Abstract
We derive a general change of variables formula for score functions, showingthat for a smooth, invertible transformation $\mathbf{y} = \phi(\mathbf{x})$,the transformed score function $\nabla_{\mathbf{y}} \log q(\mathbf{y})$ can beexpressed directly in terms of $\nabla_{\mathbf{x}} \log p(\mathbf{x})$. Usingthis result, we develop two applications: First, we establish a reverse-timeIt\^o lemma for score-based diffusion models, allowing the use of$\nabla_{\mathbf{x}} \log p_t(\mathbf{x})$ to reverse an SDE in the transformedspace without directly learning $\nabla_{\mathbf{y}} \log q_t(\mathbf{y})$.This approach enables training diffusion models in one space but sampling inanother, effectively decoupling the forward and reverse processes. Second, weintroduce generalized sliced score matching, extending traditional sliced scorematching from linear projections to arbitrary smooth transformations. Thisprovides greater flexibility in high-dimensional density estimation. Wedemonstrate these theoretical advances through applications to diffusion on theprobability simplex and empirically compare our generalized score matchingapproach against traditional sliced score matching methods.