Where is it likely that accuracy is being lost by the software?
[ ] The tests pass for a single application of M2M, as well as for the full upward pass tested at a distant point. But, exactly where does the expansion start to break down? How close to the expansion surface does it no longer provide accuracy? Knowing this will help understand if M2M surface created at wrong distance from expansion centre in order to provide required accuracy. Check against Ying Paper for reccomended distances.
[ ] Similar testing is required as above for the L2L operator, where tests superficially seem to pass
[ ] What is the quantifiable effect of choice of cutoff singular value in the M2M and L2L calculation?
[ ] I know that operator compression doesn't impact accuracy as much as error from the FMM, but how much so? Is this making a significant difference to pyexafmm, can I quantify it?
[ ] Is there a logical error in the interaction lists? Easiest to test with uniform grids/particle distributions. I don't think that there is, but if there is, it would also have an impact on accuracy.