benchmarks/scaling.py still measures only the dense backend — it predates the tensor-network, circuit, Pauli-propagation and JAX engines. With six backends that have genuinely different sweet spots, the single most useful thing the docs could add is a measured comparison.
What to do
Extend the benchmark to sweep, per backend, over qubit count and Hamiltonian structure (chain vs all-to-all, commuting vs non-commuting, high vs low temperature), recording wall-clock, memory and accuracy against the dense reference where it exists.
Then turn the numbers into a decision table in the README: given n qubits, this connectivity, this temperature, and this quantity — use backend X.
Expected findings, to check rather than assume
- tensor network wins on 1-D / low entanglement, loses on all-to-all
- Pauli propagation is topology-agnostic and cheap at high temperature, dense at low
- dense is unbeatable below ~12 qubits
- circuit is for hardware fidelity, not speed
Acceptance
- a reproducible script writing CSV + plots
- the table in the README, with the measurement conditions stated
benchmarks/scaling.pystill measures only the dense backend — it predates the tensor-network, circuit, Pauli-propagation and JAX engines. With six backends that have genuinely different sweet spots, the single most useful thing the docs could add is a measured comparison.What to do
Extend the benchmark to sweep, per backend, over qubit count and Hamiltonian structure (chain vs all-to-all, commuting vs non-commuting, high vs low temperature), recording wall-clock, memory and accuracy against the dense reference where it exists.
Then turn the numbers into a decision table in the README: given n qubits, this connectivity, this temperature, and this quantity — use backend X.
Expected findings, to check rather than assume
Acceptance