Repetition IDs for CircuitOperation

[smitsanghavi]

Useful for keying individual measurements in repeated CircuitOperations

Closes #3663

[95-martin-orion]

What is the use case for the repetitions * self.repetitions / override-all-IDs behavior? It seems like the repetitions / prefixing case should be sufficient.

Separately, we shouldn't rely on the size of this argument to determine its purpose. If we need both behaviors, we should have a separate argument (or method) for each.

[smitsanghavi]

Since we assign default repetition_ids even when the circuit is non-measurement, without override-all-IDs the following will be true

circuit_op.repeat(2).repeat(3) != circuit_op.repeat(6)
circuit_op.repeat(2).repetition_ids == ['0-0', '1-0']
circuit_op.repeat(2, ['a', 'b']).repetition_ids == ['a-0', 'b-0']

The above can be handled by special casing based on if the base or input repetition_ids are "defaults"/None. But if we are overriding in some cases, might as well make that behavior available to the caller. For e.g. if circuit_op.repeat(2).repeat(3) == circuit_op.repeat(6), it would be nice to be able to specify all 6 repetition_ids in the left-side construction.

But I agree that size-based might not be the best way to go about this. How about I add an override_all_ids: Optional[bool] arg? If specified, the repetition_ids size will be validated based on that. If unspecified, we choose the default value of override_all_ids based on repetition_ids and self.repetition_ids.

[smitsanghavi]

Made the above modifications. Let me know if this looks better.

[95-martin-orion]

In my mind, op.repeat(2).repeat(3) != op.repeat(6) is the correct behavior, since they represent different things. Conceptually, it's similar to how [[0, 1], [2, 3]] != [0, 1, 2, 3].

Regarding circuit_op.repeat(2).repetition_ids: if the original circuit_op has an explicitly-defined ID of 0 for its single repetition, I think ['0-0', '1-0'] is correct. However, if no ID has been defined, we want ['0', '1'] instead. This behavior is similar to a raw value vs. a list with one element:

a = 0
a2 = [a, a]  # [0, 0]
a2[0] == a2[1] == 0

b = [0]
b2 = [b, b]  # [[0], [0]]
b2[0][0] == b2[1][0] == 0

With this in mind, our only edge case is the single-repetition, no-ID CircuitOperation; everything else follows the "prefix" case. By extension, we can drop override_all_ids and the extra complexity associated with it.

[smitsanghavi]

Sure, I was more going for the ControlledGate type behavior we have. Chained controls with different number of qubits are automatically "flattened" and the final ControlledGate is equivalent to if all of them were applied at once.

But I guess pow semantics (repeat) don't need to be similar to that addition type operation.

Made the above change and simplified a couple of other things.

[95-martin-orion]

What do you think about making both args optional? I think it'd be nice for repeat(['a', 'b', 'c']) to work without specifying the number of IDs. (Requires addressing repetitions vs repetitions * self.repetitions first)

[smitsanghavi]

That was what I had originally implemented but because of multiple optionals, the default-ing logic was becoming too complicated and I felt it was making the API messier. But if we are making the override-all-IDs case more explicit, I should be able to add this in as well.

[smitsanghavi]

Made repetitions Optional as well. The actual conditions of what defaults to what are a little difficult to describe but added lot of docstrings to compensate. Also, even though it's verbose to explain, the actual behavior seems pretty intuitive to me.

[smitsanghavi]

Now it's even more intuitive :)

[95-martin-orion]

A few relatively minor items and this should be good to go.