Why is the Qiskit transpiler not resolving the equivalence $I = R_y(-\theta) R_y(\theta)$?

Question

Suppose you are given the following quantum circuit in Qiskit:

from numpy import pi
from qiskit import QuantumCircuit

angle = pi / 3

qc = QuantumCircuit(1)
qc.ry(angle, 0)
qc.ry(-angle, 0)

qc.draw('mpl')

The circuit simply performs two rotations around the $y$-axis with angle $\theta = \pi/3$ and then $-\theta$. These operations should be resolved in a simple identity because $I = R_y(-\theta) R_y(\theta)$. However, the Qiskit transpile function seems not to do the job as expected in this case:

from qiskit import transpile

tqc = transpile(qc, optimization_level=3)
tqc.draw('mpl')

Moreover, trying exactly the same thing with $R_x$ (qc.rx) or $R_z$ (qc.rz) rotations, the transpilation runs perfectly returning an empty circuit. Is this a bug in Qiskit? Or maybe I'm missing something?

Answer 1

It was a bug, now fixed in release 0.40.0.

In the release notes, you can read:

Fixed an issue with the transpile() function when run with optimization_level=3 and no backend, basis_gates, or target argument specified. If the input circuit contained any 2 qubit blocks which were equivalent to an identity matrix the output circuit would not be as optimized as possible and and would still contain that identity block. This could have been corrected previously by specifying either the backend, basis_gates, or target arguments on the transpile() call, but now the output will be as simplified as it can be without knowing the target gates allowed. Fixed #9217