Chapter 18 — Key Takeaways (Framework Comparison and Interoperability)

Part III's synthesis. §18.6's accounting — 90% of Parts I–II transfers — is the one to keep.

The same Bell state, five ways

  framework     lines   what the syntax makes you say
  Braket            1   almost nothing -- the shortest path to a circuit
  Qiskit            3   how many classical bits, and which measurement goes where
  Cirq              3   which qubit objects, and (implicitly) the moment structure
  PennyLane         4   what you want MEASURED, as a return value
  Q#                6   the type, the qubit lifetime, and the cleanup

None of it is verbosity for its own sake. Q#'s ResetAll exists because §15.5's runtime checks it; PennyLane's return is what gets differentiated; Qiskit's classical register is what Chapter 9's dynamic circuits need.

Five result formats for one physical outcome:

  Qiskit       {'11': 490, '00': 510}
  Cirq         {0: 503, 3: 497}                    <- INTEGER keys
  Braket       {'00': 489, '11': 511}
  PennyLane    {'00': 502.0, '11': 498.0}
  Q#           {'(Zero, Zero)': 527, '(One, One)': 473}

★ The endianness table

X on qubit 0 of a two-qubit register:

Qiskit is the outlier — 1 of 3. Parts I–II present little-endian ordering as a fact about quantum computing for thirteen chapters. It is a fact about Qiskit.

Neither convention is wrong. Little-endian makes $|q_1q_0\rangle$ read like a binary number ($q_0$ least significant) — convenient for arithmetic circuits (Ch. 22 QFT, Ch. 23 modular arithmetic). Big-endian makes the index read in the order you listed the qubits — convenient for diagrams. Neither will change. Convert at exactly one boundary.

Who speaks OpenQASM

All five, both directions — a genuine standard.

★★ What QASM transfers, and what it does not

Transfers: gate identity, qubit indices, parameters, ordering. cx q[0],q[1] is unambiguous anywhere.

Does NOT transfer — measured, Qiskit → QASM 2 → Cirq on an asymmetric state:

  qiskit           : [0, 0.7071, 0, 0.7071]
  cirq (via qasm)  : [0, 0, 0.7071, 0.7071]
  match WITHOUT reversal : False
  match WITH    reversal : True

QASM names qubits explicitly, so the program moves correctly — but where qubit 0 sits in a state vector is a per-framework decision QASM has no opinion about.

A STANDARD INTERCHANGE FORMAT DOES NOT REMOVE THE BOUNDARY. IT DEFINES WHAT IS LEFT AT IT.

Also dropped (both re-confirming Chapter 6):

  global phase   1.047198  ->  0.000000                     LOST
  parameters     ['theta[0]','theta[1]'] -> ['_theta_0_','_theta_1_']   MANGLED

Global phase loss is not cosmetic: it becomes a relative phase when the circuit is controlled — Chapter 6 watched it invert an answer ({'1':1757} → {'0':1758}) while passing every equivalence test. Parameter mangling: note the leading underscore; normalize both sides rather than reproducing the rule.

The three responsibilities of a pipeline

qasm = qasm2.dumps(circuit)                              # 1. QASM does this right
counts = cirq_histogram_to_qiskit_counts(hist, n)        # 2. ONE boundary
assert_same_state(qiskit_circuit, cirq_circuit)          # 3. ASYMMETRIC state

Step 3 is the one people skip — Bell and GHZ states are symmetric under bit reversal and prove nothing (Ch. 14 CS1).

Choosing — by capability, not preference

Only three genuine capability gaps: hardware access (Qiskit's open tier), resource estimation (Q#, uniquely), differentiability (PennyLane, by design). Everything else is ergonomics.

"A or B" is often a false choice. pennylane-qiskit differentiates circuits that execute on IBM hardware; Q#'s estimator consumes logical counts from any framework (Ch. 15 §15.9); Braket's native QASM 3 makes it a translation hub.

★★★ What actually transfers (§18.6)

Framework-specific (~10%): API names · little-endian ordering · the primitives architecture · optimization_level · NoiseModel.from_backend.

Framework-independent (~90%):

• The physics. A Bell state is a Bell state.
• The noise signatures. Chapter 11 §11.7's two-axis table reproduced identically in Aer, Cirq, and Braket. Phase damping invisible in all three — a fact about measurement, not software.
• The diagnostic procedures. Chapter 12 §12.7's noise-or-bug tree works anywhere.
• The cost models. 2q gates dominate error · depth is the enemy · T gates dominate fault-tolerant cost · a gradient costs $2n+1$ · connectivity overhead is graph embedding.
• The habits. Independent reference value · asymmetric test case · check the condition number · assert on structure · report the stack.
• The epistemics: a number can be precise, reproducible, and about something other than what you think — Ch. 11, 12, 13, 14, 15, 16, 17. Seven chapters, one lesson, none of it about Qiskit.

🔬 How much does framework choice matter?

Less than this Part's length suggests. Ranked by measured impact:

Framework choice is last, and it is not close.

Common pitfalls

• Believing a standard interchange format removes the endianness boundary.
• Verifying a translation with a Bell or GHZ state.
• Using Operator.equiv() to check a circuit you will control (it ignores global phase by design).
• Scattering bit reversals instead of converting once.
• Forgetting Cirq's QASM importer needs ply.
• Spending on the framework decision what belongs on layout, ansatz, modality, and mitigation.

Project piece added this chapter

vqelab/interop.py — reverse_bits / reverse_state_vector / needs_reversal, to_qasm and from_qasm (raising a clear error for Cirq's missing ply rather than letting ModuleNotFoundError surface from deep in a pipeline), TranslationReport which is returned rather than logged and names the consequence of each loss, and verify_translation which refuses a symmetric test case. 17 tests pass, including test_verify_translation_REFUSES_a_symmetric_test_case and test_qiskit_is_the_endianness_outlier.