On Mon, Oct 5, 2009 at 3:20 PM, Anne Archibald <peridot.faceted@gmail.com>wrote:
Just curious, Anne: have you anything in particular in mind (i.e., are
2009/10/5 David Goldsmith <d.l.goldsmith@gmail.com>: there
some small - or gaping - holes in scipy (IYO, of course) which you know could be filled by a careful implementation of something(s) extant in the literature)?
Well, not exactly - the examples I had in mind were minor and/or in the past. I ran into problems with scipy's hyp2f1, for example, so I went and looked up the best algorithm I could find for it (and I think I contributed that code). I wanted the Kuiper test as an alternative to the Kolmogorov-Smirnov test (it's invariant under cyclic permutations, and is sensitive to different features of the distribution) so I looked up the test and the special function needed to interpret its results. (I haven't contributed this to scipy yet, mostly because I chose an interface that's not particularly compatible with that for scipy's K-S test.) And on a larger scale, that's what scipy.spatial's kdtree implementation is.
For examples where I think a bit of lit review plus implementation work might help, I'd say that the orthogonal polynomials could use some work - the generic implementation in scipy.special falls apart rapidly as you go to higher orders. I always implement my own Chebyshev polynomials using the cos(n*arccos(x)) expression, for example, and special implementations for the others might be very useful.
At what order does the scipy implementation of the Chebyshev polynomials fall apart? I looked briefly at that package a long time ago, but never used it. I ask so I can check the chebyshev module that is going into numpy. Chuck