Repository navigation
Make finitediff capable of handling irregularly-spaced data #203
Description
gradient only does 2nd order, but it even works with irregularly-spaced points by solving little 3x3 inverse problems, just as I suggested in the Taxonomy paper. https://numpy.org/doc/2.1/reference/generated/numpy.gradient.html
This is pretty low priority. We'd gain an ability to handle nonuniform spacing for one param setting but not others, which is a bit confusing. Just wanted to note down it's a possible improvement.
Activity
pavelkomarov commentedon Sep 2, 2026
Looks like we could also just take on a dependency to findiff, which can do irregular grids, although we wouldn't be able to cherry pick just the derivative calculation stuff, so we'd end up with an indirect dependency on Jax. There must be other alternatives out there too. Then again, do we want to take on such a dependency for just this one method, when we could set up and solve the inverse problems ourselves in bare numpy? Taking on a heavy thing like cvxpy is one thing, because no way are we going to implement that ourselves. And even something like the nonuniform Fourier transform might be worth an extra dependency due to the difficulty of implementing it ourselves, although it only supports one method (maybe two, wavelets?). But for something relatively simple? We've got the wrinkle of repeated applications in here too, which behaves as a characterizable IIR filter when
dtis constant. How would that affect this?[-]2nd order finitediff could call np.gradient[/-][+]Make `finitediff` capable of handling irregularly-spaced data[/+]