flat blocking multi-resolution strategy - divide the volume into many sub-blocks with equal size
"octree subdivision" would be more traditional term rather then "blocking multi-resolution strategy". In fact, min|max octree has been used for decades to speedup volume ray-casting, there are tons of publications and patents on the subject...
as it says there are 2^(3 * max_l) voxels in the i-th sub-block of the certain
If ( the numeration of octree levels from root down to leaves goes as: 0 -> max_l )
then
2^(3 * max_l) is the number of all voxels in the all volume.
and
2^(3 * (max_l - l)) is the number of voxels in the subdivided volume (sub-volume) of level "l"
is the mathematic foundation of block importance
Well, if you define it as such then it is indeed a "foundation" ;o) It is really arbitrary what heuristic you apply; in this case you use "standard deviation" as a measure; apparently you may apply other measures like: (max-min)>>(max_l-l) for each sub-volume or ((max-avg)*(avg-min))>>(max_l-l) etc...