Quantum Error Mitigation: Readout, Importance Clifford Sampling

[AlejandroSopena]

General improvement of the code and docstrings.

This PR implements Iterative Bayesian Unfolding to avoid quasiprobabilities when using readout mitigation with the response matrix. It also implements the Importance Clifford Sampling method.

Checklist:

• Reviewers confirm new code works as expected.
• Tests are passing.
• Coverage does not decrease.
• Documentation is updated.

[renatomello]

@AlejandroSopena maybe this is for a different PR, but we should aim to integrate these Clifford-based methods (e.g. ICS, CDR) with the new Clifford backend from @BrunoLiegiBastonLiegi as much as it is possible / worth it.

[MatteoRobbiati]

Thanks a lot @AlejandroSopena for all these modifications and for the big refactoring of the mitigation module.

My only doubt here is about the usage of the apply_readout_mitigation. I agree with your choice, it lighten the code a lot and it is flexible enough to apply the desired readout mit technique, but I find it ambiguous when a mitigation method (e.g. CDR) is applied without applying readout mitigation, because still relying on apply_readout_mitigation.

What about renaming the apply_readout_mitigation function into something more generic like: apply_problem_configuration? Alternatively, we can make the two things explicit in a longer but clearer name: apply_problem_with_readout_conf.

A couple of minor comments follow.

[MatteoRobbiati]

I am fine with keeping this qubit_map for now. I We also close to the #1121, so I think we can then exploit that new feature to fix the circuit on a given topology.

[AlejandroSopena]

I tried to adapt the "manual" transpilation process we use in RTQEM but using the qibo transpiler. Depending on what is missing in #1121, we can merge this and then open a small pull request, or wait.

[AlejandroSopena]

Thanks a lot @AlejandroSopena for all these modifications and for the big refactoring of the mitigation module.

My only doubt here is about the usage of the apply_readout_mitigation. I agree with your choice, it lighten the code a lot and it is flexible enough to apply the desired readout mit technique, but I find it ambiguous when a mitigation method (e.g. CDR) is applied without applying readout mitigation, because still relying on apply_readout_mitigation.

What about renaming the apply_readout_mitigation function into something more generic like: apply_problem_configuration? Alternatively, we can make the two things explicit in a longer but clearer name: apply_problem_with_readout_conf.

A couple of minor comments follow.

I agree. I have changed it to apply_problem_with_readout_conf().

[AlejandroSopena]

@AlejandroSopena maybe this is for a different PR, but we should aim to integrate these Clifford-based methods (e.g. ICS, CDR) with the new Clifford backend from @BrunoLiegiBastonLiegi as much as it is possible / worth it.

Yes, I agree @renatomello . I think we can implement this in a new PR because it only involves an improvement when running the training circuits but not the final result (although a very useful improvement).

[BrunoLiegiBastonLiegi]

Thanks @AlejandroSopena, it looks good. I made some minor comments and in general I am not a huge fan of the two names of the functions _circuit_conf and apply_problem_with_readout_conf. I would probably prefer something like _execute_circuit and _get_expectation_val_with_readout_mitigation.

[BrunoLiegiBastonLiegi]

So, in principle one can consecutively apply randomized_readout and resp_mat_readout, right?

[AlejandroSopena]

No, one or the other. I have updated the code.

[MatteoRobbiati]

Thanks!
A couple of comments and some minor suggestions follow.