R-trees always give exact answers. So how can the answer then be different?
That is, unless you have some implementation errors, of course.
In my experiences, the R*-tree splitting algorithm works best. Ang-and-Tan seems to be "linear", but pretty useless for real query scenarios.