Measuring Multivariate Distribution Differences with Energy Distance

Formal Definition

Given two probability distributions F and G, draw independent random vectors X from F and Y from G. The Energy Distance D(F,G) is defined as D(F,G) = 2\,E\|X - Y\| - E\|X - X'\| - E\|Y - Y'\| Here E\|X - Y\| is the expected Euclidean distance between a point from each distribution (cross distance), while E\|X - X'\| and E\|Y - Y'\| are the expected distances between two points drawn from the same distribution (within‑distribution distances).

Principle of Energy Distance

The metric can be visualized as the net interaction energy of a system of charged particles: imagine one cloud of positively charged points and another of negatively charged points. Cross‑distribution pairs correspond to attractive interactions, and within‑distribution pairs correspond to repulsive self‑interactions. When the two clouds coincide, attractive and repulsive forces cancel, yielding zero Energy Distance; otherwise the net energy is positive.

Energy Distance measures the excess separation between two distributions beyond the natural separation within each distribution.

Illustrations with two‑dimensional distributions show that when the distributions are identical, Energy Distance equals zero; as they move apart, the cross‑distance dominates and the metric rises; when each distribution becomes more dispersed, within‑distribution distances increase and the metric trends back toward zero.

Permutation Test

To determine whether an observed Energy Distance reflects a statistically significant difference, a permutation test is used. The null hypothesis assumes the two samples come from the same distribution ( F = G). The combined sample is repeatedly shuffled, group labels are reassigned while preserving original sample sizes, and Energy Distance is recomputed each time to build an empirical null distribution. The p‑value is the proportion of permuted statistics that exceed the observed value.

Applying this test to a training‑set and test‑set revealed no evidence of a global covariate shift, though it does not rule out local extrapolation risks in sparse or tail regions of the feature space.

Conclusion

Energy Distance is a metric‑based statistical tool suitable for quantifying differences between two multivariate datasets. It is useful for data‑drift detection, verifying sample consistency in A/B tests, and comparing groups, whenever the question “do these two multivariate samples come from the same distribution?” arises.

Compared with univariate marginal tests, Energy Distance captures changes in joint relationships among variables, not just shifts in individual feature distributions. However, it detects only global distribution differences; its sensitivity to local, tail‑region discrepancies is limited, especially in high‑dimensional settings where Euclidean distances lose discriminative power. Combining Energy Distance with local density estimation or region‑wise tests can provide a more robust assessment.