Optimal Transport Bounds for Latent Density Smoothing on MTW($K>0$) Manifolds

Abstract

Smoothing an empirical distribution can fill gaps between observations, but noise added in the surrounding space also moves probability mass away from the manifold supporting the data. We study this tradeoff through Wasserstein error bounds for intrinsic heat smoothing, ambient Gaussian smoothing, and smoothing in an encoder--decoder's latent space. Under the Ma--Trudinger--Wang condition with positive cross-curvature and the stated transport-regularity assumptions, we prove that sufficiently small intrinsic smoothing improves upon the unsmoothed empirical measure. Our ambient bound quantifies how increasing the surrounding dimension restricts the smoothing scale suggested by the bound. For latent smoothing, the analysis shows how the benefit depends on reconstruction accuracy and on how the encoder--decoder pair preserves movement along the manifold and responds to noise in other directions. A quadratic approximation to the bound yields bandwidth-selection rules and conditions favoring latent smoothing in this comparison. Controlled synthetic experiments examine the predicted bandwidth and geometric trends. An exploratory MNIST study illustrates latent smoothing relative to pixel-space perturbations.