Introduction
Quantum generative models approximate a target probability distribution by optimizing the parameters of a variational quantum circuit, with applications ranging from quantum-chemistry data generation to financial modeling and combinatorial sampling. Three model families dominate this effort, namely quantum generative adversarial networks (QGANs) [1 , 2 ] , which have been used to learn and load random distributions onto quantum hardware, quantum circuit Born machines (QCBMs) [3 , 4 ] , and quantum generative diffusion models (QGDMs) [5 ] . In our own earlier work we used the QCBM as a data-driven benchmark for training shallow circuits on near-term devices [4 ] , and we built an optimized adversarial framework, Re-QGAN, around a real-amplitude generator circuit [6 ] . As such models move from simulation toward deployment, whether a trained generator has actually learned its target becomes a question of certification, which has to be settled before its samples are used downstream.
In practice, a practitioner certifies a trained generator by scoring its output distribution against the target with a distributional metric, most commonly the Wasserstein-1 distance (W1 _1 1 ) [1 ] or the maximum mean discrepancy (MMD) [7 ] . Channels drawn from the Fisher-information family, whose spectra and effective dimensions were introduced as capacity measures [8 , 9 , 10 ] , are candidate companions to such scores. Each such diagnostic is trusted once it scores well on a labeled benchmark, returning high values for good generators and low values for defective ones. A benchmark, however, is built by varying a known control, typically a perturbation applied to a good target, so that a diagnostic can score well by recovering that control variable instead of the distributional property it claims to measure. Such a diagnostic reports how the benchmark was constructed, or how the generator was trained, and says nothing about whether the output matches the target.
For the Fisher-information channel, whether it reports fit quality is settled by two known facts: on real-amplitude pure states the classical and quantum Fisher information coincide, a result due to Facchi et al. [11 ] , and a half-angle rotation with an involutive generator contributes a unit entry to the quantum Fisher diagonal [12 ] . We apply them to the universal real-amplitude state-preparation ansatz of Möttönen et al. [13 , 14 ] , used by the QCBM throughout this work and as the Re-QGAN generator [6 ] : the classical Fisher information matrix of the computational-basis distribution coincides with the quantum Fisher information matrix, and its trace equals the parameter count d = 2 n − 1 d = 2^n - 1 d = 2 n − 1 , at every parameter point outside a measure-zero set on which some basis amplitude vanishes (Proposition 1). The identity extends to all half-angle real-amplitude circuits, including the layered hardware-efficient family. As a consequence, the Fisher trace is constant on the operational parameter space, so that it carries no information about fit quality, and any certification protocol that uses it reduces to its remaining channels.
In this work, we take this identity as the starting observation and ask which channels can still certify a real-amplitude QCBM. We first specialize the identity to the deployed ansatz class, describe the exceptional set explicitly for the Möttönen circuit, and validate the identity on our training campaign. We then test the two remaining candidates, one spectral and one distributional, beginning with the eigenvalue rank of the same Fisher matrix, whose capacity interpretation has been developed for quantum neural networks and for QCBMs [10 , 15 , 16 ] , and which separates perturbed from unperturbed runs perfectly on the principal benchmark through n = 7 n = 7 n = 7 . The rank, however, counts the live nodes of the amplitude tree at the resolution of its eigenvalue threshold, which is the rank of the quantum Fisher matrix; the support-restricted classical matrix would have rank at most one. On the two-peak benchmark it equals 2 n − 1 2^n - 1 2 n − 1 whenever the fit is exact. A severity split, a pre-registered stress test on full-support defect families and a pre-registered perturbed full-support benchmark collapse it to chance, and the pattern recurs under three further optimizers. We keep it only as a flag for anisotropic verdict robustness, a role that a direct perturbation test supports. In turn, the pooled W 1 W_1 W 1 and MMD scores move with optimizer convergence and with the reference distribution against which they are computed; on the sparse benchmark they become perfect once the negative class is conditioned on convergence, and on the full-support benchmark they resolve a defect to the extent that it exceeds the fit residual. For the distributional channels that remain, we analyze an escalation rule that consults MMD only when W 1 W_1 W 1 passes narrowly, and we prove a finite-sample bound on its false-pass probability that improves quadratically on the single-metric rule at the same threshold for defects near it. When it does consult MMD, the rule has a higher false-fail rate than the single-metric rule. For two families of bitstring defects on which both W 1 W_1 W 1 and MMD have closed forms, we compute exactly the tolerance-mismatch region, the set of generators that the transport threshold admits and the discrepancy threshold rejects, together with the window of scales and tolerance ratios in which it is non-empty. The window requires certification-grade tolerances and narrows as the qubit number increases. Inside it the escalation acts on both families as a single shifted transport threshold, and the two families close it
at tolerance ratios that differ by a fixed factor. Against an unknown defect family at a fixed operating point it remains open only at the smallest qubit number we study. On the Hamming cost we prove two further exact limits of any such protocol. The discrepancy is redundant for every pair of distributions once its tolerance exceeds 2 ( 1 − e − γ ) τ W 1 2 2(1 - e^{-\gamma})\tau_{W_1}^2 2 ( 1 − e − γ ) τ W 1 2 . For every normalized kernel, some parity defect of fixed transport size has squared discrepancy at most β 2 / ( 2 n − 1 ) \beta^2/(2^n - 1) β 2 / ( 2 n − 1 ) , so that no fixed kernel is uniformly sensitive to correlation defects and certification must name the correlation orders it is meant to detect. That kernel discrepancies weight high-order correlations weakly is known. Rudolph et al. [17 ] draw out its consequence for training gradients, and Kurkin et al. [18 ] exhibit pairs at unit total variation whose discrepancy is exponentially small for every stationary characteristic kernel with unit diagonal on the bitstring space. Our parity statement covers every unit-diagonal positive-semidefinite kernel, stationary or not, takes an explicit parity family at fixed transport distance as its alternatives, and gives an exact value with an attaining kernel (Sec. V F). In this way the present work complements the empirical multi-metric advocacy of Gili et al. [19 ] and the alternative-loss exploration of Coyle et al. [20 ] with a structural account of why the trace channel would fail and with finite-sample bounds and design-recovery controls for the protocol that replaces it. Section II fixes notation, the ansatz, the decision rules and the experimental configuration. We prove and validate the Fisher-trace identity in Sec. III, test the rank and the pooled distributional scores in Sec. IV, and define, bound and test the escalation protocol in Sec. V. Section VI discusses alternatives, noise and limitations, and Sec. VII concludes. The proofs and the concentration machinery are collected in Appendices A and B.
Background and experimental design
We consider quantum generative models whose output is a probability distribution q ∈ Δ N q \in\Delta_N q ∈ Δ N , the probability simplex over the outcome space X = { 0 , 1 } n \mathcal{X} = \{0,1\}^n X = { 0 , 1 } n with N = 2 n N = 2^n N = 2 n elements. The distribution is obtained by measuring an n n n -qubit parameterized quantum circuit (PQC) in the computational basis. The target distribution p ∈ Δ N p \in\Delta_N p ∈ Δ N is either known analytically or estimated from samples. The certification question is then: given p p p and q q q , is q q q an acceptable approximation of p p p ?
Wasserstein-1 distance
The Wasserstein-1 (earth mover’s) distance between distributions p p p and q q q on X \mathcal{X} X , equipped with a ground-cost matrix C ∈ R ≥ 0 N × N C \in\mathbb{R}_{\ge0}^{N \times N} C ∈ R ≥ 0 N × N , is defined as the optimal transport cost:
W 1 ( p , q ; C ) = min Γ ≥ 0 ⟨ C , Γ ⟩ s.t. Γ 1 = p , Γ T 1 = q (1) W_1(p,q;C) = \min_{\Gamma\ge0} \langle C,\Gamma\rangle\quad\text{s.t.}\quad\Gamma\mathbf{1}=p,\quad\Gamma^{\mathsf{T}}\mathbf{1}=q \tag*{(1)} W 1 ( p , q ; C ) = Γ ≥ 0 min ⟨ C , Γ ⟩ s.t. Γ 1 = p , Γ T 1 = q (1)
where Γ \Gamma Γ is the transport plan (coupling), 1 \mathbf{1} 1 is the all-ones vector so that the two constraints fix the row and column marginals of Γ \Gamma Γ to p p p and q q q , and ⟨ ⋅ , ⋅ ⟩ \langle\cdot,\cdot\rangle ⟨ ⋅ , ⋅ ⟩ denotes the Frobenius inner product. Equivalently, by Kantorovich duality,
W 1 ( p , q ; C ) = max f ∑ x ∈ X f ( x ) ( p ( x ) − q ( x ) ) , (2) W_1(p,q;C)=\max_f\sum_{x\in\mathcal{X}}f(x)(p(x)-q(x)), \tag*{(2)} W 1 ( p , q ; C ) = f max x ∈ X ∑ f ( x ) ( p ( x ) − q ( x )) , (2)
where the maximum is over C C C -Lipschitz potentials: f ( x ) − f ( y ) ≤ C ( x , y ) f(x)-f(y)\le C(x,y) f ( x ) − f ( y ) ≤ C ( x , y ) for all x , y x,y x , y . We denote throughout by TV ( p , q ) = 1 2 ∥ p − q ∥ 1 = 1 2 ∑ x ∣ p ( x ) − q ( x ) ∣ \operatorname{TV}(p,q)=\frac{1}{2}\lVert p-q\rVert_1=\frac{1}{2}\sum_x\lvert p(x)-q(x)\rvert TV ( p , q ) = 2 1 ∥ p − q ∥ 1 = 2 1 ∑ x ∣ p ( x ) − q ( x )∣ the total-variation distance, which enters the closed forms of § 4 D.
The ground cost determines which distributional differences W 1 W_1 W 1 sees, and we use three: the Hamming cost C 1 ( x , y ) = d H ( x , y ) C_1(x,y)=d_H(x,y) C 1 ( x , y ) = d H ( x , y ) ; a weighted variant C 2 ( x , y ) = ∑ i w i ∣ x i − y i ∣ C_2(x,y)=\sum_i w_i\lvert x_i-y_i\rvert C 2 ( x , y ) = ∑ i w i ∣ x i − y i ∣ with qubit weights w i > 0 w_i>0 w i > 0 ; and a circuit-structure cost C 3 ( x , y ) = d H ( x , y ) + α W ∑ i ∣ χ i ( x ) − χ i ( y ) ∣ C_3(x,y)=d_H(x,y)+\alpha_W\sum_i\lvert\chi_i(x)-\chi_i(y)\rvert C 3 ( x , y ) = d H ( x , y ) + α W ∑ i ∣ χ i ( x ) − χ i ( y )∣ with nearest-neighbor parities χ i ( x ) = x i ⊕ x i + 1 \chi_i(x)=x_i\oplus x_{i+1} χ i ( x ) = x i ⊕ x i + 1 . All three are metrics (C3 adds parity pseudometrics to the Hamming metric), so the duality (2) and the centering argument of Proposition 6 apply to each. The maximum cost D max = max x , y C ( x , y ) D_{\max}=\max_{x,y}C(x,y) D m a x = max x , y C ( x , y ) determines the scale of W 1 W_1 W 1 and enters the sample-complexity bounds (§ 4 ).
