On Sun, Oct 19, 2014 at 3:37 AM, Charles R Harris <charlesr.harris@gmail.com> wrote:
On Sat, Oct 18, 2014 at 7:17 PM, Nathaniel Smith <njs@pobox.com> wrote:
So here are my concerns:
- We decided to revert the changes to np.gradient in 1.9.1 (at least by default). I'm not sure how much of that decision was based on the issues with masked arrays, though, which turns out to be a different issue entirely. The previous discussion conflated the two issues above.
My concern was reproducibility. The old behavior wasn't a bug, so we should be careful that old results are reproducible.
Yep.
- We decided to gate the more-accurate boundary calculation behind a kwarg called "edge_order=2". AFAICT now that I actually understand what this code is doing, this is a terrible name -- we're using a 3-coefficient kernel to compute a quadratically accurate approximation to a first-order derivative.
Accuracy is a different problem and depends on the function being interpolated. As you point out above, the order refers to the *rate* of convergence. Normally that is illustrated with a loglog plot of the absolute value of the error against h, resulting in a straight line once the function is sufficiently smooth over the range of h. The end points are a bit special because of the lack of bracketing. Relative to the interior, one is effectively extrapolating rather than interpolating and things can become a bit unstable. Hence it is useful to have a safe option, here linear extrapolation, as an alternative to a higher order method.
This sounds plausible! But given that we have to (a) pick defaults, (b) pick which ones are included at all, and (c) document the difference in such a way that users can make an informed choice, then I'd kinda like... more precision :-). I'm sure there are situations where one or the other is better, but which situations are those? Do you know some way to tell which is which? Does one situation arise substantially more often than the other?
There probably exist other kernels that are also quadratically accurate. "Order 2" simply doesn't refer to this in any unique or meaningful way. And it will be even more confusing if we ever add the option to select which kernel to use on the interior, where "quadratically accurate" is definitely not enough information to uniquely define a kernel. Maybe we want something like edge_accuracy=2 or edge_accuracy="quadratic" or something? I'm not sure.
- If edge_order=2 escapes into a release tomorrow then we'll be stuck with it.
Order has two common meanings in the numerical context, either degree, or, for polynomials, the number of coefficients (degree + 1). I've noticed that the meaning has evolved over the years, and these days an equivalence with degree seems to be pretty common. In the present context, it refers to the power of the h term in the error term of the Taylor's series approximation of the derivative. The precise meaning needs to be elucidated in the notes, as the order doesn't map one-to-one into methods.
Surely this is still an argument for using a word that requires less elucidation, like degree or accuracy? (I'm particularly concerned because "2nd order derivative" has a *very* well known meaning that's very importantly different.) -- Nathaniel J. Smith Postdoctoral researcher - Informatics - University of Edinburgh http://vorpus.org