A quantum generative model's Fisher trace is pinned to its parameter count, so certification shifts to a two-metric escalation rule
Synopsis
For quantum circuits that prepare real-amplitude states from half-angle rotations, the classical Fisher information matrix equals the quantum one off a measure-zero set, so its trace is fixed at the parameter count 2n−1 and says nothing about fit quality. The analysis turns to the Fisher eigenvalue rank and to two distributional metrics, Wasserstein-1 and maximum mean discrepancy: the rank only counts the support of the converged distribution and falls to chance on a full-support benchmark. For two analytically solvable bitstring defect families, the tolerance-mismatch window in which the two metrics disagree opens only at certification-grade tolerances and closes as the qubit count grows.
Interpretation
Whether a trained quantum generative model has learned its target is usually decided by a diagnostic validated on a labeled benchmark; for a real-amplitude quantum circuit Born machine, the Fisher trace is shown to carry no fit-quality information. Fisher spectra and the effective dimensions built on them had been proposed as expressibility and trainability measures, with the trace a candidate certification channel; for circuits with real amplitudes and half-angle rotation generators, the trace equals the parameter count at almost every parameter point. Across 2,520 noiseless statevector training runs, removing the probability floor returns the trace to the parameter count in every run at every qubit count to within 10^-8; the clamped estimator lands within 0.03% to 1.1% of the cap in the mean and within 0.01 of it in 92% to 97% of runs.
The eigenvalue rank of the same Fisher matrix separates perturbed from unperturbed runs perfectly on a sparse benchmark, but it counts the live nodes of the converged distribution's support at the eigenvalue-threshold resolution, not defect size. On the two-peak benchmark an exact fit gives a rank equal to the 2n−1 live nodes of the amplitude tree; once the clean target itself has full support, both classes are full rank and the label information disappears. On the pre-registered perturbed full-support benchmark the rank AUROC is 0.500, 0.500, 0.500, 0.494, 0.525 and 0.494 at n = 3–8, with 597 of 600 fits full rank; on the sparse benchmark the rank AUROC reaches 1.000 through n = 7 and falls to 0.959–0.966 at n = 8.
Certification therefore falls to two distributional metrics, Wasserstein-1 distance and maximum mean discrepancy, composed into an escalation rule: W1 decides first, and an MMD vote is added only when W1 falls inside a window below the threshold. A single-metric rule suppresses false passes only at the rate of one estimation error; with disjoint sample sets the escalation rule turns the joint misvote into a product of two errors, lowering the false-pass bound from about 0.158 to 0.05 at the same per-metric sample size. On two defect families with closed forms, W1 responds linearly to contamination and MMD quadratically; at certification-grade tolerances τMMD = 10^-4 and τW1 = 0.05 the mismatch window is open at n = 3–6 with relative width falling from 0.51 to 0.03, and closed at n ≥ 7.
For every unit-diagonal positive-semidefinite kernel, some parity defect at fixed transport distance has squared discrepancy at most β²/(2n−1), so no kernel is uniformly sensitive to all correlation orders. Kernel discrepancies were already known to weight high-order correlations weakly; what is added is an exact worst-parity minimax value over all unit-diagonal kernels, stationary or not, attained by the centered kernel. Under the Hamming cost and the RBF–Hamming kernel, once τMMD ≥ 2(1−e^-γ)τW1² the discrepancy vote agrees with the transport vote on population values for every pair; on the fixed-threshold panel, the independent discrepancy vote changed no verdict in any of the 41,771 in-window draws at a thousand shots or more.
Perspective
The results rest on noiseless statevector simulation at n = 3–8, with the Möttönen real-amplitude state-preparation ansatz trained by CMA-ES and the rank control replicated under other optimizers only at n ≤ 5. The two defect-family windows are exact closed forms but are stated for the Hamming cost and the RBF–Hamming kernel; the fixed-threshold validation uses 600 perturbed full-support fits whose clean target does not converge within budget at n ≥ 7.
The zero-amplitude locus Z is explicit for the Möttönen ansatz, but how it relates to the parameters the optimizer actually reaches remains open. A distribution-level depolarizing rate of one part in a thousand already breaks the Fisher-trace identity on sparse targets, and no false-pass guarantee has been established under device noise; the escalation's advantage over a single transport threshold did not reproduce on independent targets, and certification-grade tolerances demand more shots than the tested budgets.