Maximum mean discrepancy
The maximum mean discrepancy (MMD) between p p p and q q q , with respect to a positive-definite kernel k : X × X → R k:\mathcal{X}\times\mathcal{X}\to\mathbb{R} k : X × X → R , is defined according to [7 ]
MMD k 2 ( p , q ) = E x , x ′ ∼ p [ k ( x , x ′ ) ] + E y , y ′ ∼ q [ k ( y , y ′ ) ] − 2 E x ∼ p , y ∼ q [ k ( x , y ) ] . (3) \operatorname{MMD}_k^2(p,q)=\mathbb{E}_{x,x'\sim p}[k(x,x')]+\mathbb{E}_{y,y'\sim q}[k(y,y')]-2\mathbb{E}_{x\sim p,y\sim q}[k(x,y)]. \tag*{(3)} MMD k 2 ( p , q ) = E x , x ′ ∼ p [ k ( x , x ′ )] + E y , y ′ ∼ q [ k ( y , y ′ )] − 2 E x ∼ p , y ∼ q [ k ( x , y )] . (3)
Equivalently, MMD k ( p , q ) = ∥ μ p − μ q ∥ H k \operatorname{MMD}_k(p,q)=\lVert\mu_p-\mu_q\rVert_{\mathcal{H}_k} MMD k ( p , q ) = ∥ μ p − μ q ∥ H k , where μ p = E x ∼ p [ k ( x , ⋅ ) ] \mu_p=\mathbb{E}_{x\sim p}[k(x,\cdot)] μ p = E x ∼ p [ k ( x , ⋅ )] is the kernel mean embedding in the reproducing kernel Hilbert space (RKHS) H k \mathcal{H}_k H k .
We use the radial basis function (RBF) kernel on Hamming distance:
k ( x , y ) = exp ( − γ d H ( x , y ) ) , γ = 1 / n . (4) k(x,y)=\exp(-\gamma d_H(x,y)),\qquad\gamma=1/n. \tag*{(4)} k ( x , y ) = exp ( − γ d H ( x , y )) , γ = 1/ n . (4)
For a characteristic kernel such as the RBF, MMD k ( p , q ) = 0 \operatorname{MMD}_k(p,q)=0 MMD k ( p , q ) = 0 if and only if p = q p=q p = q [7 ] . The kernel bandwidth γ = 1 / n \gamma=1/n γ = 1/ n is chosen so that the kernel has non-trivial sensitivity across the full range of Hamming distances at each qubit count.
In the statevector campaign MMD 2 \operatorname{MMD}^2 MMD 2 is evaluated exactly from the output distributions (Sec. II H). The finite-shot study of § 4 J estimates it from samples with the unbiased U U U -statistic of Proposition 8 and reports the plug-in (V V V -statistic) estimator on the same draws for comparison; the shot-budget study of Appendix B § 5.8.4 quantifies the difference between the two. In every setting the value is thresholded against a tolerance τ M M D \tau_{\mathrm{MMD}} τ MMD . Throughout this paper τ M M D \tau_{\mathrm{MMD}} τ MMD acts on the squared statistic, as in Algorithm 1: the loose benchmark-comparison tolerance is MMD k 2 ≤ 10 − 2 \operatorname{MMD}_k^2\le10^{-2} MMD k 2 ≤ 1 0 − 2 and the certification-grade tolerance is MMD k 2 ≤ 10 − 4 \operatorname{MMD}_k^2\le10^{-4} MMD k 2 ≤ 1 0 − 4 , equivalently MMD k ≤ 0.1 \operatorname{MMD}_k\le0.1 MMD k ≤ 0.1 and MMD k ≤ 0.01 \operatorname{MMD}_k\le0.01 MMD k ≤ 0.01 . The labels “loose” and “certification-grade” refer to τ M M D \tau_{\mathrm{MMD}} τ MMD alone; the transport threshold is τ W 1 = 0.05 \tau_{W_1}=0.05 τ W 1 = 0.05 in every analysis. We do not use a permutation test, the customary two-sample tool, because its p p p -value tests the sharp null p = q p=q p = q . Certification asks whether the discrepancy exceeds an operational tolerance, and a generator can differ from its target at any significance level while sitting well inside any tolerance a practitioner would set. The false-pass guarantee of § 4 B therefore covers the threshold comparison.
Classical Fisher-trace diagnostic
The classical Fisher trace measures the generator’s sensitivity to parameter perturbations at the current operating point. For a PQC with parameters θ ∈ R d \theta\in\mathbb{R}^d θ ∈ R d producing output distribution p θ p_\theta p θ over X \mathcal{X} X , the classical Fisher information matrix (CFIM) is
F j k ( θ ) = ∑ x ∈ X 1 p θ ( x ) ∂ p θ ( x ) ∂ θ j ∂ p θ ( x ) ∂ θ k . (5) F_{jk}(\theta)=\sum_{x\in\mathcal{X}}\frac{1}{p_\theta(x)}\frac{\partial p_\theta(x)}{\partial\theta_j}\frac{\partial p_\theta(x)}{\partial\theta_k}. \tag*{(5)} F j k ( θ ) = x ∈ X ∑ p θ ( x ) 1 ∂ θ j ∂ p θ ( x ) ∂ θ k ∂ p θ ( x ) . (5)
The diagnostic is the trace of the CFIM:
CFI w ( θ ) = Tr ( F ( θ ) ) = ∑ k = 1 d ∑ x ∈ X 1 p θ ( x ) ( ∂ p θ ( x ) ∂ θ k ) 2 . (6) \operatorname{CFI}_w(\theta)=\operatorname{Tr}(F(\theta))=\sum_{k=1}^{d}\sum_{x\in\mathcal{X}}\frac{1}{p_\theta(x)}\left(\frac{\partial p_\theta(x)}{\partial\theta_k}\right)^2. \tag*{(6)} CFI w ( θ ) = Tr ( F ( θ )) = k = 1 ∑ d x ∈ X ∑ p θ ( x ) 1 ( ∂ θ k ∂ p θ ( x ) ) 2 . (6)
Here, the CFIM in Eq. (5) is dominated by the Löwner (positive-semidefinite, PSD) partial order by the quantum Fisher information matrix (QFIM) computed from the full density operator [21 , 22 ] , an inequality we exploit in § 2 to derive an analytical cap on CFI w \operatorname{CFI}_w CFI w . In the noiseless statevector setting, the Jacobians ∂ p θ / ∂ θ k \partial p_\theta/\partial\theta_k ∂ p θ / ∂ θ k are computed exactly by the parameter-shift rule (Sec. II H). The Fisher trace is a property of the generator at its current parameters. Fisher spectra and the effective dimensions built on them have been proposed as expressibility and trainability measures [8 , 9 , 10 ] ; whether the trace can be a certification channel alongside the distributional ones is what § 2 tests on the Möttönen ansatz used in this work.
The Möttönen real-amplitude state-preparation ansatz
We adopt the universal real-amplitude state-preparation circuit of Möttönen et al. [13 , 14 ] , both as the QCBM generator and as the working example for the Fisher-trace identity of § 2 . It is also the generator architecture of our earlier work on quantum generative adversarial networks (Re-QGAN) [6 ] . The circuit is recursive: a length-n n n block is constructed from two length-( n − 1 ) (n-1) ( n − 1 ) blocks joined by a single CNOT between the first and last qubit, with the base case providing one or two R Y R_Y R Y rotations on a single or two-qubit block. Applying such blocks in sequence on the first k k k qubits, k = 1 , … , n k = 1, \ldots, n k = 1 , … , n , gives the full circuit, whose stage k k k is the uniformly controlled rotation of Appendix A 2. The parameter count satisfies the recurrence T ( n ) = 2 T ( n − 1 ) + 1 T(n) = 2T(n - 1) + 1 T ( n ) = 2 T ( n − 1 ) + 1 with T ( 1 ) = 1 T(1) = 1 T ( 1 ) = 1 , yielding exactly d = 2 n − 1 d = 2^n - 1 d = 2 n − 1 rotation parameters and 2 n − n − 1 2^n - n - 1 2 n − n − 1 CNOT gates for an n n n -qubit ansatz. This count matches the dimension of the real-amplitude pure-state sphere S R 2 n − 1 ⊂ R 2 n S_{\mathbb{R}}^{2^n-1} \subset\mathbb{R}^{2^n} S R 2 n − 1 ⊂ R 2 n , so the ansatz is universal for real-amplitude states with the minimal parameter count compatible with universality.
The identity of § 2 uses three properties of this construction. First, the gate set { R Y , C N O T } \{R_Y, \mathrm{CNOT}\} { R Y , CNOT } has real matrix entries in the computational basis, and the initial state ∣ 0 ⟩ ⊗ n \lvert0\rangle^{\otimes n} ∣ 0 ⟩ ⊗ n is real, so the prepared state ∣ ψ ( θ ) ⟩ \lvert\psi(\theta)\rangle ∣ ψ ( θ )⟩ is real for every θ \theta θ . Second, the parameter count matches the real-amplitude manifold dimension exactly, so that the parameterization is minimal and no parameter is redundant on a generic open set. Third, each rotation R Y ( θ k ) = exp ( − i θ k Y / 2 ) R_Y(\theta_k) = \exp(-\mathrm{i}\theta_kY/2) R Y ( θ k ) = exp ( − i θ k Y /2 ) enters with the half-angle convention and a Hermitian-unitary generator Y Y Y , which together force the diagonal entries of the quantum Fisher matrix to equal one identically. As we show in § 2 , reality and half-angle structure suffice for Proposition 1 ; minimality only fixes the parameter count at 2 n − 1 2^n - 1 2 n − 1 .
Certification as a decision problem
We frame certification as a hypothesis test: given distributions p p p (target) and q q q (generator output), the certifier must decide PASS or FAIL. A false pass (type II error) occurs when a defective generator is declared certified; a false fail (type I error) occurs when a good generator is rejected. The relative severity of the two errors depends on the application; in safety-critical or scientific workflows, false passes are typically the more costly.
Given thresholds τ W 1 \tau_{\mathrm{W1}} τ W1 and τ M M D \tau_{\mathrm{MMD}} τ MMD , we consider four decision rules: the W1-only baseline D 1 D_1 D 1 (PASS iff W 1 ≤ τ W 1 W_1 \le\tau_{\mathrm{W1}} W 1 ≤ τ W1 ), the OR rule D 2 D_2 D 2 (either metric passes), the conservative AND rule D 3 D_3 D 3 (both must pass), and the escalation rule D 4 D_4 D 4 . The last of these uses W 1 W_1 W 1 as the primary decision channel and applies the AND rule only when the W 1 W_1 W 1 value falls within the escalation window τ W 1 − η < W 1 ≤ τ W 1 \tau_{\mathrm{W1}} - \eta< W_1 \le\tau_{\mathrm{W1}} τ W1 − η < W 1 ≤ τ W1 on the PASS side of the threshold. In this way, the rule keeps the sample efficiency of D 1 D_1 D 1 in the clear case, because the MMD sample set is drawn only if the window is entered, and the false-pass guarantee of D 3 D_3 D 3 in the ambiguous case, at the cost of a second sample set on escalated runs. Rule D 4 D_4 D 4 is the protocol we analyze; its advantage over D 1 D_1 D 1 is confined to the tolerances and scales established in § 4 , and the Fisher trace enters no decision rule (§ 2 ). Table I collects the thresholds, margins and reference conventions used from here on.
