Bridging Differential Privacy and Random Triangles

Abstract

The classical analysis of the Gaussian mechanism in differential privacy reduces privacy loss for a pair of neighboring datasets to a scalar random variable. While this scalar characterization is sufficient for privacy accounting, each perturbation instance also induces a high-dimensional random triangle formed by the sensitivity vector and the two corresponding noise vectors. In this work, we develop two complementary geometric representations of these random triangles. The first representation maps the normalized squared edge lengths to a simplex. We derive its exact joint density, characterize its elliptical support, and reconstruct the classical privacy loss random variable from the simplex coordinates. The second representation maps the spectral shape of each normalized triangle to a hemisphere. We derive the corresponding density and coordinate mappings, recover the same privacy loss, and characterize an equatorial drift together with band concentration as the dimension increases. These results provide two exact geometric coordinate systems that complement the scalar privacy loss and connect differential privacy with the probabilistic analysis of random shapes.