I guess you've thought of using complex numbers? Assuming your actual observations can never be complex...
Like... having an array with only the 3 dense axes and encoding the "hidden dimensions" in the imaginary part of a given observation. Maybe the best encoding is an "id" (the relational approach?) rather than a packed integer, the latter could let you down at a future date, unless you can put a cast-iron upper bound to the number of planes in each hidden dimension. Nice to project-out the (quite literally) real data. Or the (imaginary) pseudo-coordinates. Bit nasty if you need to do arithmetic on the whole array. On Sun, Jul 3, 2011 at 2:40 PM, david alis <[email protected]> wrote: > sorry about the typo - it's 80 not 800. > > "A J-sparse array implementation would have 10 sparse axis. > This means that for every observation there would 10 extra numbers (i.e. > integers). > i.e. for each 8 bytes of useful data there needs to be 80 bytes of support > (J64)." > ---------------------------------------------------------------------- > For information about J forums see http://www.jsoftware.com/forums.htm > ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