Table I. Symbols used in the decision analysis.
Generative model and training configuration
The QCBM output distribution, built from the Möttönen ansatz U ( θ ) U(\theta) U ( θ ) of § 2 , follows the Born rule in the computational basis:
q θ ( x ) = ∣ ⟨ x ∣ U ( θ ) ∣ 0 ⟩ ⊗ n ∣ 2 , x ∈ { 0 , 1 } n . (7) q_\theta(x) = \lvert\langle x\lvert U(\theta)\lvert0\rangle^{\otimes n}\rvert^2,\quad x \in\{0,1\}^n . \tag*{(7)} q θ ( x ) = ∣⟨ x ∣ U ( θ ) ∣ 0 ⟩ ⊗ n ∣ 2 , x ∈ { 0 , 1 } n . (7)
We train by minimizing W 1 ( p , q θ ; C ) W_1(p,q_\theta;\mathcal{C}) W 1 ( p , q θ ; C ) with CMA-ES [23 ] , a derivative-free optimizer well suited to the nonconvex objectives of PQC optimization. CMA-ES also sidesteps the gradient estimation whose shot cost can plateau render prohibitive, and provides robust convergence with explicit function-evaluation budgets; it does not escape the underlying landscape flatness itself, which afflicts gradient-free optimizers as well [24 ] .
Benchmark suite
We define three benchmark families. Benchmark B1 is a GHZ-type synthetic target, referred to throughout as the convex-blend family . The clean target is the two-peak Born distribution p 0 = 0.6 δ x , 0 n + 0.4 δ x , 1 n p_0 = 0.6\delta_{x,0^n} + 0.4\delta_{x,1^n} p 0 = 0.6 δ x , 0 n + 0.4 δ x , 1 n of the state 0.6 ∣ 0 ⟩ ⊗ n + 0.4 ∣ 1 ⟩ ⊗ n \sqrt{0.6}\lvert0\rangle^{\otimes n} + \sqrt{0.4}\lvert1\rangle^{\otimes n} 0.6 ∣ 0 ⟩ ⊗ n + 0.4 ∣ 1 ⟩ ⊗ n , and the perturbed targets are the distribution-level convex blends
p ϵ = ( 1 − ϵ ) p 0 + ϵ ν , ϵ ∈ { 0.00 , 0.02 , 0.05 , 0.10 , 0.20 } , (8) p_{\epsilon} = (1-\epsilon)\,p_0 + \epsilon\,\nu,\qquad\epsilon\in\{0.00, 0.02, 0.05, 0.10, 0.20\}, \tag*{(8)} p ϵ = ( 1 − ϵ ) p 0 + ϵ ν , ϵ ∈ { 0.00 , 0.02 , 0.05 , 0.10 , 0.20 } , (8)
where ν \nu ν is a random background distribution drawn once per seed (i.i.d. uniform weights on { 0 , 1 } n \{0,1\}^n { 0 , 1 } n , normalized). The blend spreads support over all 2 n 2^n 2 n outcomes while keeping the two peaks dominant, and the perturbation level ϵ \epsilon ϵ supplies ground-truth quality labels for ROC analysis. Benchmark B2 is a quantum-circuit-generated target produced by a random Clifford+T T T circuit, a structured but non-trivial distribution whose Shannon entropy grows with n n n while staying well below the uniform maximum (Table II). Finally, benchmark B3 is a high-entropy synthetic target on { 0 , 1 } n \{0,1\}^{n} { 0 , 1 } n , near-uniform with Shannon entropy just below the maximum n ln 2 n \ln2 n ln 2 (Table II; per-outcome deviations from uniform of at most a few parts in 10 3 10^{3} 1 0 3 ), which stress-tests certification under targets with broad support. B3 is the only benchmark not produced by a circuit, and we flag the conclusions that rest on it alone where they are drawn (Sec. VI).
Table 2. Shannon entropy H H H (nats) of the quantum-circuit target B2 and the near-uniform target B3 across qubit counts, with the uniform maximum n ln 2 n \ln2 n ln 2 for reference. B2 stays well below maximal. B3 agrees with n ln 2 n \ln2 n ln 2 to within rounding, so the last column gives its deficit n ln 2 − H B 3 n \ln2 - H_{\mathrm{B3}} n ln 2 − H B3 in units of 10 − 3 10^{-3} 1 0 − 3 nats.
Experimental design
The full campaign comprises 2,520 experiments (420 per qubit count, six qubit counts), structured around the three benchmarks B1/B2/B3 and three ground-cost functions C1/C2/C3 introduced above. For B1, each cost contributes five perturbation levels times 20 seeds, giving 100 runs per cost and 300 runs per qubit count; for B2 and B3, each cost contributes 20 seeds, giving 60 runs per cost and 120 runs per qubit count when combined. All experiments use noiseless statevector simulation (default.qubit backend in PennyLane [25 ] ). CMA-ES function-evaluation budgets scale with qubit count: 5,000 (n = 3 n = 3 n = 3 ), 8,000 (n = 4 n = 4 n = 4 ), 12,000 (n = 5 n = 5 n = 5 ), 20,000 (n = 6 n = 6 n = 6 ), 30,000 (n = 7 n = 7 n = 7 ), and 50,000 (n = 8 n = 8 n = 8 ); all six scales follow the same benchmark, cost, and perturbation structure and appear in Tables III, Table VI , and Table VII . In the campaign MMD is computed exactly from the statevector output distributions, with no shot noise; sampled estimators enter only in § 4 J and Appendix B § 5.8.4 . The Fisher trace is computed with the Jacobian J = ∂ p / ∂ θ J = \partial p/\partial\theta J = ∂ p / ∂ θ evaluated exactly by the parameter-shift rule ∂ p / ∂ θ k = 1 2 [ p ( θ + π 2 e k ) − p ( θ − π 2 e k ) ] \partial p/\partial\theta_k = \frac{1}{2}[p(\theta+ \frac{\pi}{2}e_k) - p(\theta- \frac{\pi}{2}e_k)] ∂ p / ∂ θ k = 2 1 [ p ( θ + 2 π e k ) − p ( θ − 2 π e k )] , which is available because each variational angle enters a single half-angle generator (condition (cond2) of § 2 ) and carries no step size and no O ( h 2 ) O(h^2) O ( h 2 ) error. The amplitude Jacobian of Sec. IV …
Table VI. TABLE VI. W 1 ( C 1 ) W_1(C_1) W 1 ( C 1 ) on B1 against both references (mean and 95% CI over 20 seeds), abridged to ϵ ∈ { 0 , 0.05 , 0.20 } \epsilon\in\{0, 0.05, 0.20\} ϵ ∈ { 0 , 0.05 , 0.20 } ; the full response is in Figure 2 . At ϵ = 0 \epsilon= 0 ϵ = 0 the two references coincide and the columns are identical.
Table VII. AUROC on B1 perturbation labels (100 runs per cell) for the rank diagnostic, W1 _1 1 and MMD under both references, and the debiased Sinkhorn divergence (σ = 0.5 \sigma= 0.5 σ = 0.5 ), with seed-cluster bootstrap 95% CIs over 2,000 resamples of the 20 seeds of a cell. Metrics within a row are compared through the paired differences of Sec. IV B. The rank uses no reference and its intervals are degenerate wherever the classes are gap-separated.
Sec. IV A uses the same shift on the real statevector, ∂ ψ / ∂ θ k = [ ψ ( θ + π 2 e k ) − ψ ( θ − π 2 e k ) ] / ( 2 2 ) \partial\psi/\partial\theta_k = [\psi(\theta+ \frac{\pi}{2}e_k) - \psi(\theta- \frac{\pi}{2}e_k)]/(2\sqrt{2}) ∂ ψ / ∂ θ k = [ ψ ( θ + 2 π e k ) − ψ ( θ − 2 π e k )] / ( 2 2 ) , the coefficient differs because here the shift acts on a half-angle rotation of the state, while the probability rule shifts an expectation value. The inverse-probability weight in F c l = J T diag ( 1 / p ) J F_{\mathrm{cl}} = J^{\mathsf{T}}\operatorname{diag}(1/p)J F cl = J T diag ( 1/ p ) J is clamped at a probability floor of 10 − 10 10^{-10} 1 0 − 10 to regularize rare near-zero outcomes, a regularization distinct from Jacobian accuracy (§ 2 C). On the converged parameters the parameter-shift and a central-difference (h = 10 − 4 h = 10^{-4} h = 1 0 − 4 ) Jacobian reproduce the trace and rank to numerical precision.
Two reference distributions can be used to score a run. The trained-on reference is the target the run was optimized against, which is the perturbed target p ϵ p_{\epsilon} p ϵ whenever ϵ > 0 \epsilon> 0 ϵ > 0 ; a score against it is a training residual and measures fit quality. In contrast, the clean reference is the unperturbed target p 0 p_0 p 0 , against which the benchmark’s notion of a defect is defined, so a score against it measures defect magnitude. The two coincide only at ϵ = 0 \epsilon= 0 ϵ = 0 , and they can disagree sharply: a perturbed run that is fitted perfectly has a training residual of zero while carrying the largest defect the benchmark injects. Unless a table states otherwise, we report both. Neither alone answers the certification question: the trained-on column shows what the optimizer achieved, the clean column shows what a certifier would see in deployment, where the target claimed for the generator is p 0 p_0 p 0 . All 2,520 runs completed with zero execution failures.2 A further 420-run batch at n = 9 n = 9 n = 9 is excluded from every analysis: at that scale CMA-ES fits neither the unperturbed nor the perturbed targets within any budget we tried, so the batch only locates a trainability frontier.
