Cirq simulation regressions

[mpharrigan]

This issue tracks problems that had to be worked around to bump our Cirq version from 1.4 onwards in #1761. Note that many unit tests in the qualtran.bloqs quantum standard library use a Cirq simulator for numerical simulation. A way forward might be to re-work these unit tests to use native Qualtran tensor contraction.

Note: the performance regressions are likely a consequence of the kludge needed in #1761.

Performance regressions

• qualtran.bloqs.hamiltonian_simulation.hamiltonian_simulation_by_gqsp_test.test_hamiltonian_simulation_by_gqsp
• qualtran.bloqs.rotations.hamming_weight_phasing_test.test_hamming_weight_phasing
  • n=6, 7 now marked as slow

Tests marked slow that likely got slower

• qualtran.bloqs.phase_estimation.qubitization_qpe_test.test_qubitization_phase_estimation_of_walk (also see cirq.unitary does not work for controlled QubitizationWalkOperator operator #1495)
• qualtran.bloqs.phase_estimation.text_book_qpe_test.test_textbook_phase_estimation_qubitized_walk (also see cirq.unitary does not work for controlled QubitizationWalkOperator operator #1495)
• qualtran.bloqs.qubitization.qubitization_walk_operator_test.test_qubitization_walk_operator

Breakages

• qualtran.bloqs.reflections.reflection_using_prepare_test.test_reflection_using_prepare now gets SIGKILL -ed. See test_reflection_using_prepare broken with cirq 1.5 #1737
• qualtran.bloqs.rotations.hamming_weight_phasing_test.test_hamming_weight_phasing n=8 causes SIGKILL
• qualtran.bloqs.rotations.phasing_via_cost_function_test.test_hamming_weight_phasing_using_phase_via_cost_function
• qualtran.bloqs.rotations.quantum_variable_rotation_test.test_qvr_phase_gradient_cost_reg_greater_than_b_grad

Related issues

• cirq.unitary fails for bloq with And followed by And^dagger #1488 -- If Cirq is allowed to decompose gates, it often gets down to And(uncompue=True) which says that it's not unitary. Support Numpy 2 via supporting Cirq 1.5+ #1761 "fixes" this by doing tensor contraction in Qualtran and giving a big tensor to Cirq. This causes performance regression
• Tensor contraction 4-5x slower than cirq.unitary for hamiltonian simulation bloq #1336 -- The Cirq simulator is often faster for small-qubit-count circuits since it decomposes and simulates in one fell swoop instead of constructing a tensor network and then having quimb contract it