2. Identity of indiscernibles: E(X,Y) = 0 if and only if X and Y are identically distributed

  3. Symmetry: E(X,Y) = E(Y,X)

  4. Triangle inequality: E(X,Z) ≤ E(X,Y) + E(Y,Z)

Academic References

  1. Székely, G. J., & Rizzo, M. L. (2013). Energy statistics: A class of statistics based on distances. Journal of Statistical Planning and Inference, 143(8), 1249-1272.

  2. Szekely, G. J. (2003). E-statistics: The energy of statistical samples. Bowling Green State University, Department of Mathematics and Statistics Technical Report, 3(05), 1-18.

  3. Rizzo, M. L., & Székely, G. J. (2016). Energy distance. Wiley Interdisciplinary Reviews: Computational Statistics, 8(1), 27-38.

Conclusion

The EnergyDistanceMatrix class provides a robust implementation of the energy distance metric, making it particularly useful for: - Testing multivariate goodness-of-fit - Measuring statistical distance between samples - Detecting differences in distributions without assuming a parametric form - Cluster analysis and pattern recognition tasks

This implementation is optimized for efficiency while maintaining numerical stability, making it suitable for both small and large-scale statistical applications within the distancia package.