Every table and figure below reports the campaign as it was run, with two exceptions stated where they occur (Table XIII , and three of the nine campaign fits reused as finite-shot pairs in § 4 J); a later warm restart of a subset of runs is reported as a robustness check in Sec. IV B. Statistical analysis uses bootstrap resampling for 95% confidence intervals on all reported quantities, with 5,000–10,000 resamples for means and 2,000 seed-cluster resamples for AUROC cells, as described below. The area under the receiver operating characteristic (AUROC) is the probability that a diagnostic scores a randomly chosen perturbed run as worse than a randomly chosen unperturbed one; we compute it with the perturbation level as a proxy quality label, ϵ = 0 \epsilon= 0 ϵ = 0 (good) against ϵ > 0 \epsilon> 0 ϵ > 0 (degraded). The 20 seeds of a cell are shared across the five perturbation levels, since the background ν \nu ν of seed i i i is the same at every ϵ \epsilon ϵ , and across the three ground costs, so runs are not independent across levels or costs. Every AUROC interval is therefore a seed-cluster bootstrap, which resamples the 20 seed indices and carries all levels and costs of a seed together. The run-level stratified bootstrap, which treats the 100 runs of a cell as independent, moves no endpoint by more than 0.04 and changes the median interval width by 3%. Two diagnostics or two costs are compared through the paired difference of their AUROCs on the same runs or the same seeds, with its own seed-cluster interval. A difference whose interval covers zero is reported as a non-detection, and the precision of every such comparison is set by its 20 unperturbed runs. Two experiments were specified before execution in dated documents in the public repository (see Data availability). The first is the adversarial cross-family stress test of Sec. IV A: its hypotheses, its defect families (correlation shift and mode collapse), its design (10 seeds per family, qubit count, and level) and its pass threshold, a rank AUROC ≥ 0.90 \ge0.90 ≥ 0.90 , were fixed before the runs. The second, specified after the campaign analysis, is the perturbed full-support benchmark of Sec. IV C, with three predictions fixed before launch and its extension to n = 7 , 8 n=7,8 n = 7 , 8 fixed in a dated addendum. Three further controls, the convergence conditioning, the severity split, and the Powell and Adam replications, were specified after the campaign had been seen and before the analysis they support, and were run once without tuning. The eigenvalue-threshold and kernel-bandwidth sweeps are post hoc. In the next section, we test the Fisher trace, the first of the candidate certification channels.
Table XIII. Classical Fisher trace relative to the parameter count under a depolarizing channel of rate p d e p = 10 − 3 p_{\mathrm{dep}}=10^{-3} p dep = 1 0 − 3 , obtained by post-processing the parameter-shift Jacobians of the converged unperturbed C 1 C_1 C 1 fits, evaluated at the warm-restarted parameters of Sec. IV B; this is the one table not computed on the campaign as run.
The Fisher-trace identity for real-amplitude circuits
In this section we specialize known results to the deployed ansatz family and show that the trace is pinned to the parameter count, first in a structural form and then for the Möttönen circuit [13 , 14 ] used by the QCBM of Sec. II and as the Re-QGAN generator. The ingredients are standard, since Facchi et al. [11 ] established the classical–quantum Fisher equality for real-amplitude pure states, and the Cramér–Rao matrix inequality and the unit quantum-Fisher diagonal of half-angle involutive generators are textbook results [12 , 26 , 27 ] .
The structural identity
Two conditions suffice for the identity, real amplitudes and half-angle generators. A third, minimality, only fixes the parameter count at 2 n − 1 2^n-1 2 n − 1 .
Proposition 1 (Structural Fisher-trace identity). Let ∣ ψ ( θ ) ⟩ = U ( θ ) ∣ 0 ⟩ ⊗ n |\psi(\boldsymbol{\theta})\rangle=U(\boldsymbol{\theta})|0\rangle^{\otimes n} ∣ ψ ( θ )⟩ = U ( θ ) ∣0 ⟩ ⊗ n be the output state of an n n n -qubit parameterized quantum circuit with parameters θ ∈ R d \boldsymbol{\theta}\in\mathbb{R}^d θ ∈ R d , and let p θ ( x ) = ∣ ⟨ x ∣ ψ ( θ ) ⟩ ∣ 2 p_{\boldsymbol{\theta}}(x)=|\langle x|\psi(\boldsymbol{\theta})\rangle|^2 p θ ( x ) = ∣ ⟨ x ∣ ψ ( θ )⟩ ∣ 2 be the computational-basis probability distribution. Let F C l ( θ ) F_{\mathrm{Cl}}(\boldsymbol{\theta}) F Cl ( θ ) be the classical Fisher information matrix of p θ p_{\boldsymbol{\theta}} p θ , Eq. (5) with the sum running over outcomes with p θ ( x ) > 0 p_{\boldsymbol{\theta}}(x)>0 p θ ( x ) > 0 , and F Q ( θ ) F_{\mathrm{Q}}(\boldsymbol{\theta}) F Q ( θ ) the quantum Fisher information matrix of ∣ ψ ( θ ) ⟩ |\psi(\boldsymbol{\theta})\rangle ∣ ψ ( θ )⟩ . Suppose the circuit satisfies the following two conditions:
(cond1 _1 1 ) (Real amplitudes.) The state ∣ ψ ( θ ) ⟩ |\psi(\boldsymbol{\theta})\rangle ∣ ψ ( θ )⟩ has real amplitudes in the computational basis for every θ \boldsymbol{\theta} θ .
(cond2 _2 2 ) (Half-angle rotation generators.) Each parameter θ k \theta_k θ k enters U ( θ ) U(\boldsymbol{\theta}) U ( θ ) exactly once, as a half-angle rotation R G k ( θ k ) = exp ( − i θ k G k / 2 ) R_{G_k}(\theta_k)=\exp(-i\theta_kG_k/2) R G k ( θ k ) = exp ( − i θ k G k /2 ) , where G k G_k G k is a Hermitian unitary acting on some subset of qubits, i.e. an involution G k 2 = I G_k^2=I G k 2 = I (canonically a Pauli string, as for single-qubit Y Y Y rotations of the Möttönen ansatz).
Then the following two statements hold for the Fisher matrices of the circuit.
(a) For every θ ∈ R d \boldsymbol{\theta}\in\mathbb{R}^d θ ∈ R d , Tr F Q ( θ ) = d \operatorname{Tr}F_{\mathrm{Q}}(\boldsymbol{\theta})=d Tr F Q ( θ ) = d and
Tr F C l ( θ ) ≤ d . (9) \operatorname{Tr}F_{\mathrm{Cl}}(\boldsymbol{\theta})\le d. \tag*{(9)} Tr F Cl ( θ ) ≤ d . (9) (b) Let ψ θ ( x ) = ⟨ x ∣ ψ ( θ ) ⟩ \psi_{\boldsymbol{\theta}}(x)=\langle x|\psi(\boldsymbol{\theta})\rangle ψ θ ( x ) = ⟨ x ∣ ψ ( θ )⟩ and let Z = { θ : ψ θ ( x ) = 0 and ∂ k ψ θ ( x ) ≠ 0 for some x , k } \mathcal{Z}=\{\boldsymbol{\theta}:\psi_{\boldsymbol{\theta}}(x)=0\text{ and }\partial_k\psi_{\boldsymbol{\theta}}(x)\ne0\text{ for some }x,k\} Z = { θ : ψ θ ( x ) = 0 and ∂ k ψ θ ( x ) = 0 for some x , k } be the zero-amplitude set. For every θ ∉ Z \boldsymbol{\theta}\notin\mathcal{Z} θ ∈ / Z ,
F C l ( θ ) = F Q ( θ ) , Tr F C l ( θ ) = d , F_{\mathrm{Cl}}(\boldsymbol{\theta})=F_{\mathrm{Q}}(\boldsymbol{\theta}),\qquad\operatorname{Tr}F_{\mathrm{Cl}}(\boldsymbol{\theta})=d, F Cl ( θ ) = F Q ( θ ) , Tr F Cl ( θ ) = d , and for θ ∈ Z \boldsymbol{\theta}\in\mathcal{Z} θ ∈ Z the inequality (9) is strict.
If, in addition, the parameterization is minimal, that is, if (cond3 _3 3 ) the parameter count satisfies d = 2 n − 1 d=2^n-1 d = 2 n − 1 , matching the dimension of the real-amplitude pure-state sphere S R 2 n − 1 ⊂ R 2 n S_{\mathbb{R}}^{2^n-1}\subset\mathbb{R}^{2^n} S R 2 n − 1 ⊂ R 2 n (modulo global sign), then d d d in (9) and (10) reads 2 n − 1 2^n-1 2 n − 1 .
Proof sketch. The full argument is in Appendix A. Reality of the state vector forces the phase overlap ⟨ ψ ∣ ∂ k ψ ⟩ \langle\psi|\partial_k\psi\rangle ⟨ ψ ∣ ∂ k ψ ⟩ to vanish, which is the computational-basis instance of the Facchi et al. [11 ] equality. The half-angle involutive generator then fixes the tangent norm at ∥ ∂ k ψ ∥ 2 = 1 / 4 \|\partial_k\psi\|^2=1/4 ∥ ∂ k ψ ∥ 2 = 1/4 , so every diagonal entry of the quantum Fisher matrix equals one and Tr F Q = d \operatorname{Tr}F_{\mathrm{Q}}=d Tr F Q = d exactly. The quantum–classical Cramér–Rao matrix inequality gives part (a). For part (b), writing p θ = ψ θ 2 p_{\boldsymbol{\theta}}=\psi_{\boldsymbol{\theta}}^2 p θ = ψ θ 2 turns each classical entry into 4 ∑ x ∂ j ψ θ ( x ) ∂ k ψ θ ( x ) 4\sum_x\partial_j\psi_{\boldsymbol{\theta}}(x)\partial_k\psi_{\boldsymbol{\theta}}(x) 4 ∑ x ∂ j ψ θ ( x ) ∂ k ψ θ ( x ) over outcomes with p θ ( x ) > 0 p_{\boldsymbol{\theta}}(x)>0 p θ ( x ) > 0 , and the quantum entry is the same sum over all x x x . The two matrices differ by the zero-amplitude terms of Eq. (37) , which vanish off Z \mathcal{Z} Z and have positive trace on it. ▫
Part (b) turns the inequality (9) into an identity on the operational parameter space, because the set Z \mathcal{Z} Z has measure zero. Indeed, each amplitude ψ θ ( x ) \psi_{\boldsymbol{\theta}}(x) ψ θ ( x ) is a real-analytic function of θ \boldsymbol{\theta} θ , and an amplitude that vanishes identically has vanishing derivatives and contributes nothing to Z \mathcal{Z} Z . Hence Z \mathcal{Z} Z lies in the union of the zero sets of the amplitudes that do not vanish identically, and the zero set of a real-analytic function that is not identically zero has Lebesgue measure zero. The strict inequality on Z \mathcal{Z} Z is a property of the convention that sums Eq. (5) over outcomes with p θ ( x ) > 0 p_{\boldsymbol{\theta}}(x)>0 p θ ( x ) > 0 . Wherever ψ θ ( x ) ≠ 0 \psi_{\boldsymbol{\theta}}(x)\ne0 ψ θ ( x ) = 0 the classical summand equals ( ∂ k p θ ( x ) ) 2 / p θ ( x ) = 4 ( ∂ k ψ θ ( x ) ) 2 (\partial_k p_\theta(x))^2/p_\theta(x) = 4(\partial_k\psi_\theta(x))^2 ( ∂ k p θ ( x ) ) 2 / p θ ( x ) = 4 ( ∂ k ψ θ ( x ) ) 2 , which has a finite limit as the amplitude vanishes, so the continuous extension of F c l F_{\mathrm{cl}} F cl equals F Q F_Q F Q at every θ \boldsymbol{\theta} θ and F c l F_{\mathrm{cl}} F cl as defined is discontinuous on Z \mathcal{Z} Z . The clamped estimator of Sec. II H does not take this limit, which is why it registers a deficit near Z \mathcal{Z} Z (§ 2 C).
