I wonder how good compression rations this really has... It reminds me of experiments I did with wavelets for image compression some.. 22 years ago. The inver

I wonder how good compression rations this really has... It reminds me of experiments I did with wavelets for image compression some.. 22 years ago. The inverse transform starts with a small pixel image and then uses just as many coefficients to transform it to an image twice as large (in width or height). This is done over and over again. The point is that most of your data consists of these coefficients, and most of them are close to zero: so you can flush them to zero. Then the problem is mostly a matter of how to encode where you don't have zeroes: i.e. you have like a bitmap and an array of the non-zeroes. There were different algorithms that encoded non-zeroes, more or less conservatively, but they did often exploit the property that these non-zeroes typically were quite clustered — which is the opposite of a typical hash function used for a bloom filter. Image compression this way had obviously very poor locality, first just for the transform and then for the coefficient compression, so it was quite slow. Therefore it did feel like a dead end.