Set grad_method=None for ControlledSequence, Reflection, AmplitudeAmplification, and Qubitization.

[astralcai]

Context:
Templates that are not actually supported by parameter_shift should have grad_method=None so that they are decomposed by _expand_transform_param_shift

Description of the Change:

1. Adds the data of components of the templates to the data of the templates such that trainable parameters are tracked
2. Adds grad_method=None for ControlledSequence, Reflection, AmplitudeAmplification, and Qubitization.

Related GitHub Issues:
Fixes #5802
[sc-64967]

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.66%. Comparing base (38b3e74) to head (ba16407).
Report is 250 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #5806      +/-   ##
==========================================
- Coverage   99.66%   99.66%   -0.01%     
==========================================
  Files         415      415              
  Lines       39601    39309     -292     
==========================================
- Hits        39469    39176     -293     
- Misses        132      133       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[trbromley]

CC @KetpuntoG

[KetpuntoG]

CC @KetpuntoG

ControlledSequence, AmplitudeAmplification and Reflection from a theoretical point of view should be differentiable with parameter shift rule. Is this PR due to a PL limitation?

[astralcai]

CC @KetpuntoG

ControlledSequence, AmplitudeAmplification and Reflection from a theoretical point of view should be differentiable with parameter shift rule. Is this PR due to a PL limitation?

I've done some investigation but haven't found out why directly taking parameter shift on the templates return incorrect results. See #5802 for more details.

[albi3ro]

CC @KetpuntoG

ControlledSequence, AmplitudeAmplification and Reflection from a theoretical point of view should be differentiable with parameter shift rule. Is this PR due to a PL limitation?

To clarify, this doesn't say "don't use parameter shift with these templates". It is "decompose before applying parameter shift".

Theoretically, we could develop parameter shift rules for the templates and add them to the grad_recipe. But that seems like a lot of unnecessary work when we can just decompose them. Since this is fixing a bug where pennylane is giving wrong answers and blocking a P0, we should proceed with this fix for now. But if you are interested in developing a customized parameter shift recipe for the templates, I am willing to support.

Would there be a benefit to being able to natively apply parameter shift to these templates instead of to their decomposed form? For some templates, like CommutingEvolution, it might lead to fewer total parameter shift tapes.

[KetpuntoG]

CC @KetpuntoG

ControlledSequence, AmplitudeAmplification and Reflection from a theoretical point of view should be differentiable with parameter shift rule. Is this PR due to a PL limitation?

To clarify, this doesn't say "don't use parameter shift with these templates". It is "decompose before applying parameter shift".

Theoretically, we could develop parameter shift rules for the templates and add them to the grad_recipe. But that seems like a lot of unnecessary work when we can just decompose them. Since this is fixing a bug where pennylane is giving wrong answers and blocking a P0, we should proceed with this fix for now. But if you are interested in developing a customized parameter shift recipe for the templates, I am willing to support.

Would there be a benefit to being able to natively apply parameter shift to these templates instead of to their decomposed form? For some templates, like CommutingEvolution, it might lead to fewer total parameter shift tapes.

Ah okey! That's perfect: first decompose, then shift-rule
Thanks for the clarification

[KetpuntoG]

Thanks! It looks good to me :)

[dwierichs]

Looks good to me! 🎉 Thanks!