The value of the trace depends on the coordinates. The matrix identity F c l = F Q F_{\mathrm{cl}} = F_Q F cl = F Q of part (b) is covariant under reparameterization, since both matrices transform as Λ − ⊤ F Λ − 1 \Lambda^{-\top}F\Lambda^{-1} Λ − ⊤ F Λ − 1 under ϑ = Λ θ \boldsymbol{\vartheta} = \Lambda\boldsymbol{\theta} ϑ = Λ θ . The value Tr F Q = d \operatorname{Tr}F_Q = d Tr F Q = d , however, belongs to the circuit angles, in which the half-angle convention fixes every diagonal entry at one, and a measure that normalizes by the trace, such as the effective dimension of Abbas et al. [8 ] , inherits this coordinate choice. For the Möttönen node angles of Appendix A 2 the trace is pinned as well: on the 2 k − 1 2^{k-1} 2 k − 1 angles of stage k = 1 , … , n k=1,\ldots,n k = 1 , … , n , Λ \Lambda Λ is a ± 1 \pm1 ± 1 matrix with Λ k Λ k T = 2 k − 1 I \Lambda_k\Lambda_k^{\mathsf{T}} = 2^{k-1}I Λ k Λ k T = 2 k − 1 I , so Λ − T Λ − 1 \Lambda^{-\mathsf{T}}\Lambda^{-1} Λ − T Λ − 1 equals 2 − ( k − 1 ) I 2^{-(k-1)}I 2 − ( k − 1 ) I there and Tr F ϑ = ∑ k = 1 n 2 − ( k − 1 ) 2 k − 1 = n \operatorname{Tr}F_{\boldsymbol{\vartheta}} = \sum_{k=1}^{n}2^{-(k-1)}2^{k-1}=n Tr F ϑ = ∑ k = 1 n 2 − ( k − 1 ) 2 k − 1 = n , which we confirmed numerically at n = 3 , 4 n=3,4 n = 3 , 4 . Because the trace is constant in both coordinate systems, the trace normalization of the effective dimension is the identity map for this ansatz, and the effective dimension reduces to a function of the shape of the Fisher spectrum alone.
Specialization to the Möttönen QCBM
The Möttönen ansatz of Sec. II D satisfies conditions (cond1 _1 1 )–(cond3 _3 3 ) by inspection, since the gate set { R Y , CNOT } \{R_Y,\operatorname{CNOT}\} { R Y , CNOT } has real matrix entries in the computational basis and the initial state ∣ 0 ⟩ ⊗ n \lvert0\rangle^{\otimes n} ∣ 0 ⟩ ⊗ n is real (cond1 _1 1 ), the R Y R_Y R Y rotations are half-angle rotations with Hermitian-unitary generator Y Y Y (cond2 _2 2 ), and the parameter count matches the dimension of the real-amplitude sphere (cond3 _3 3 ).
For the Möttönen real-amplitude state-preparation ansatz ∣ ψ ( θ ) ⟩ = U ( θ ) ∣ 0 ⟩ ⊗ n \lvert\psi(\boldsymbol{\theta})\rangle= U(\boldsymbol{\theta})\lvert0\rangle^{\otimes n} ∣ ψ ( θ )⟩ = U ( θ ) ∣ 0 ⟩ ⊗ n with parameter count d = 2 n − 1 d=2^n-1 d = 2 n − 1 , the trace C F I w ( θ ) = Tr F c l ( θ ) \mathrm{CFI}_{w}(\boldsymbol{\theta})=\operatorname{Tr}F_{\mathrm{cl}}(\boldsymbol{\theta}) CFI w ( θ ) = Tr F cl ( θ ) of the classical Fisher information matrix of the computational-basis output distribution p θ p_{\boldsymbol{\theta}} p θ obeys, for all θ ∈ R d \boldsymbol{\theta}\in\mathbb{R}^d θ ∈ R d ,
C F I w ( θ ) ≤ d = 2 n − 1 , (10) \mathrm{CFI}_{w}(\boldsymbol{\theta}) \le d = 2^n-1, \tag*{(10)} CFI w ( θ ) ≤ d = 2 n − 1 , (10)
with equality for every θ ∉ Z \boldsymbol{\theta}\notin\mathcal{Z} θ ∈ / Z . The Fisher trace of the QCBM is thus fixed by the circuit architecture alone: at almost every parameter point it takes the value 2 n − 1 2^n-1 2 n − 1 irrespective of the target distribution and of how well the circuit fits it.
Condition (cond3 _3 3 ) enters only to identify d d d with 2 n − 1 2^n-1 2 n − 1 , so the identity holds for any real-amplitude half-angle ansatz. This includes the layered hardware-efficient real-amplitude ansatz built from L + 1 L+1 L + 1 alternating R Y R_Y R Y layers separated by CNOT entanglers (d = n ( L + 1 ) d=n(L+1) d = n ( L + 1 ) parameters), which is over-parameterized when d > 2 n − 1 d>2^n-1 d > 2 n − 1 . We verified this numerically at random parameters across ( n , L ) (n,L) ( n , L ) combinations spanning n = 3 n=3 n = 3 –8 8 8 and L = 1 L=1 L = 1 –5 5 5 , including over-parameterized cases, and found Tr F c l / d = 1.000 \operatorname{Tr}F_{\mathrm{cl}}/d=1.000 Tr F cl / d = 1.000 to numerical precision in every case.
Table 3. Clamped Fisher-trace estimator across the campaign (420 runs per n n n ). C F I w max = 2 n − 1 \mathrm{CFI}_{w}^{\max}=2^n-1 CFI w m a x = 2 n − 1 is the bound of Eq. (10) , attained off Z \mathcal{Z} Z ; Saturated (%) counts runs within 0.01 of it. Off the zero-amplitude set the saturation is an identity, so the table validates the estimator; the sub-cap entries are the clamp, Eq. (11) .
Empirical saturation and the clamped estimator
Recomputing the campaign’s converged parameter points with the estimator’s probability floor removed returns Tr F c l ( θ ) = d \operatorname{Tr}F_{\mathrm{cl}}(\boldsymbol{\theta})=d Tr F cl ( θ ) = d in every run at every qubit count, to within 10 − 8 10^{-8} 1 0 − 8 , as Proposition 1 (b) requires, so Table III validates the estimator. With the floor in place, the clamped trace 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 at each scale.
The sub-cap entries originate in the clamp alone, since neither the zero-amplitude set nor the optimizer produces them. Estimation near Z \mathcal{Z} Z is the operational regime, because a converged fit to the two-peak B1 target drives 2 n − 2 2^n-2 2 n − 2 amplitudes toward zero (median minimum basis probability 2.8 × 10 − 8 2.8\times10^{-8} 2.8 × 1 0 − 8 at n = 6 n=6 n = 6 ; § 2 H, Appendix A1). The exact parameter-shift Jacobian of § 2 H removes finite-difference cancellation there. What remains is the inverse-probability weighting of rare outcomes. Writing p f l o o r = 10 − 10 p_{\mathrm{floor}}=10^{-10} p floor = 1 0 − 10 for the probability floor, the clamped estimator replaces 1 / p θ ( x ) 1/p_\theta(x) 1/ p θ ( x ) by 1 / p f l o o r 1/p_{\mathrm{floor}} 1/ p floor on outcomes below the floor, which underweights precisely those outcomes. Off Z \mathcal{Z} Z the estimate therefore falls short of d d d by
d − Tr F ^ c l = 4 ∑ x : 0 < p θ ( x ) < p f l o o r ( 1 − p θ ( x ) p f l o o r ) ∑ k ( ∂ k ψ θ ( x ) ) 2 , (11) d-\operatorname{Tr}\widehat{F}_{\mathrm{cl}}=4\sum_{x:0<p_\theta(x)<p_{\mathrm{floor}}}\left(1-\frac{p_\theta(x)}{p_{\mathrm{floor}}}\right)\sum_k(\partial_k\psi_\theta(x))^2, \tag*{(11)} d − Tr F cl = 4 x : 0 < p θ ( x ) < p floor ∑ ( 1 − p floor p θ ( x ) ) k ∑ ( ∂ k ψ θ ( x ) ) 2 , (11)
which is strictly positive exactly when some outcome falls below the floor. We verify Eq. (11) run by run across the campaign, and we find that every sub-cap run carries at least one outcome below the floor, no run without such an outcome falls below the cap, and the observed deficit matches the prediction to a maximum relative residual of 1.2 × 10 − 8 1.2\times10^{-8} 1.2 × 1 0 − 8 . The smallest ∣ ψ θ ( x ) ∣ \lvert\psi_\theta(x)\rvert ∣ ψ θ ( x )∣ carrying non-negligible ∣ ∂ k ψ θ ( x ) ∣ \lvert\partial_k\psi_\theta(x)\rvert ∣ ∂ k ψ θ ( x )∣ separates the two populations: its median is ∼ 10 − 6 \sim10^{-6} ∼ 1 0 − 6 on the sub-cap runs against 10 − 5 10^{-5} 1 0 − 5 to 10 − 2 10^{-2} 1 0 − 2 on the saturated ones. For the Möttönen ansatz the set Z \mathcal{Z} Z is a union of hyperplane families in the node angles of the circuit’s binary tree, and exact sparse fits lie on it (Appendix A 2). How Z \mathcal{Z} Z relates to the set of parameters the optimizer reaches remains open.
The cost of observing the identity from measurement shots is quantified in Appendix A 1.
The trace is therefore constant on the operational parameter space and carries no certification signal; we retain it only as a check on the estimator.
Testing the rank and the pooled scores
With the trace removed, two candidate channels remain: the eigenvalue spectrum of the same Fisher matrix and the pooled distributional scores. We score both against the benchmark labels and then subject each to controls that separate distributional quality from the experimental design.
The Fisher rank counts the support of the converged distribution
Proposition 1 pins the trace of F c l ( θ ) F_{\mathrm{cl}}(\theta) F cl ( θ ) but not its spectrum. The capacity interpretation of its rank is due to Larocca et al. [15 ] and, for the QCBM specifically, to Delgado et al. [16 ] ; here we ask whether it certifies. We recomputed the spectrum of F c l ( θ ) F_{\mathrm{cl}}(\theta) F cl ( θ ) at the converged parameters of every publication run and report the numerical rank with eigenvalue threshold 10 − 6 10^{-6} 1 0 − 6 relative to the largest eigenvalue, denoted rank 10 − 6 ( F c l ( θ ) ) \operatorname{rank}_{10^{-6}}(F_{\mathrm{cl}}(\theta)) rank 1 0 − 6 ( F cl ( θ )) .
Table 4. What the rank counts, on B1 at cost C 1 C_1 C 1 . 2 n − 1 2^n - 1 2 n − 1 is the number of live nodes of the amplitude tree (Appendix A 2) for an exact fit to the two-peak target, the rank of the quantum Fisher matrix there; “at 2 n − 1 2^n - 1 2 n − 1 ” is the fraction of converged unperturbed runs (training residual below 0.005; below 0.02 at n = 8 n = 8 n = 8 , where none reaches 0.005) whose rank 10 − 6 \operatorname{rank}_{10^{-6}} rank 1 0 − 6 equals it, with the number of such runs in parentheses. The classes are gap-separated (largest unperturbed rank below smallest perturbed rank) through n = 7 n = 7 n = 7 .
