WorksMathematics
Sharp Local Entropy Bounds in the Wasserstein Space
Abstract
Fréchet regression extends regression analysis to metric space-valued responses such as probability distributions. Its asymptotic theory often relies on local entropy conditions, which control the complexity of small balls in the response space. A commonly used condition requires the associated local entropy integral to remain bounded as the balls shrink, but this requirement can fail when the response space is infinite-dimensional. We determine the exact rate at which this integral diverges in the Wasserstein space of probability distributions supported on a compact interval. Specifically, for a broad class of ball centers, we establish matching upper and lower bounds for the local covering entropy and show that the local entropy integral grows at the rate of the reciprocal square root of the ball radius. This rate persists when the response space is restricted to a large class of continuous distributions. The same rate holds for uniform local entropy integrals, with the supremum taken over ball centers. These results quantify the local complexity of the Wasserstein space and provide sharp benchmarks for entropy assumptions in statistical problems with distribution-valued responses.