Rust exeception raised when transpiling PaulEvolutionGate with measurements against backend

[nonhermitian]

Environment

• Qiskit version: 1.3.1
• Python version:
• Operating system: linux

What is happening?

backend = provider.backend('ibm_nazca')

hamiltonian = SparsePauliOp.from_list([('IXX', 1), ('IYY', 1), ('IZZ', 1),
                                       ('XXI', 1), ('YYI', 1), ('ZZI', 1)])

N = 8
qc = QuantumCircuit(3)
# Initial state 110
qc.x([1, 2])
op = PauliEvolutionGate(hamiltonian, time=1)
qc.append(op.power(N), [0,1,2])
qc.measure_all()

transpile(qc, backend)

gives

thread '' panicked at crates/circuit/src/interner.rs:147:14:
the caller is responsible for only using interner keys from the correct interner
note: run with RUST_BACKTRACE=1 environment variable to display a backtrace

PanicException: the caller is responsible for only using interner keys from the correct interner

How can we reproduce the issue?

run above

What should happen?

it should work. It does if I first compile against no backend, and then use those circuits as input to the backend transpile

Any suggestions?

No clue

[Cryoris]

This is caused by an issue in the UnitarySynthesis for 3+ qubit blocks (when calling op.power you convert the PauliEvo gate into a UnitaryGate), which is fixed by #13591.

Out of interest: Are you using power to repeat the circuit here or are you interested in the unitary of that power? I'm asking because power converts the gate to a matrix and then computes the matrix power. To repeat the gate object, you can use op.repeat(N).

[nonhermitian]

I am interested in applying the gate multiple times, which is what I expect the power op to do. There is no mention of any implementation difference between power and repeat, so an user is most likely unaware of the points you mention.

[jakelishman]

Gate.power:

Raise this gate to the power of exponent.

Implemented either as a unitary gate (ref. UnitaryGate) or as an annotated operation (ref. AnnotatedOperation).

Gate.repeat:

Creates an instruction with self repeated :math:n times.

Do you have suggestions for how we could make this implementation difference clear for you?

[nonhermitian]

Perhaps explicitly mention the power method is numerical and thus only works for small numbers of qubits, and then point users to the repeat method in the docstring for power?

Could also do a similar cross-reference for the repeat docstrings.