Figure 1 reports the rank-diagnostic AUROC on the unperturbed-vs-perturbed labels of benchmark B1 alongside the corresponding W 1 W_1 W 1 and MMD AUROCs at cost C 1 C_1 C 1 . The rank diagnostic separates the two classes perfectly through n = 7 n = 7 n = 7 at costs C 1 C_1 C 1 and C 2 C_2 C 2 and reaches 0.999 at C 3 C_3 C 3 . The distributional scores W 1 ( C 1 ) W_1(C_1) W 1 ( C 1 ) and MMD track each other closely and fall to AUROC ≈ 0.70 \approx0.70 ≈ 0.70 –0.78 over n = 5 n = 5 n = 5 –7 7 7 , with intervals spanning 0.2 to 0.4 (Table VII ). At n = 8 n = 8 n = 8 the rank falls to 0.959–0.966 at every cost. Through n = 7 n = 7 n = 7 at C 1 C_1 C 1 and C 2 C_2 C 2 the classes are gap-separated outright, every perturbed run outranking every unperturbed one (Table IV), so the rank interval is degenerate there while the distributional intervals are too wide to order the metrics at all. The separating cuts, including the amplitude-tree count 2 n − 1 2^n - 1 2 n − 1 , were identified after the campaign had been inspected, so the perfect scores carry no distribution-free lower confidence bound. Such a bound requires a cut fixed in advance, and the 80 perturbed runs of a cell, four levels of the same 20 seeds, would enter it as 20 clusters. The degenerate bootstrap interval [ 1.000 , 1.000 ] [1.000,1.000] [ 1.000 , 1.000 ] of a gap-separated cell describes the observed sample. Doubling the unperturbed class to 40 seeds at every scale leaves all 360 added runs rank-deficient and, at n ≤ 7 n \le7 n ≤ 7 , below the smallest perturbed rank of their cells, the same observation on a second sample. The W 1 W_1 W 1 and MMD decline in Table VII is the convergence labeling artifact of Sec. IVB and is removed by the convergence control developed there.
Figure 1. AUROC vs. qubit count on B1 at cost C 1 C_1 C 1 for the rank diagnostic, W 1 W_1 W 1 and MMD under both references, and the debiased Sinkhorn divergence (regularization σ = 0.5 \sigma= 0.5 σ = 0.5 ). The rank reaches 1.000 through n = 7 n = 7 n = 7 and 0.966 at n = 8 n = 8 n = 8 (per-cost values and intervals in Table VII ); the distributional depression over n = 5 n = 5 n = 5 –7 7 7 is the labeling effect removed in Figure 3 , and the two conventions separate only at n = 8 n = 8 n = 8 . Shading: 95% seed-cluster bootstrap bands of the clean-reference W 1 W_1 W 1 (blue) and MMD (red) curves, which span 0.18–0.40 at n ≥ 5 n \ge5 n ≥ 5 (Table VII ).
What the rank counts follows from the amplitude tree of Appendix A 2, and it requires distinguishing three matrices: the support-restricted classical matrix and the quantum Fisher matrix, which coincide off the zero-amplitude locus and separate on it, and the clamped estimator, which differs from both wherever an outcome lies below the floor. At an exact fit to the two-peak target every amplitude outside the two peaks vanishes, so the support-restricted classical Fisher matrix, Eq. (5) summed over the outcomes of positive probability, is built from two rows whose derivatives sum to zero and has rank at most one, the general bound being rank F c l ≤ s − 1 \operatorname{rank} F_{\mathrm{cl}} \le s - 1 rank F cl ≤ s − 1 on a support of size s s s . The quantum Fisher matrix, the continuous extension of F c l F_{\mathrm{cl}} F cl through the locus by Proposition 1 , does not drop the pruned outcomes. An exact fit prunes every subtree that carries no mass, and the amplitude Jacobian vanishes on the angles inside a pruned subtree. The parameters that survive are the live nodes on the two root-to-leaf paths, 2 n − 1 2n - 1 2 n − 1 of them, which is the rank of F Q F_Q F Q at that point (rank 5, 7 and 9 at n = 3 n = 3 n = 3 , 4 and 5 in the amplitude-tree model, against rank 1 for the support-restricted matrix). The clamped estimator of Sec. II H is neither of these matrices exactly. No run of the campaign has an amplitude that vanishes to machine precision, so the support-restricted matrix coincides with F Q F_Q F Q on every run; the clamp replaces 1 / p θ ( x ) 1/p_\theta(x) 1/ p θ ( x ) by 1 / p f l o o r 1/p_{\mathrm{floor}} 1/ p floor on the 599 runs that carry an outcome below the floor (§ 2 C). We recomputed the rank at the stored parameters of all 2,520 runs from the exact amplitude Jacobian, with and without the clamp. The clamped rank equals the rank of F Q F_Q F Q on 2,431 runs, including every run at n ≤ 5 n \le5 n ≤ 5 and every B2 and B3 run. On the 89 sub-floor B1 runs at n = 6 n = 6 n = 6 to 8 where the two differ, the clamped rank is smaller by one or two in a dimension of at least 63. Proposition 5 (Appendix A 3) turns the trace deficit of Eq. (11) into a per-run certificate of agreement. It certifies 2,024 of the 2,520 runs, every run without a sub-floor outcome and 103 of the 599 with one. All 89 disagreements lie among the 496 uncertified runs, of which 407 agree nonetheless. The smallest observed decision margin in a certified run is 10 − 9 λ max 10^{-9}\lambda_{\max} 1 0 − 9 λ m a x , about four orders of magnitude above the nominal double-precision eigenvalue-error scale; this is a numerical robustness check without interval arithmetic. The fraction at 2 n − 1 2n - 1 2 n − 1 in Table IV is the same for both matrices at every n n n , and the rank AUROC on B1 at C1 agrees to three decimals through n = 7 n = 7 n = 7 and reads 0.966 against 0.964 at n = 8 n = 8 n = 8 . The reported clamped ranks are therefore an empirical approximation to the rank of F Q F_Q F Q , exact in 96.5% of the runs and off by at most two in the rest, and through it to the support at the resolution of the eigenvalue threshold. They need not equal the rank of the support-restricted classical matrix in general, although on this campaign, whose amplitudes are all nonzero, that matrix coincides with F Q F_Q F Q . The common trace cap of Eq. (10) does not by itself make the spectra or the eigendirections of the two matrices identical. Every rank and eigenvector reported in this paper is the clamped estimator. The converged unperturbed runs realize the live-node count exactly at small n n n (Table IV): every run with residual below 0.005 has rank 5 at n = 3 n = 3 n = 3 and 7 at n = 4 n = 4 n = 4 , 86% have rank 9 at n = 5 n = 5 n = 5 , and no unperturbed run at any scale attains full rank. From n = 6 n = 6 n = 6 the count is exceeded, by a factor of two at n = 6 n = 6 n = 6 and four at n = 7 n = 7 n = 7 , because CMA-ES leaves background mass above the 10 − 6 10^{-6} 1 0 − 6 eigenvalue threshold on outcomes that an exact fit would zero, and the rank then reads how far the optimizer got toward exact sparsity.
Perturbed targets have full support, and the harder fit leaves them at or near full rank; converged unperturbed fits of the full-support targets B2 and B3 are full rank at every n n n . The rank is therefore a property of the output distribution, its support at the resolution of the eigenvalue threshold, and the benchmark label coincides with that support on B1 because the blend is the only full-support class. To make the resolution quantitative, let L c ( q ) L_c(q) L c ( q ) be the number of internal nodes of the amplitude tree, at depths 0 0 0 to n − 1 n - 1 n − 1 , that are prefixes of some outcome with q ( x ) > c q(x) > c q ( x ) > c . The cutoff c c c is an absolute mass on the leaves, distinct from the relative eigenvalue cutoff of the rank, and a node whose subtree mass exceeds c c c only through several sub-cutoff leaves is not counted; at c = 10 − 6 c = 10^{-6} c = 1 0 − 6 the two counts agree on 81% of campaign runs and on every B3 run, and differ by up to 20 nodes on B1 runs at n = 6 n = 6 n = 6 and 8. L c = 2 n − 1 L_c = 2n - 1 L c = 2 n − 1 for an exact two-peak fit and L c = d L_c = d L c = d for full support. At c = 10 − 6 c = 10^{-6} c = 1 0 − 6 , the eigenvalue threshold itself, L c L_c L c predicts rank 10 − 6 \operatorname{rank}_{10^{-6}} rank 1 0 − 6 across all 2,520 runs with Pearson r = 0.9996 r = 0.9996 r = 0.9996 , equality in 75% of runs, a median deviation of zero and a 90th-percentile deviation of 0.02 d d d , so it is a close proxy for the rank. On the converged unperturbed B1 runs at n = 6 n = 6 n = 6 and 7 it reads 20 and 54 against ranks 22 and 56, so the excess over 2 n − 1 2n - 1 2 n − 1 is background mass above 10 − 6 10^{-6} 1 0 − 6 . The converged B2 fits are full rank for the same reason: at n = 3 n = 3 n = 3 the B2 target has entropy 0.19 nats but carries mass above 10 − 6 10^{-6} 1 0 − 6 on every outcome, so L 10 − 6 = d L_{10^{-6}} = d L 1 0 − 6 = d in 90% of converged B2 fits and in every converged B3 fit. The trace still saturates the cap of Eq. (10) on every run, so rank deficiency concentrates the eigenvalue mass on the live subspace.
Three controls confirm that the rank reads the support and is blind to the defect. First, a median split on ϵ \epsilon ϵ among the perturbed runs, asking each diagnostic to grade defect severity, reduces the mean rank AUROC at n = 3 n = 3 n = 3 –6 to 0.54–0.56 while W 1 W_1 W 1 and MMD keep 0.89–0.95 against the clean reference, the appropriate convention since severity is defined against the clean target. Against the trained-on target the distributional means fall to 0.67–0.68; the rank uses no reference and is unmoved. Pooled at C1 the rank reads 0.510 [0.450, 0.570], chance, and 0.511 [0.447, 0.572] on the restarted campaign. Of all quantities in this section, the rank alone is moved neither by convergence conditioning nor by the restart policy.
Second, on the pre-registered stress test, whose families keep the support fixed across levels, the perfect AUROC collapses entirely: exactly 0.50 on the correlation-shift family at every tested n n n (both classes have full support and the rank is uniformly full) and 0.55 falling to 0.37 (CI [0.23, 0.53]) on mode collapse, against 1.000 on B1.
