Unrecoverable internal error when running Cartan decomposition on several circuits

[alexcowtan]

Information

• Qiskit Terra version: 0.10.0
• Python version: 3.6.9
• Operating system: MacOS Mojave

What is the current behavior?

When run on optimisation level 3, Qiskit tries to run a KAK/Cartan decomposition on 2-qubit subcircuits. It frequently suffers from an unrecoverable internal error when doing this. The error is produced non-deterministically; sometimes the circuit transpilation completes, and sometimes it throws the error. The fact that this is random is surprising to me.

Steps to reproduce the problem

from qiskit import QuantumCircuit
from qiskit.compiler import transpile
from qiskit.transpiler import CouplingMap

qsc = QuantumCircuit.from_qasm_file("my_circuit.qasm")
basis_gates = ['u1','u2','u3','cx']

qsc2 = transpile(qsc,basis_gates=basis_gates,coupling_map=None,optimization_level=3)

I can send a wide set of circuits from the IBM routing benchmark sets which produce this error, if this is helpful.

What is the expected behavior?

---

Suggested solutions

---

[kdk]

I was able to replicate this through python -m unittest test/python/quantum_info/test_synthesis.py after recently upgrading to Mojave (this doesn't seem to be an issue on High Sierra.)

The raised error comes from https://github.com/scipy/scipy/blob/f1648dd9e36403f14ef844e3495d74925ff07557/scipy/linalg/decomp.py#L474 which suggests its a scipy/LAPACK issue.

I didn't see the same issue after making a new venv and reinstalling terra. Can you try installing in a new venv and seeing if the error still occurs? If so, can you also include the output of numpy.show_config()?

[nonhermitian]

I think we have seen similar eigh behavior here: #2793

This is an openblas issue with NumPy.

[kdk]

Closing as I'm no longer able to replicate this after #3884. Please reopen if you still see the issue.