Revisit units in API

[yaelbh]

Based on previous discussions, we currently have the following situation in our T1/T2Ramsey API:

• Input:
  • The user can provide delay times in different units.
  • Frequency in T2Ramsey is only in Hz.
  • Initial guess units are identical to the units of the rest of the input: always Hz for frequency, and for T1/T2* the guess unit is the same as the one of the delays.
• Output is always in SI.

I remember that we considered doing everything in SI, including input, and decided not to. But I don't remember what were the arguments.

@nkanazawa1989, working on #427, has now new arguments for allowing initial guesses to be only in SI. I don't like having one rule for delay times and another rule for the initial guess, I find it an inconsistent and confusing API.

Therefore I'd like to revisit the decision to allow delay times not in SI. If we decide that it's really important, then we should keep it also for the initial guess, in spite of @nkanazawa1989's reservations. However we may also decide that removing units from the API and working in SI all over is cleaner.

I'll be happy to hear opinions of users.

[nkanazawa1989]

Pros:

• User can provide parameters in arbitrary units.
• User can provide initial guesses in the same units.

Cons:

• We need more code to maintain
• There will be many edge cases owing to the logic. something like

# multiple conversion factors, we need to integrate unit conversion into data processor.

exp1 = T1(0, delays, unit="s")
data1 = exp1.run(backend)

exp2 = T2(0, delays, unit="dt")
data2 = exp1.run(backend, data1)

• If we have multiple units in experiment parameters, we need to manage mapping of parameter to units. T2Ramsey cannot set units for frequency offset because of lacking of this mapping. Once we implement this mapping, we need to write more unittest.
• (For T1/T2) When we provide delays in dt the initial guess of T1/T2 should be provided in units of dt but it might be unusual for experimentalists to guess these parameters in dt.

From development point of view, the unit support is too much overhead in return for a little user benefit. I think dropping this doesn't hurt usability, i.e.

assert T1(0, delays, units="us") == T1(0, delays*1e-6)

When we drop the unit support, we need to update T1, T2Ramsey, Spectroscopy and Hamiltonian tomography experiment (code and unittest will be simplified).

[yaelbh]

This issue is being worked on in #516