[mlpack] Numeric sensitivity of RangeSearch

[Ryan Curtin]

On Tue, Mar 25, 2014 at 10:34:12PM +0800, Li Dong wrote:
> Hi Ryan,
> 
> Thanks for your quick fixing! It looks correct now. The tutorial is
> the key resources for our users~ It would be better if you could give
> some explanation for this change on the tutorial.

Ok, the updated tutorial has been deployed to the website.

A long time ago, the way trees and dual-tree algorithms worked in mlpack
was different, and all of the pruning rules that were used for nearest
neighbor search and range search worked in such a way that they did not
depend on the triangle inequality.  This meant that the squared
Euclidean distance worked as a distance measure, and could provide a bit
of speedup because it avoided a call to sqrt().

However, more recent improvements included prunes that do depend on the
triangle inequality, which means that an actual metric is required, and
the squared Euclidean distance will no longer work.

I hope that's a better explanation.  Let me know if you'd like me to
clarify further.

Thanks,

Ryan

-- 
Ryan Curtin    | "Leave your stupid comments in your pocket!"
ryan at ratml.org |   - Mark