Thu Feb 21 08:33:49 CST 2013
On 21 Feb 2013 13:51, <email@example.com> wrote:
> On Thu, Feb 21, 2013 at 8:47 AM, Charles R Harris
> <firstname.lastname@example.org> wrote:
> > On Thu, Feb 21, 2013 at 2:54 AM, Nathaniel Smith <email@example.com> wrote:
> >> I'm not an expert on spline implementation in general, but for
> >> regression applications, we need to be able to evaluate the individual
> >> functions in an arbitrary spline basis at irregular x points, which is
> >> well supported by the current API. (splev only computes linear
> >> of the full basis set, so to get individual basis functions we have to
> >> multiple evaluations with linear combinations like [0, 0, 1].) So
> >> this use case in mind would be nice :-).
> > <snip>
> > Can you be more specific here? Design matrices?
> The spline analog of np.vander.
> Once we have the basis function as a design matrix, we can use all the
> regular tools for linear model estimation (statsmodels).
> number of knots usually considerably smaller than number of
> observations for penalized or smoothing splines.
> That's what I see, and Nathaniel might have in mind.
Yes, that's what I meant.
Bonus points if we can easily get compatible implementations of R's bs()
and ns(). (bs() is trivial, we already have it, just inconveniently; I'm
pretty sure ns() is the same thing with those linear boundary conditions
someone else mentioned up-thread.)
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the SciPy-Dev