More realistic noise models for Fake Backends

[ajrazander]

Environment

• Qiskit Terra version: 0.21.2
• Python version: 3.8.8
• Operating system: Mac Darwin

What is happening?

The simulated backends under qiskit.providers.fake_provider can have dramatically less noisy output than the actual backends. It can rather misleading at times.

How can we reproduce the issue?

See the jupyter notebook in this PR: https://github.com/qiskit-research/qiskit-research/pull/33 and look at the figure under the section "Only mitigation by transpiration". Running the notebook with the FakeLagos() backend gives dramatically better results than running the notebook with the real backend ibm_lagos().

Here's an example of the output difference between the ibm_lagos device and it's fake backend. The fake backend produces this first figure.

The red line is the exact, noiseless oscillation, and the blue line is the output from the simulated backend. They don't match exactly because there is some noise from the fake backend--which is desired. Note that the dynamics basically match the expected values (red).

Running the same circuits through the actual backend, there's almost no dynamics due to noise.

A moderate difference between the simulated noise model and the real hardware is expected. In this case, however, the reported gate errors and qubit parameters of the FakeLagos match fairly well with the real hardware at the time these plots were computed, and yet there is a stark difference in output.

What should happen?

The fake backends should have a more faithful noise model to the hardware. This doesn't need to be overly exact, but based off the data shown above, there seems to be some important noise factors missing from the fake backend noise models.

Any suggestions?

No response