qiskit.transpile with optimization_level=3 does not optimize away an identity operation like cp with phase 0

[dixr]

Environment

• Qiskit version: 1.2.4
• Python version: 3.11.3
• Operating system: Ubuntu Linux

What is happening?

Why does qiskit.transpile() function with optimization level 3 not remove a cp (controlled phase) gate when its parameter is 0? Since a cp(0) gate is equivalent to an identity operation, it should be optimized away.

How can we reproduce the issue?

import qiskit
import qiskit_aer
circuit = qiskit.QuantumCircuit(2)
circuit.cp(0, 1, 0)
circuit = qiskit.transpile(circuit, qiskit_aer.AerSimulator(), optimization_level=3)
circuit.data

Output:

[CircuitInstruction(operation=Instruction(name='cp', num_qubits=2, num_clbits=0, params=[0.0]), qubits=(Qubit(QuantumRegister(2, 'q'), 1), Qubit(QuantumRegister(2, 'q'), 0)), clbits=())]

What should happen?

The gate should be removed.

Any suggestions?

No response

[Cryoris]

Short answer: Yeah we currently don't do value-based optimization for gates that are already in the target basis (note that the AerSimulator accepts any standard gate in its basis). To resolve this you could set a specific basis that forces Qiskit to re-synthesize the operation. For example

import qiskit
circuit = qiskit.QuantumCircuit(2)
circuit.cp(0, 1, 0)
circuit = qiskit.transpile(circuit, basis_gates=["u", "cx"], optimization_level=3)  

That should also be in the correct format for the AerSimulator, I think.

Longer answer: This is generally an optimization we miss, here's a more basic example:

import qiskit

circuit = qiskit.QuantumCircuit(1)
circuit.ry(0, 0)
circuit = qiskit.transpile(circuit, basis_gates=["ry"], optimization_level=3, callback=callback)
print(circuit)  # Ry(0) present!

circuit = qiskit.transpile(circuit, basis_gates=["u"], optimization_level=3, callback=callback)
print(circuit)  # empty!

In the second example, the Ry(0) gets converted to U(0,0,0) and subsequently Optimize1qGatesDecomposition removes the gate. We could maybe expand that workflow to remove 0-angle rotations with the same tolerance that it uses in converting Euler angles to the identity 🤔

[Cryoris]

Update: this should get fixed by #12384.