In statistics and information theory, divergences like Kullback-Leibler (KL), chi-squared, and squared Hellinger distance quantify how far apart two probability distributions are. It is often useful to bound one divergence by another, but such bounds typically degrade as the ambient dimension grows. This paper focuses on Gaussian location mixtures, where a mixing distribution over locations is convolved with a standard Gaussian, and asks whether dimension-free comparisons are possible.
The authors prove that for compactly supported mixing distributions, both the KL divergence and the chi-squared divergence are bounded above by a constant times the squared Hellinger distance. The constant depends only on the radius of the support, not on the ambient dimension or the number of mixture components. This strengthens earlier comparison inequalities and settles an open question posed by Jia et al. in 2023.
The proof relies on a Gaussian interpolation argument that connects the derivatives of these divergences to differences between posterior means. This technique may be of independent interest for other high-dimensional problems. The result is particularly relevant for high-dimensional statistics and machine learning, where dimension-free guarantees are highly desirable.