Composite analysis can not handle single shot memory data

[eggerdj]

Informations

• Qiskit Experiments version: 660766a
• Python version: 3.8.5
• Operating system: Linux

What is the current behavior?

The following fails:

specs = []
for qubit in [0, 1, 2]:
    freq01 = backend.defaults().qubit_freq_est[qubit]
    specs.append(QubitSpectroscopy(qubit, frequencies=np.linspace(freq01-20e6, freq01+20e6, 51), backend=backend))

multi_spec = ParallelExperiment(specs, backend=backend)
multi_spec.set_run_options(meas_level=MeasLevel.KERNELED, meas_return="single")

mexp_data = multi_spec.run()

With error

ValueError: x and y must be the same size

when the analysis attempts to create a scatter plot of the data.

Steps to reproduce the problem

Execute the code above.

What is the expected behavior?

The code above should run and produce meaningful results.

Suggested solutions

I suspect that the cause of this is the data processing. Nothing is telling the data processing chain for each individual experiment how to extract the data from the "memory" in the experiment_data object. This points to a missing node in the data processing chain.

[yaelbh]

1. If the spectroscopy experiment has default transpile or run options, they don't propagate to the composite experiment itself, therefore you have to manually set them in the composite experiment.
2. It is possible that an explicit call to block_for_results is still required for composite experiments.

P.S. I see that you took care of the run options. Didn't check if there are particular transpile options.

I don't know about the data processor, but as far as I remember the code was changed to support measurement level 1, including the case of parallel experiments. Parallel readout angle (with measurement level 1 like here, but meas_return=avg) works well.

[eggerdj]

@yaelbh have you tried with meas_return=single ?

[yaelbh]

No, and if I remember correctly the readout angle experiment is written in a way that assume avg.

[wshanks]

The problem is that _marginalize_data in composite_analysis.py does not handle single shot data for memory. Here is where it marginalizes memory data: https://github.com/Qiskit/qiskit-experiments/blob/db3962c57f115b5d2065d5526d5459e5700d6c28/qiskit_experiments/framework/composite/composite_analysis.py#L152-L158

If you measure 3 qubits like in the example above and measure 1024 shots, composite_clbits is [[0], [1], [2]] and datum['memory'].shape is (1024, 3, 2), so instead of picking out the appropriate qubit the function picks out the first shot leaving something with shape (1, 3, 2) instead of (1024, 1, 2). When this is passed to the curve fitting function, you get the shape mismatch error.

[chriseclectic]

Composite analysis currently only supports level 2 measurement data. The handling of memory was intended for single-shot level-counts which returned a list like "memory": ["000", "000", "111", ...] when running using the run option memory=True (using memory key for both level 1 and 2 is confusing here).

Adding support for marginalizing level 1 data is a feature that needs to be added to this class.

[wshanks]

Ah, okay. It seems that the support single-shot level 2 also incidentally enabled support for average level 1 data.

[coruscating]

Fixed by #801.