NB: this problem actually happens inside tensorflow, and results in samples not exactly from the true pdf. The principle is, however, the same in numpy, and my aim is to understand the following warning. NB: this problem actually happens inside tensor