Third, every run behind these numbers uses one optimizer, so we retrained B1 at n = 3 , 4 , 5 n = 3, 4, 5 n = 3 , 4 , 5 on the same stored targets under three alternatives. Powell’s method keeps the W 1 W_1 W 1 objective but replaces the search. Adam on an MMD objective with exact gradients changes both. Adam on the W 1 W_1 W 1 objective itself, with the exact gradient given by the optimal dual potential of the Kantorovich problem, changes only the search. Table V reports the four arms. Adam on W1 _1 1 reproduces the CMA-ES result exactly, a rank AUROC of 1.000 with no unperturbed fit at full rank at any of the three scales. Adam on MMD leaves every unperturbed fit at full rank at n = 3 n = 3 n = 3 and scores at chance there. The count the rank reaches is thus set by the objective: a W1 _1 1 fit drives the background of a sparse target to zero and an MMD fit does not. Powell and Adam on W1 _1 1 converge poorly at n = 5 n = 5 n = 5 , with median unperturbed residuals of 0.29 and 0.41, so their values there partly re-express the convergence confound.
Table V. Rank AUROC on B1 at three relative eigenvalue thresholds, cost C 1 C_1 C 1 , with the range over the three ground costs in brackets; and the optimizer replication of the rank confound at n = 3 n = 3 n = 3 , 4 and 5, listed in that order within each cell (rank AUROC, fraction of unperturbed fits at full rank, median unperturbed training residual). CMA-ES is the campaign optimizer; Powell keeps the W1 _1 1 objective; Adam uses an MMD objective with exact gradients.
Once the parameter count is large enough, every optimizer we tried leaves unperturbed fits rank-deficient and perturbed fits substantially higher-rank. The scale at which this sets in depends on the optimizer and on the objective (Table V ).
Rank deficiency also occurs among runs that the protocol of § 4 certifies; Sec. VI tests whether it marks fragile verdicts.
Every rank verdict depends on a single hyperparameter, so we vary it (Table V ). Recomputing the diagnostic at relative eigenvalue thresholds of 10 − 5 10^{-5} 1 0 − 5 and 10 − 7 10^{-7} 1 0 − 7 alongside the 10 − 6 10^{-6} 1 0 − 6 used throughout leaves the pattern unchanged; the only weakening is at n = 3 n = 3 n = 3 for 10 − 7 10^{-7} 1 0 − 7 , where the parameter count is small enough that a single admitted eigenvalue changes the score, and the n = 8 n = 8 n = 8 breakdown is present at every threshold.
In its present formulation the rank is a simulation-time diagnostic, because it requires resolving the CFIM spectrum down to a small relative eigenvalue, a strictly harder estimation task than the scalar trace, whose shot cost already grows steeply with n n n (Figure 7 ). Two non-saturating Fisher-spectrum diagnostics refine the rank, namely the effective dimension of Abbas et al. [8 ] and the entropy-based effective rank of Roy and Vetterli [28 ] . Recomputed on the same converged CFIM spectra, they inherit the same design-recovery confound. On the B1 perturbation labels the Roy–Vetterli effective rank and a participation-ratio effective dimension reach AUROC = 0.96–1.00 across n = 3 n = 3 n = 3 –6, against 1.000 for the numerical rank.
Figure 7. Interquartile width of the finite-shot Fisher trace under the Krichevsky–Trofimov estimator, normalized by the cap, over 8 random saturated parameter points and 40 shot replicates (at n = 7 n=7 n = 7 one of the eight draws fell more than 0.01 below the cap under the probability floor and was excluded, leaving 7); the dotted line marks a ± 1 % \pm1\% ± 1% band. The budget required to resolve the cap grows from S ≈ 3 × 10 4 S \approx3 \times10^{4} S ≈ 3 × 1 0 4 at n = 3 n=3 n = 3 to 3 × 10 5 3 \times10^{5} 3 × 1 0 5 at n = 8 n=8 n = 8 .
The pooled distributional scores track convergence and reference
Figure 2 and Table VI report the training residual W 1 ( C 1 ) W_1(C_1) W 1 ( C 1 ) on B1. At n ≥ 4 n \ge4 n ≥ 4 the unperturbed entries are elevated (0.058–0.170) because CMA-ES does not always fit the clean target within budget, so the residual is non-monotone in ϵ \epsilon ϵ and reflects fit quality. In contrast, defect magnitude is the distance to the clean target, which rises monotonically with ϵ \epsilon ϵ . The elevated unperturbed residuals are the convergence confound analyzed below.
Figure 2. W 1 ( C 1 ) W_1(C_1) W 1 ( C 1 ) vs. perturbation level on B1, both references, with 95% bootstrap CIs over 20 seeds (symmetric-log axes). The trained-on residual (top) is non-monotone in ϵ \epsilon ϵ and ordered by fit difficulty; the clean-reference distance (bottom) rises monotonically with ϵ \epsilon ϵ .
Table VII reports AUROC with the perturbation level as a proxy quality label, with a seed-cluster bootstrap 95% confidence interval on every entry. Each cell has
20 negatives and 80 positives, so a single unperturbed run misranked against the whole positive class moves the AUROC by 1 / 20 = 0.05 1/20 = 0.05 1/20 = 0.05 ; the quantization visible in the recurring values 0.950 and 0.900 is exactly this granularity, and the cluster intervals span 0.15 to 0.40 at n ≥ 5 n \ge5 n ≥ 5 . Metrics within a scale are compared only through their paired seed-cluster differences, reported below, since the marginal intervals carry the between-seed variation that a paired comparison removes.
Within those intervals the pooled scores do not decay monotonically with n n n (clean-reference W 1 ( C 1 ) W_1(C_1) W 1 ( C 1 ) : 1.000, 0.902, 0.712, 0.772, 0.775, 0.920 at n = 3 n = 3 n = 3 –8). They are, however, depressed over n = 5 n = 5 n = 5 –7 by an amount the intervals resolve. The depression is a labeling effect: the label calls every ϵ = 0 \epsilon= 0 ϵ = 0 run “good”, the optimizer does not always fit the unperturbed target, and a run labeled good while distributionally wrong penalizes a metric for correctly flagging it. Restricting the negative class to converged unperturbed runs, by dropping the ϵ = 0 \epsilon= 0 ϵ = 0 runs whose training residual exceeds a fit-quality threshold τ f i t = 0.02 \tau_{\mathrm{fit}} = 0.02 τ fit = 0.02 , lifts every ( n , c o s t ) (n, \mathrm{cost}) ( n , cost ) cell to AUROC = 1.000 for both metrics, at every scale from n = 3 n = 3 n = 3 to n = 8 n = 8 n = 8 (Figure 3 ). After the cut, no converged unperturbed run scores worse than any perturbed one. The result is flat for τ f i t \tau_{\mathrm{fit}} τ fit between 0.005 and 0.10 and is reproduced by a fit-quality criterion independent of the scoring metric, the total-variation residual. Because the W 1 W_1 W 1 residual at ϵ = 0 \epsilon= 0 ϵ = 0 is itself the score being evaluated against the clean reference, we also cut the negatives on the other channel: restricting to unperturbed runs with M M D k 2 ≤ 10 − 4 \mathrm{MMD}_k^2 \le10^{-4} MMD k 2 ≤ 1 0 − 4 and scoring W 1 W_1 W 1 against the clean reference gives 1.000 at every n n n (negatives 20, 18, 14, 15, 15, 16), and the MMD AUROC under the W 1 W_1 W 1 cut is likewise 1.000 everywhere. At the looser M M D k 2 ≤ 10 − 3 \mathrm{MMD}_k^2 \le10^{-3} MMD k 2 ≤ 1 0 − 3 cut the W 1 W_1 W 1 AUROC is 1.000 through n = 7 n = 7 n = 7 and 0.966 at n = 8 n = 8 n = 8 with 19 negatives, so the perfect score at n = 8 n = 8 n = 8 requires negatives restricted to near-perfect fits. The optimizer’s own stopping rule cannot supply the cut, because CMA-ES exhausts its budget in 20 of 20 unperturbed runs at n = 5 n = 5 n = 5 and in 19 of 20 at n = 6 n = 6 n = 6 and 7. The cut does, however, spend negatives, leaving 20, 18, 14, 15, 15 and 13 converged negatives per C 1 C_1 C 1 cell at n = 3 n = 3 n = 3 through 8. As in Sec. IV A, the cut lies in an observed gap and selects the negatives by their residual after the fact, so the degenerate intervals describe the surviving sample. At n = 8 n = 8 n = 8 under cost C 3 C_3 C 3 only four negatives survive, too few to support a claim.
Figure 3. Convergence control on the B1 W 1 ( C 1 ) W_1(C_1) W 1 ( C 1 ) clean-reference AUROC under both campaign variants. Pooled (dashed) it moves by up to 0.23 between the campaign as run and the restarted one; conditioned on converged negatives (residual ≤ 0.02 \le0.02 ≤ 0.02 ) it is 1.000 everywhere and identical under both. Small numerals give the surviving negatives, which bound the precision of each conditioned cell.
A post hoc intervention, which we keep out of the primary analysis, leads to the same conclusion. Part of the campaign was later re-optimized by warm-restarting CMA-ES from each run’s incumbent parameters under a uniform residual criterion. The restart targets CMA-ES under-convergence, which elevates unperturbed residuals and can mimic a degradation of discrimination with scale. A run was eligible when its training residual exceeded τ W 1 = 0.05 \tau_{W_1} = 0.05 τ W 1 = 0.05 , and the driver warm-started CMA-ES from the run’s incumbent parameters under the campaign’s stopping rule. The as-run parameters analyzed here are byte-preserved pre-restart snapshots, and we verified per run that the stored metrics reproduce from the stored parameters, so the two campaign variants are exactly reconstructible from the archive.
The restart driver was applied to both label classes and to all three benchmarks, but its coverage was not uniform. Of the runs meeting the restart criterion it reached 99% of unperturbed B1 runs against 48% of perturbed ones, and the shortfall concentrates at n = 7 n = 7 n = 7 and 8, where perturbed coverage falls to 34% and 36% while unperturbed coverage stays at 100% and 93%. Through n ≤ 5 n \le5 n ≤ 5 coverage is complete in both classes; beyond that an intervention that improves one class’s convergence more than the other’s acts directly on the label structure. We therefore report the campaign as run and treat the restarted one as a robustness check. The restart raises the pooled distributional AUROC in every cell, by +0.05 to +0.34, with the largest shifts where its coverage is most asymmetric. It moves the rank AUROC by at most +0.02, and by nothing through n = 7 n = 7 n = 7 , while the convergence-conditioned AUROC stays at 1.000 everywhere.
