6 problems

Let $\Omega$, $\rho_0$, and $\rho_1$ be as in the preceding convergence theorem, and let $F_{\theta_j}:\Omega\to\Omega$ be a sequence of invertible RealNVP neural networks whose pu…

A flow is a family of maps with group properties, and neural conjugate flows represent dynamics by composing an invertible map, a flow, and its inverse. Training may either impose…

Consider normalizing flows built from NSF layers, where the tails of each NSF layer are linear. Tail-linearity conjecture. The observed improved performance in low-sample tail regi…

Let a random variable have a bounded copula density, and let a Lipschitz-continuous diffeomorphism be applied to it. Bounded-copula-density conjecture. The transformed random varia…

Probabilistic-circuit integration conjecture. Although not tested in the experiments, integrating PCs with better Flow models, such as IDF++, could further improve performance.

Non-universality conjecture. Since the step functions used to establish $L^p$-universality bypass the uniformity of the transformation by affine coupling flows, the difficulty is i…