20-60x the bandwidth requirements is one of the things that can reasonably be expected to improve with effort. Look at 2D video and how much better h264 is than

20-60x the bandwidth requirements is one of the things that can reasonably be expected to improve with effort. Look at 2D video and how much better h264 is than, say, GIF. And the higher the dimensionality, the more data but also the more low-hanging fruit for compression. We had 100:1 lossy lightfield compression 20 years ago; we can likely put another zero on that, maybe more with machine learning. That's what the linked paper is about - effective representations for lightfield data. The field is just kicking off.