Add fft submodule to qml.math

[josh146]

No description provided.

[github-actions]

Hello. You may have forgotten to update the changelog!
Please edit .github/CHANGELOG.md with:

• A one-to-two sentence description of the change. You may include a small working example for new features.
• A link back to this PR.
• Your name (or GitHub username) in the contributors section.

[mariaschuld]

Nice! I wonder if the numpy_mimic is causing the tests to fail?

And what would we need to do, just add tests? Or are the math module functions untested?

[josh146]

Nice! I wonder if the numpy_mimic is causing the tests to fail?

Tests are failing because I stupidly passed self to super(), I forget that you don't have to and that is automatic 😆

Just fixed that, so tests should pass now.

And what would we need to do, just add tests? Or are the math module functions untested?

• Finish ensuring that all TF/Torch FFT functions that are needed are accessible. For now, I imagine just the main ones (fft, ifft, fft2 etc.) as well as some of the helper function (fftshift) should be sufficient?
• Writing tests 🙂

[codecov]

Codecov Report

Merging #1440 (615ecd8) into master (03e7f50) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master    #1440   +/-   ##
=======================================
  Coverage   99.82%   99.82%           
=======================================
  Files         329      329           
  Lines       28822    28848   +26     
=======================================
+ Hits        28771    28797   +26     
  Misses         51       51           

Impacted Files	Coverage Δ
pennylane/math/__init__.py	100.00% <100.00%> (ø)
pennylane/math/single_dispatch.py	99.72% <100.00%> (+0.01%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[josh146]

@glassnotes @mariaschuld @dwierichs... is this still needed? I have completely forgotten the context to this PR :(

[dwierichs]

It seems like it is almost done? We have not talked much about differentiability of parameter shift rules or reconstruction functionalities with respect to the frequency or reconstruction/evaluation points since summer. It might be nice to leave this open and merge it in once we want some functionality like this, but for now it does not appear to be needed?

[josh146]

@dwierichs yes as far as I can tell, it only needs tests. Okay, let's leave it open for now

[josh146]

@dwierichs is this something that is now worth adding?

[dwierichs]

I tried to write tests for this.
Can't get the right result in Tensorflow so far, and the following code breaks my Python instance:

import torch
def real_fft(x):
    return torch.real(torch.fft.fft(x))

x = torch.tensor([0.5], requires_grad=True)
torch.autograd.functional.jacobian(real_fft, x)

[josh146]

@dwierichs did you get the tests working based on the recent commit?

[dwierichs]

@josh146 Yes. Complex-valued derivatives are a bit strange in Tensorflow, to me at least. And the torch problem was self-made, I had some local version collision.
This PR should be ready to go soonish, modulo copying the now working tests to the other three registered functions.

[dwierichs]

This could use some reviews now 😅 @josh146

[josh146]

Nice work @dwierichs! All looks good to me, and very thorough test coverage. Unfortunately, I cannot approve as I opened this PR 😆