Approximating something with “first two moments”? Neural Network Paper

Question

I am reading a paper about Unsupervised Part-Based Disentangling of Object Shape and Appearance [https://arxiv.org/pdf/1903.06946.pdf] and at some point, they a