Combine Pauli measurements and postselection

[ddddddanni]

No description provided.

[eliottrosenberg]

Thanks, Danni! This is a really great start! In addition to the comments below, it would be great if we could return the fraction of the bitstrings that survive the postselection, and maybe even more granular information in addition, like how many violate each of the symmetries. Maybe we can also optionally return the raw bitstrings to the users in case they want to do more analysis on them.

[eliottrosenberg]

-> "The final state generated by the circuit is an eigenvector ... with the specified eigenvalues."

[eliottrosenberg]

Maybe this should default to None?

[eliottrosenberg]

I realized that if I try to construct a dictionary with PauliSums as keys, I get the error TypeError: unhashable type: 'PauliSum'. Do you have any suggestions of what to do here? Should we use tuples or make a new data structure?

[ddddddanni]

Curious is PauliSum just a combination of PauliStrings? If so, can we remove the PauliSum from the postselection_symmetries: dict[ops.PauliString | ops.PauliSum, int], and make it postselection_symmetries: dict[ops.PauliString, int] instead?
For users who originally want to make

psum = cirq.PauliSum.from_pauli_strings([
    pauli_1,
    pauli_2,
    pauli_3
])
postselection_symmetries[psum] = 1

they can split the psum into paulistrings and make the postselection_symmetries dict by

postselection_symmetries[pauli_1] = 1
postselection_symmetries[pauli_2] = 1
etc

Wdyt?

[eliottrosenberg]

There are some cases where the symmetry has to be specified by a PauliSum. For example, if total number is conserved, then $\sum_i Z_i$ is conserved, but each $Z_i$ individually is not necessarily.

[ddddddanni]

Ahh got it, thanks for the clarification! In this case, I think we can define postselection_symmetries as a list of tuples such as Sequence[tuple[cirq.PauliString | cirq.PauliSum, int]]?

[ddddddanni]

Thanks, Danni! This is a really great start! In addition to the comments below, it would be great if we could return the fraction of the bitstrings that survive the postselection, and maybe even more granular information in addition, like how many violate each of the symmetries. Maybe we can also optionally return the raw bitstrings to the users in case they want to do more analysis on them.

I actually have

@attrs.frozen
class PostFilteringSymmetryCalibrationResult:
    """Result of post-selection symmetry calibration.
    Attributes:
        raw_bitstrings: The raw bitstrings obtained from the measurement.
        filtered_bitstrings: The bitstrings after applying post-selection symmetries.
    """

    raw_bitstrings: np.ndarray
    filtered_bitstrings: np.ndarray

Maybe I can add:

1. fraction_of_survived_bistrings: float
2. symmetry_to_filtered_bistrings: dict{paulistring|paulisum, list[bitstrings]}?

[eliottrosenberg]

Do we need this input given that we have circuits_to_pauli_params?

[eliottrosenberg]

I guess the user doesn't use this function directly, so it's ok

[eliottrosenberg]

@ddddddanni ah, I missed that I can do .calibration_result.filtered_bitstrings and .calibration_result.calibration_result.filtered_bitstrings. That's great! Thanks!

[mhucka]

@ddddddanni Just wanted to gently ask what the status of this is. No pressure; I know this is a draft.

[ddddddanni]

@ddddddanni Just wanted to gently ask what the status of this is. No pressure; I know this is a draft.

Thanks for checking!! I was working on the other PRs and Q3 okrs. I plan to get back to this PR this week.

[mhucka]

@ddddddanni No problem at all. Thanks for your quick reply!

[eliottrosenberg]

@ddddddanni This looks really great! I tested it in this colab, and it is easy to use and seems to work as intended! My only nit is that the docstrings could be a bit clearer.

[eliottrosenberg]

This isn't exactly a dictionary, so let's modify the docstring.

[ddddddanni]

Done in #7746!

[eliottrosenberg]

We can also skip this if we're postselecting on at least one of the qubits in each observable, since then the postselection already does some of the readout mitigation. I think you are already skipping this in the analysis, but we can also skip it in the data taking (maybe you are already doing this), and it should be clear in the docstring that we are doing so.

[ddddddanni]

By "skip this if we're postselecting on at least one of the qubits in each observable" - do you mean skip readout data taking (generating random bitstrings, finding e0 and e1 for each qubits and calculating confusion matrices)? Yes we are not doing it if we use posteselecting method.