Verify values of parametrized rzz gates

[yaelbh]

Summary

Fixes #2000.

This is only a draft. I forgot to write the tests before the code. So the current push contains the logic, but the tests will be added shortly.

Details and comments

@SamFerracin I have a question about the noise learner. I see that NoiseLearner._run calls validate_isa_circuits, like BasePrimitiveV2. I added in BasePrimitivesV2 a call to a new function, validate_rzz_pubs. But I'm not familiar enough with the noise learner to figure if a similar call should be added there too. I tried to add, but NoiseLearner._run receives circuits, unlike BasePrimitiveV2, which receives pubs. Should I make _run receive pubs and add the call to the new function?

[yaelbh]

Following #2035, we should also take into account parameter expressions.

[yaelbh]

A couple of comments:

1. The check for fixed angles will take place in both is_isa_circuit and is_rzz_pub. In is_isa_circuit - because not everyone has a full pub and can call is_rzz_pub - see the noise learner. In is_rzz_pub - because this way the function makes more sense and is more complete, this will be less confusing for developers, who may naturally expect fixed angles to be covered.
2. Parameter expressions are much more complicated, and will be computationally more expensive. An expression can consist of multiple parameters. We will have to evaluate each expression with all the parameter values for its parameters. At least for now this PR continues to skip parameter expressions.

[t-imamichi]

LGTM. But since I'm not an official approver, you need to ask someone else to review it too.

[kt474]

LGTM, can you add a release note?

[ihincks]

There is an asymmetry here in that rounding errors are compensated on the upper end, but not the lower end. for example, -1e-12 is invalid whereas np.pi/2 + 1e-12 is valid.

I'm also a bit surprised that the level is at 1e-3, very far away from machine epsilon. For example, is there a problem moving this to 1e-10 which is still 100x machine epsilon?

Suggested change

	angle < 0.0 or angle > 1.001 * np.pi / 2
	not -EPSILON < angle < np.pi / 2 + EPSILON

[wshanks]

This mirrors line 82 above. I think I contributed to the original choice of 1.001. The hardware is optimized for pi/2 because the main focus is the cz gate. It doesn't blow up if you go just past pi/2 but you don't want to go much past it (not calibrated and will eventually hit some hardware/physical limit). So it is not critical that the limit be tight on pi/2 but whatever is likely not to run into rounding issues should be okay.

[yaelbh]

Then, what's the decision, for both ends?

[yaelbh]

Once there is a decision, I'll of course fix (if needed) not only here, but also in several other places where this check takes place (I've spotted that this check appears in three different places, maybe I'll write a tiny function for it).

[wshanks]

I am okay with 1e-10. I think we only document pi/2 as the limit and here we are just trying to avoid being too strict where in practice the user won't notice the difference.

There was also the point about the asymmetry with negative angles. I am not sure what to do about that right now. The backend will error right now if it gets a negative number. We could change the backend to round small negative numbers to 0 if we wanted symmetry of behavior for the bounds.

[yaelbh]

Based on this conversation, I'm keeping the current behavior for the negative angles. Let me know if it's a problem.

[wshanks]

All conversations must be resolved before merging. Can we mark this as resolved @ihincks? I think the remaining options would be to round up small negative angles/parameters to 0 (a change of scope since the validation code only checks without modifying right now) or to change the backend code to accept small negative angles (out of scope here for now).

[yaelbh]

@kt474 @wshanks It's ready for another review. Note that I added release notes.

[wshanks]

This looks good to me except for a minor formatting change on the release note. I put in two other minor suggestions.

[wshanks]

Looks good to me!