Resolve issue2578

[gluonhiggs]

Description

Here is the original issue.
Basically, the error happens technically due to the fact that rzz is not recognized by QASM.
In general, this can also happen to other gates that are pre-built in any framework but not existing in QASM.
So, the idea is that we should decompose() some parts of the argument circuit such that these gate are more recognizable to QASM, and then decompose() one more time to obtain the decomposition by the basis gates.

My original thought was:

Basically, the error happens technically due to the fact that rzz is not recognized by QASM.
In general, this can also happen to other gates that are pre-built in any framework but not existing in QASM.
So, the idea is that we should decompose() the argument circuit to semi-basic gate using QFT such that these gate are recognizable to QASM, and then decompose() one more time to obtain the decomposition by the basis gates. Turns out it was not working even passing all tests.

---

License

• I license this contribution under the terms of the GNU GPL, version 3 and grant Unitary Fund the right to provide additional permissions as described in section 7 of the GNU GPL, version 3.

Before opening the PR, please ensure you have completed the following where appropriate.

• I added unit tests for new code.
• I used type hints in function signatures.
• I used Google-style docstrings for functions.
• I updated the documentation where relevant. no doc updated.
• Added myself / the copyright holder to the AUTHORS file

[github-actions]

Hello @gluonhiggs, thank you for submitting a PR to Mitiq! We will respond as soon as possible, and if you have any questions in the meantime, you can ask us on the Unitary Fund Discord.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 98.72%. Comparing base (9561b9f) to head (272b0d5).
Report is 2 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2579   +/-   ##
=======================================
  Coverage   98.72%   98.72%           
=======================================
  Files          92       92           
  Lines        4168     4169    +1     
=======================================
+ Hits         4115     4116    +1     
  Misses         53       53           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[andreamari]

First, thanks @gluonhiggs for the quick fix!
I am not sure about this line where you are decomposing simple one-qubit gates such as "cx", "rx", "ry". Those gates are supported by Mitiq/Cirq and we risk to over-decompose the circuit.

I understand that tests are passing, but this change may unnecessarily increase the depth of the output circuit and this is a problem for the efficiency of the quantum computation. For example, if you input a short qiskit circuit made by a single "rzz" gate and many "cx", "rx", "ry" , Mitiq would decompose all gates producing an extremely long output circuit.

I don't know if it works, but I would suggest,it after the existing line circuit=circuit.decompose(),
to add something like this:

circuit = circuit.decompose(gates_to_decompose=["rzz", "rzx", "rxx"], reps=10)

In this way, Mitiq will first apply a general first decomposition step as usual.
Afterwards, if there are still "rzz", "rzx", "rxx" gates left in the circuit, it will simplify only those specific gates.

What do you think?

[cosenal]

For issues such as #2578 and #558, we should just find the maximum set of gates that are supported by both Qiskit's QASM2 serializer and Cirq's QASM2 parser, and transpile to that set. @andreamari wdyt?

[gluonhiggs]

@cosenal I thought about that and tried to transpile not to the maximum set of the common gates, but to the set of the gates that can be used to construct the others, i.e., ["u1", "u2", "u3", "cx", "rx", "ry", "rz"], then decompose(). I also thought that this would be efficient because transpile comes with optimization_method. However, the same error still happens. I even tried with a broader set, e.g., ["u3", "u2", "u1", "cx", "id", "x", "y", "z", "h", "s", "t", "rx", "ry", "rz"], but nothing has changed. Generally, I think it would be hard to determine the set because there will be more n-qubit gates, right?
@andreamari If we choose the set gate to decompose like ["rzz", "rzx", "rxx"], I'm wonder if we also have to consider other combinations of x, y, and z here like ryy, ryz, and so on? Because I tried to replace rzz by ryy in the circuit (in the case ["rzz", "rzx", "rxx"]), there will be an error.

test_qc1 = QuantumCircuit(2)
test_qc1.rx(0.1, 0) 
test_qc1.ryy(0.1, 0, 1)
print(test_qc1)
print(convert_to_mitiq(test_qc1))

And i think over decompose will happen if we add more rii gates (where i in {x, y, z})?

[cosenal]

What I am proposing is something like

circuit = qiskit.transpile(
    circuit,
    basis_gates=QASM2_GATE_SET,
    optimization_level=0,       
)
 mitiq_circuit = from_qasm(qasm2.dumps(circuit))

without the decompose(), because that would bring the gate set to another set that may not be supported by QASM serializer/parser.

I tried with QASM2_GATE_SET = ["u1", "u2", "u3", "cx", "rx", "ry", "rz"] and it seems to work fine. I don't think that's the maximal set though. Having the maximal QASM-supported set there is the best we can do. And something smaller than the maximal will be sub-optimal when it comes to the size of the output circuit (as Andrea pointed out.)

[gluonhiggs]

Yeah. It works. Let me have some tests. @andreamari what do you think? I think we can increase the level of optimization or expose the option to the interface.
Update 1: @cosenal i have tried different variants of this approach, but some tests can't be passed.
Update 2: Because transpile is for transforming a circuit to satisfy specific constraints including hardware, I don't think we should use transpile here. To the best of my knowledge, decompose is just a "pure" break of circuits to some basis gates, and to optimize the decomposition (which is generally a difficult problem, even though we have algorithms to do that for some circuits), we either should choose a good set gates_to_decompose globally optimal (what we got so far is "u1", "u2", "u3", "cx", "rx", "ry"), or we suggest a change on qiskit repo. I'd love to hear your thoughts on this.