T1 with measurement level 1

[yaelbh]

Make T1 work with level 1 measurements

[yaelbh]

Will do after #427

[yaelbh]

Need to remember to repeat (maybe afterwards and not in parallel) for all experiments for whom level 1 is relevant

[yaelbh]

It looks like there are many ways to approach the task of using IQ measurements in a T1 experiment:

1. Using meas_return=avg: with measurement level 2, every T1 circuit is assigned with a probability of measuring 1. In our case, of measurement level 1, we'll replace this probability with the distance of the average from the 0 center, divided by the distance between the centers. As far as I see, this is what's done in experiments that are already using measurement level 1, with the addition of a linear regression that aligns all the averages on the line between the centers.
2. Using meas_return=single, I see two options:
   • Consider only the best shots and continue as usual.
   • Take into account the shot reliability when fitting (assign more weight to reliable shots).

Is there any preference to any of these options? Is there a reason why Option 1 was selected for Rabi? And inside this option, in the Rabi case, how is accuracy improved (if at all) by the choice to perform a linear regression? Should we implement one or more of the options, and if more than one option, then which one should be the first?

[blakejohnson]

I would encourage you to take path (1) because it works even if the measurement process is not yet well calibrated, where it is difficult to assign a "shot reliability". In approach (1) I suggest using PCA to find the optimal mapping onto the real-number line as it is more robust than simple linear regression.

[wshanks]

In my opinion, this should be handled at the data processor level rather than within the experiment. An experiment should specify a data format that it wants to start from and the user should supply a data processor that can produce that format and specify the measurement level that that data processor needs to start from.

It would be nice to make measurement level and data processor handling more automatic (like set measurement level 1 and an appropriate data processor in one place without needing to update the way each experiment in a script is run), but so far that kind of global configuration has been avoided in qiskit-experiments.

[yaelbh]

With meas_level=1 and meas_return='avg', the curve analysis mechanism automatically creates a data processor that performs normalization followed by SVD. So, just by looking at the code, it looks possible that T1 with IQ is already supported, and all the user has to do is to set meas_level and meas_return. However, when I try it, it fails with a crash. We'll check why.

[yaelbh]

I think it's been done.