q-tensor
Tensor-network claims for noisy quantum simulation conflate unique-record harvesting with real execution speedup, and nobody had measured exactly where batching actually wins.
On the exact public 50-qubit/200-gate Figure-3 workload, tensor-network proportional sampling measured a 6.875× speedup over CUDA-Q TensorNet at equal shots, and a preregistered IBM Kingston study showed the frozen calibration-informed model beating the ideal noiseless model at all four tested depths.
The problem
Q-Tensor started from NVIDIA Research’s open-source PTSBE work and asked three practical questions: are the sampling semantics statistically correct, where does tensor-network batching actually become faster, and can a frozen calibration-informed trajectory model predict real quantum hardware better than an ideal noiseless model? The upstream artifact had a real trap: CUDA-Q reference files are cached by trajectory serial number rather than by the circuit/noise configuration that generated them, so a clean run can pass while an immediate rerun against stale cached reference data fails.
Approach
Q-Tensor separates two sampling semantics that should not be conflated: proportional sampling (trajectory multiplicities preserved, outputs represent physical sampling weights) versus non-proportional unique-record harvesting (deliberately expands distinct labeled outcomes for dataset generation — useful, but not an IID proportional sample and not a raw simulator speedup). It uses categorical Kraus sampling and preserves trajectory multiplicity for its scientific and benchmark claims, with all success criteria preregistered before hardware runs.
Evidence
Corrected proportional sampling passed 48/48 seeded TN/reference, CUDA-Q/reference, and TN/CUDA-Q finite-shot checks. The central finding is a regime split, not a blanket win: small proportional workloads favored CUDA-Q (no crossover through the validated 11-qubit sweep), while the exact 50-qubit/200-gate Figure-3 workload measured 40.419 s vs 277.865 s — a 6.875× equal-shot TN speedup. Separately, a prospectively frozen calibration-informed model had lower total-variation distance to real IBM Kingston hardware than the ideal noiseless model at all four tested depths (3, 6, 9, 12 CZ gates), with the 95% bootstrap interval entirely above zero at three depths; the IBM job used 6 quantum seconds. A corrected exploratory Aer comparison was statistically unresolved (TVDs differed by <0.001 at every depth) and is reported as such.