Appendix H: The Quantum Software Ecosystem

A map of what exists, what it is for, and which parts are load-bearing. This appendix ages faster than any other in the book — treat it as a snapshot with reasoning attached, and check the reasoning rather than the list.

Full-stack frameworks

The honest summary from Chapter 18: they are more similar than the marketing suggests, and the differences that matter are endianness, scheduling model, and differentiability — not gate names.

Simulators

★ The 30–35 qubit boundary decides advantage claims. Chapter 35: below it, a result is checkable classically and therefore is not evidence of advantage. Above it, you cannot verify your own answer.

Chemistry

• PySCF — the classical workhorse. Integrals, Hartree–Fock, and the CCSD(T) baseline you must beat.
• PennyLane qml.qchem — what Chapter 36 used; builds qubit Hamiltonians in one call.
• OpenFermion — Google's fermionic mappings and Hamiltonian manipulation.
• Qiskit Nature — IBM's chemistry stack. Restructured more than once; check the version.

⚠️ Chapter 36's finding applies to all of them: the active space you choose dominates every downstream engineering decision, and none of these tools will tell you it was wrong.

Optimization

• cvxpy with SCS or Clarabel — the SDP solver behind Goemans–Williamson (Chapter 37).
• networkx — graph construction and instance families.
• Gurobi / CPLEX / HiGHS — the incumbents. Case Study 37.1 turns on a solver log line printing a 0.046% optimality gap that nobody had read.
• Qiskit Optimization — QUBO/Ising converters.

Machine learning

• PennyLane — the framework Part VI is written in.
• scikit-learn — the baselines. LogisticRegression, SVC, kNN, RandomForest. Chapter 32–35 measured quantum models tying or losing to these on every dataset.
• TensorFlow Quantum — Cirq-based; less active than it was.

Error correction

• Stim — Clifford simulation for QEC, extremely fast. The standard tool.
• PyMatching — minimum-weight perfect matching decoder.
• Union-find and neural decoders — the real-time decoding frontier, and Chapter 40 §40.1's highest-demand skill.

Cryptography

• cryptography ≥ 46 — ships ML-KEM natively (asymmetric.mlkem). Chapter 38 measured ML-KEM-768 at 201.7 µs against X25519's 76.6 µs.
• liboqs / Open Quantum Safe — the broader post-quantum toolkit.

Cloud platforms

Chapter 39's warning: aggregators make different modalities look interchangeable behind one API, and the invoice differs by 3,718× for identical work.

What to actually learn

If you are choosing where to spend effort, Chapter 40 §40.1's headcount ordering is the guide:

1. Qiskit — the largest ecosystem and the most jobs.
2. PennyLane — if you do anything variational.
3. Classical software engineering — testing, reproducibility, CI. Chapter 39 §39.8 found quantum computing behind on this, not ahead.
4. Stim + a decoder — if error correction interests you, which is where the field's central problem is.
5. A classical solver — Gurobi, PySCF, scikit-learn. You cannot evaluate a quantum claim without knowing what it must beat, and this is the single most common gap.

What this appendix will get wrong

Every specific version number here will be stale, several of these projects will be renamed or abandoned, and at least one will be replaced by something better.

What will not go stale: the reasoning. Endianness will still differ between frameworks. Simulability will still end around 30–35 qubits. Classical baselines will still be the thing you have to beat, and there will still be nobody in the building who has run them.

See also: Chapter 1 (the landscape), Chapter 18 (interoperability measured), Chapter 39 (platforms), Chapter 40 (careers), Appendix C (installing this).