Basically, if you have a multi-scale problem and you think the coordinates for the scaling are embedding in the data itself ... what do you do? Is the idea that with these normalizing flows you always take some hierarchical structure and start having a go at limiting scale? So you could start with ZERO dims (just learn shift and scale constants), then pick your least variable feature and try to descale the other based on that? Or do you just randomly reorder and pick the first n.
Am looking for a coherent reason to pick anything in the absence of concrete model knowledge.