A Normalizing Flow is a powerful density estimator based on neural networks. Essentially, it is simply a chain of variable transforming the density to learn $p(x)$ into any target distribution (e.g. uniform or standard Gaussian)
In the paper "Neural Autoregressive Flows" by Huang et. al. the authors prove that their proposed flow is a universal density estimator in the sense that any density supported on an open set $U$ can be arbitrarily learned using their flow.
As the target distribution of the normalizing flow, they use a uniform distribution in $(0,1)^m$ where $m$ is the dimensionality of the data. However, $(0,1)^m$ is diffeomorphic to $\mathbb{R}^m$ such that the results still holds if one uses a standard Gaussian target distribution instead of uniform. But this is a contradiction, as not every open set is diffeomorphic to $\mathbb{R}^m$, see e.g. discussion.
So how is it possible that at the one hand not every open set $U$ is diffeomorphic to $\mathbb{R}^m$, but at the other hand a density supported on any open set $U$ can be learned using a normalizing flow?