8
$\begingroup$

I've read in many papers on higher-dimensional nearest neighbor search that KD-Trees are exponential in K, but I can't seem to determine why.

What I'm looking for is a solid runtime-complexity analysis which explains this aspect of the problem.

| cite | improve this question | | | | |
$\endgroup$
  • $\begingroup$ Quick thought is that k is effectively the dimension of the problem and so it suffers from the "curse of dimensionality." $\endgroup$ – Michael Klein Feb 21 '16 at 6:36
1
$\begingroup$

kNN tends to be exponential because the search space increases with $2^k$. Imagine you partition the space around your search point into quadrants. For k=1 you just have to search two 'quadrants' (higher and lower values), for k=2 it's 4 quadrants, for k=3 it's 8 quadrants, i.e. exponential growth of search space. That is what the kD-tree suffers from, because it has to search $2^k$ sub-branches.

Other trees perform much better, for example the CoverTree. I also found that the PH-Tree works quite well, it seems to consistently take twice as long as the CoverTree for datasets between k=8 and k=27 (I didn't have datasets with higher k).

| cite | improve this answer | | | | |
$\endgroup$
  • $\begingroup$ Note that you can use LaTeX here to typeset mathematics in a more readable way. See here for a short introduction. $\endgroup$ – Raphael Feb 22 '16 at 13:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.