Further Reading: PennyLane

Tagged Tier 1 (confident it exists and recommended) and Tier 2 (real and worth seeking, but verify the current version or URL).

The primary sources

• The PennyLane documentation and demos. Unusually good — the demo library is a genuine teaching resource rather than marketing, with worked implementations of VQE, QAOA, quantum kernels, and the barren plateau measurement this chapter reproduces. Start with the "Key Concepts" pages, then the demos closest to your problem. Tier 1.
• The PennyLane API reference for qml.grad, diff_method, and the device classes. In particular the differences between parameter-shift, backprop, adjoint, and finite-diff, and which are available on hardware. §16.4's cost analysis only applies to parameter-shift. Tier 1.
• The pennylane-qiskit and pennylane-cirq plugin documentation. How to run a QNode on the backends from Chapters 7–14, which is arguably PennyLane's best use: a differentiation layer on top of another framework. Tier 1.
• The PennyLane templates catalogue. StronglyEntanglingLayers, BasicEntanglerLayers, AngleEmbedding, AmplitudeEmbedding, and the rest. Read the circuit diagrams — Case Study 1 exists because the first gate of Rot was not something I had looked at. Tier 1.

On the parameter-shift rule

• Mitarai, Negoro, Kitagawa, and Fujii, "Quantum Circuit Learning" (2018), Physical Review A 98, 032309. Where the parameter-shift rule enters the quantum machine learning literature. Tier 1.
• Schuld, Bergholm, Gogolin, Izaac, and Killoran, "Evaluating analytic gradients on quantum hardware" (2019), Physical Review A 99, 032331. The definitive treatment, by the PennyLane team: the derivation, the generalization to gates whose generators have more than two eigenvalues, and the variance analysis explaining why this beats finite differences under shot noise. The paper to read if you read one. Tier 1.
• Wierichs, Izaac, Wang, and Lin, "General parameter-shift rules for quantum gradients" (2022), Quantum 6, 677. The four-term and general rules — what happens when §16.3's two-term derivation does not apply. Tier 1.
• Any treatment of the numerics of finite differencing. The truncation-versus-cancellation trade-off in §16.3's table is classical numerical analysis, and the general treatment is clearer than any quantum-specific version. Tier 2.

On barren plateaus

• McClean, Boixo, Smelyanskiy, Babbush, and Neven, "Barren plateaus in quantum neural network training landscapes" (2018), Nature Communications 9, 4812. The original result. Proves the $\mathcal{O}(2^{-n})$ variance decay that §16.6 measures, via a 2-design argument. Read it for the precise statement of the conditions — the theorem's assumptions are where all the proposed mitigations try to escape. Tier 1.
• Cerezo, Sone, Volkoff, Cincio, and Coles, "Cost function dependent barren plateaus in shallow parametrized quantum circuits" (2021), Nature Communications 12, 1791. The local-observable result: global cost functions have barren plateaus even at shallow depth, while local ones do not, for sufficiently shallow circuits. This is why §16.6 measures $Z_0Z_1$ rather than a global product, and it is the most practically actionable mitigation in the literature. Tier 1.
• Grant, Wossnig, Ostaszewski, and Benedetti, "An initialization strategy for addressing barren plateaus in parametrized quantum circuits" (2019), Quantum 3, 214. The identity-block initialization strategy — Exercise 16.25's subject. Tier 1.
• Wang et al., "Noise-induced barren plateaus in variational quantum algorithms" (2021), Nature Communications 12, 6961. A second, independent mechanism: hardware noise flattens the landscape too. Sobering, because unlike the original result this one cannot be escaped by clever ansatz design. Connects directly to Chapter 13. Tier 1.
• Recent work on classical simulability of circuits that avoid barren plateaus. An uncomfortable argument worth knowing: some of the structure that avoids plateaus may also make the circuit classically simulable, which would undercut the point. Tier 2 — active and contested.

On variational algorithms generally

• Cerezo et al., "Variational Quantum Algorithms" (2021), Nature Reviews Physics 3, 625. The comprehensive survey: VQE, QAOA, ansatz design, optimizers, and the obstacles. The single best orientation to the field. Tier 1.
• Bittel and Kliesch, "Training variational quantum algorithms is NP-hard" (2021), Physical Review Letters 127, 120502. A different obstacle from the barren plateau: even with good gradients, the optimization itself is hard in the worst case. Tier 1.
• The literature on classical shadows and randomized measurements. A different route to expectation values with better shot scaling, increasingly relevant given §16.4's cost accounting. Tier 2.

Forward references

• Chapter 24 — VQE built properly, with the mitigations, and an honest revisit of this chapter's measurement.
• Chapter 32 — Part VI, where PennyLane becomes the primary framework.
• Chapter 33 — quantum classifiers, built on everything here.
• Appendix E — PennyLane alongside the other four frameworks.

Where to go next. If one thing: Schuld et al. on analytic gradients. It contains the derivation, the generalization, and the variance argument, and it will make §16.3 feel like a consequence rather than a trick.

If two: add McClean et al. on barren plateaus, and read its assumptions carefully — every proposed escape route is an attempt to violate one of them.

Then Chapter 17 — Amazon Braket, and hardware that is not superconducting, where the gate sets and connectivity assumptions of Part II stop holding.