Benchmarks

All results are wall-clock means over ≥20 runs with warmup, on the same hardware (Apple Silicon M-series). Generated by cd bench && conda run -n qudit python run.py → docs/public/bench.json.

Category	What's measured
Ideal Circuit	GHZ statevector sim, 2–15 qubits, 8 frameworks
Noisy Circuits	Depol / AD / phase-damp — stochastic vs deterministic vs other frameworks
QEC Recovery	Petz / Leung / Cafaro construction + application on 3 codes
Gradient Descent	VQE loop — autograd vs parameter-shift across 6 frameworks
QUBO / QAOA	QUBO conversion, QAOA forward pass, ClockSolver loop
Metrics	Fidelity, entropy, mutual info, trace/Bures distance — scaling with $d$

Charts use log scale by default. qudit series are shown with solid bold lines; other frameworks use dashed lines. Lower is always better.