1: \begin{abstract}

2: 		Coupling-based normalizing flows (e.g.~RealNVP) are a popular family of normalizing flow architectures that work surprisingly well in practice.

3: 		This calls for theoretical understanding. Existing work shows that such flows \textit{weakly} converge to arbitrary data distributions \cite{teshima_coupling-based_2020}.

4: 		However, they make no statement about the stricter convergence criterion used in practice, the maximum likelihood loss.

5: 		For the first time, we make a quantitative statement about this kind of convergence:

6: 		We prove that all coupling-based normalizing flows perform whitening of the data distribution (i.e.~diagonalize the covariance matrix) and derive corresponding convergence bounds that show a linear convergence rate in the depth of the flow.

7: 		Numerical experiments demonstrate the implications of our theory and point at open questions.

8: 	\end{abstract}

9: