Folks, I'm trying to implement by brute force some median-based linear regression techinques, requiring a lot of median computations (for n elements, I have about n*n medians to compute). And as I have a lot of data to process that way, I'm a bit pecky about optimization. I've noticed that for 1D and 2D arrays, the basic numpy version of median is far faster than the nD version of scipy.stats (you can still choose to compute along any of the 2 axes with numpy.median by playing with transpose). Would it be useful to modify the scipy.stats version, or should it be left to the user ? And on related matter, numpy based this time: the numpy.median version seems a bit strange: what's the idea behind using `msort` twice (around line 810 in the 0.9.5 version) ? Thanks for you feedback in advance -- Pierre GM