The choice of reference distribution moves the pooled score independently of both. At n = 8 n = 8 n = 8 the trained-on and clean scores separate for MMD, 0.735 [0.590, 0.869] against 0.896 [0.781, 0.979] at C 1 C_1 C 1 . The marginal intervals overlap, so the comparison needs the paired statistic: a seed-cluster bootstrap of the AUROC difference on the same runs gives +0.161 [0.084, 0.243] at C 1 C_1 C 1 , +0.144 [0.082, 0.214] at C 2 C_2 C 2 and +0.104 [0.052, 0.173] at C 3 C_3 C 3 for MMD, and +0.062 [0.018, 0.121] at C 1 C_1 C 1 for W 1 W_1 W 1 , each excluding zero. At n ≤ 7 n \le7 n ≤ 7 the two conventions differ by at most 0.06 outside n = 3 n = 3 n = 3 . As noted in Sec. II H, a training residual measures fit quality, and at n = 8 n = 8 n = 8 the perturbed targets are the harder ones to fit, so scoring against the trained-on target increasingly rewards the wrong runs. The best-scoring W 1 W_1 W 1 ground cost also changes with scale (Table VII ); the paired seed-cluster difference between C 1 C_1 C 1 and C 2 C_2 C 2 on the clean-reference W 1 W_1 W 1 AUROC excludes zero at n = 5 n = 5 n = 5 alone, − 0.24 -0.24 − 0.24 [− 0.46 -0.46 − 0.46 , − 0.04 -0.04 − 0.04 ], and covers zero at every other scale, so no cost is best at every scale. The same paired statistic orders the two metrics: W 1 W_1 W 1 scores above MMD against the clean reference in all 18 B1 cells, by 0.02 to 0.07, and the difference excludes zero in 8 of them, at every cost for n = 7 , 8 n = 7, 8 n = 7 , 8 and at C 3 C_3 C 3 for n = 5 , 6 n = 5, 6 n = 5 , 6 . The gap is an order of magnitude below the convergence effect of Figure 3 , and certification depends on calibration at a fixed operating point (§ 4 ), which this ordering leaves open.
The Sinkhorn divergence [29 ] interpolates between W 1 W_1 W 1 and MMD and dominates pure MMD on QCBM training in Coyle et al. [20 ] , so it is the obvious candidate to replace the pair; Table VII and Figure 1 report it at entropic regularization σ = 0.5 \sigma= 0.5 σ = 0.5 . Sinkhorn sits between the clean-reference W 1 W_1 W 1 and MMD values at every scale through n = 7 n = 7 n = 7 and slightly below both at n = 8 n = 8 n = 8 (0.881 against 0.896 and 0.920), and in every cell its value lies within 0.02 of one of the two. The W 1 W_1 W 1 /Sinkhorn verdict disagreement is likewise negligible below n = 6 n = 6 n = 6 (0.0% at n = 3 , 4 , 5 n = 3, 4, 5 n = 3 , 4 , 5 ), rising to 1.0%, 5.0% and 17.0% at n = 6 , 7 , 8 n = 6, 7, 8 n = 6 , 7 , 8 as the fits degrade. Sinkhorn thus tracks the pair at every scale, and σ ∈ { 0.1 , 1.0 } \sigma\in\{0.1, 1.0\} σ ∈ { 0.1 , 1.0 } leaves this unchanged. We ran no paired comparison between it and either metric. The MMD scores are insensitive to the kernel bandwidth within the cluster intervals: recomputing the B1 C 1 C_1 C 1 MMD AUROC at γ ∈ { 1 / ( 2 n ) , 1 / n , 2 / n } \gamma\in\{1/(2n), 1/n, 2/n\} γ ∈ { 1/ ( 2 n ) , 1/ n , 2/ n } and at the median-heuristic bandwidth changes it by at most 0.02 at every n ≤ 7 n \le7 n ≤ 7 and by at most 0.05 at n = 8 n = 8 n = 8 , so the γ = 1 / n \gamma= 1/n γ = 1/ n choice explains neither the scale-dependent dip nor the n = 8 n = 8 n = 8 MMD shortfall.
The distributional scores remain usable as certification channels, provided the scoring convention is stated as part of the claim. Three conventions follow. The negative class must be conditioned on convergence, since the cut lifts every B1 cell to 1.000. Scores must be taken against the clean reference, since the two conventions separate by 0.16 at n = 8 n = 8 n = 8 (paired interval [0.08, 0.25]). Any post hoc re-optimization must be reported with its coverage by label class, since that lever alone moves the pooled figure (Figure 3 ). These controls explain why the pooled scores decline. Because the convergence cut selects on a residual that is itself a distributional score, deployment discrimination is tested separately, without conditioning, at fixed thresholds on the full-support benchmark (§ 4 K).
A full-support labeled benchmark {#sec:A full-support labeled benchmark}
Every AUROC above is computed on B1, whose perturbation label coincides with the support of the target, since the blend is the only full-support class. To separate the two, we built a labeled benchmark whose clean target already has full support: the near-uniform B3 target of Sec. II G blended with the same random background as B1, p ϵ = ( 1 − ϵ ) p B 3 + ϵ p_\epsilon= (1-\epsilon)p_{\mathrm{B3}}+\epsilon p ϵ = ( 1 − ϵ ) p B3 + ϵ , at the five levels of Eq. (8) , cost C 1 C_1 C 1 , n = 3 n = 3 n = 3 –8, 20 seeds per level and the publication budgets. Three predictions were fixed before launch in a dated document (see Data availability): (P1) the rank AUROC lies in [0.35, 0.65] at every n n n ; (P2) the convergence cut lifts the clean-reference W 1 W_1 W 1 and MMD AUROCs to at least 0.90 in every cell with at least 10 surviving negatives; (P3) converged clean fits are full rank. The runs at n = 7 n = 7 n = 7 and 8 were launched after the outcome at n ≤ 6 n \le6 n ≤ 6 was known, under a dated addendum that keeps the design and the three predictions. Table VIII reports the outcome.
Table VIII. Perturbed full-support benchmark (B3 blends, cost C 1 C_1 C 1 , 100 runs per n n n ). AUROC on the ϵ = 0 \epsilon= 0 ϵ = 0 vs ϵ > 0 \epsilon> 0 ϵ > 0 labels for the rank and for the clean-reference W1 _1 1 and MMD, with seed-cluster bootstrap 95% CIs; the same scores after the convergence cut on the negatives (W1 _1 1 residual ≤ 0.02 \leq0.02 ≤ 0.02 ; the cross-metric cut MMDk 2 ≤ 10 − 4 _k^2 \leq10^{-4} k 2 ≤ 1 0 − 4 for W1 _1 1 ), with surviving negatives in parentheses; the median training residual of the unperturbed fits; the exact clean-reference W1 _1 1 of the injected defect at ϵ = 0.02 \epsilon= 0.02 ϵ = 0.02 (median over seeds); and the per-level AUROC (W1 _1 1 /MMD) of the ϵ = 0.02 \epsilon= 0.02 ϵ = 0.02 and ϵ = 0.20 \epsilon= 0.20 ϵ = 0.20 levels against ϵ = 0 \epsilon= 0 ϵ = 0 . A conditioned cell is left empty when no negative survives the cut.
P1 and P3 hold at every n n n . The rank AUROC is 0.500, 0.500, 0.500, 0.494, 0.525 and 0.494 at n = 3 n = 3 n = 3 –8, 597 of the 600 fits are full rank, and the median rank is d d d in both classes, so the rank carries no label information once the support is fixed, as the support reading of Sec. IV A requires. P2 fails at n = 6 n = 6 n = 6 , where the conditioned MMD AUROC is 0.863 (W 1 W_1 W 1 : 0.937), and it cannot be tested at n = 7 n = 7 n = 7 and 8, where the clean near-uniform target does not converge within the publication budget (median residual 0.043 and 0.046) and at most one unperturbed fit survives the cut. The pooled clean-reference scores, 0.981, 0.991, 0.890, 0.811, 0.696 and 0.734 for W 1 W_1 W 1 and 0.968, 0.973, 0.816, 0.706, 0.529 and 0.604 for MMD, stay below the conditioned B1 values even where the unperturbed fits converge (median residual 0.000, 0.002, 0.004 and 0.014 at n = 3 n = 3 n = 3 –6). The per-level columns show why.
On the near-uniform target the injected defect is small: at ϵ = 0.02 \epsilon= 0.02 ϵ = 0.02 its clean-reference W1 _1 1 is 0.005–0.007 at every n n n , against roughly 0.02–0.03 on B1, and from n = 6 n = 6 n = 6 on it lies below the unperturbed fit residual, where the per-level W1 _1 1 AUROC falls to 0.598, 0.495 and 0.453. At ϵ = 0.20 \epsilon= 0.20 ϵ = 0.20 the defect (0.055–0.069) exceeds the median residual at every n n n , and the W1 _1 1 AUROC is 1.000 at n = 3 n = 3 n = 3 –6 and 0.978 and 0.975 at n = 7 n = 7 n = 7 and 8, where the defect exceeds the residual by only 20–30%. Across the 24 level cells the per-level AUROC rises with the ratio of the defect to the median residual (Spearman rank correlation 0.86 for W1 _1 1 and 0.93 for MMD). The distributional channels therefore resolve a defect on a full-support target to the extent that it exceeds the fit residual, and
the perfect conditioned scores of Sec. IV B are a property of the sparse benchmark, on which the smallest injected defect is two to five times the residual. The conditioning removes mislabeled negatives; it cannot create resolution below the residual. The trained-on reference scores at 0.39–0.62 throughout, near chance, as a fit-quality measure should when both classes fit comparably well. In the next section, we build a certification protocol on the two distributional channels and bound its false-pass probability.
The bi-metric escalation protocol
The protocol of this section uses the two distributional channels that Secs. III and IV leave, with the rank retained as a flag and the scoring convention fixed by the controls of Sec. IV B. The pair is a design choice among the admissible distributional channels (Sec. VI); we study it because both metrics admit closed forms on the defect families below and their composition admits a finite-sample bound. We bound the false-pass probability of the protocol, compute where it improves on single-metric certification, and test each claim on trained generators and at finite shot budgets. An AUROC of one guarantees that some threshold separates the classes; a fixed pair ( τ W 1 , τ M M D ) (\tau_{\mathrm{W1}}, \tau_{\mathrm{MMD}}) ( τ W1 , τ MMD ) may still fail to. Every claim below is therefore made at a stated operating point, and Secs. V D–V G show that the operating point, through a single tolerance ratio, decides which defects the escalation can see.
Three populations recur below and are distinct. A perturbed pair carries the benchmark label ϵ > 0 \epsilon> 0 ϵ > 0 . A transport-defective pair has W 1 ( p , q ; C ) > τ W 1 \mathrm{W}_1(p,q;C) > \tau_{\mathrm{W1}} W 1 ( p , q ; C ) > τ W1 , the sense of the false-pass bound of Sec. V B, and a transport-good pair has W 1 ≤ τ W 1 \mathrm{W}_1 \le\tau_{\mathrm{W1}} W 1 ≤ τ W1 . The acceptance set of the escalation rule is { W 1 ≤ τ W 1 − η } ∪ { τ W 1 − η < W 1 ≤ τ W 1 , M M D k 2 ≤ τ M M D } \{\mathrm{W}_1 \le\tau_{\mathrm{W1}}-\eta\}\cup\{\tau_{\mathrm{W1}}-\eta< \mathrm{W}_1 \le\tau_{\mathrm{W1}},\mathrm{MMD}_k^2 \le\tau_{\mathrm{MMD}}\} { W 1 ≤ τ W1 − η } ∪ { τ W1 − η < W 1 ≤ τ W1 , MMD k 2 ≤ τ MMD } . A PASS certifies the transport tolerance everywhere and the discrepancy tolerance only inside the window, so a pair with W 1 ≤ τ W 1 − η \mathrm{W}_1 \le\tau_{\mathrm{W1}}-\eta W 1 ≤ τ W1 − η and M M D k 2 > τ M M D \mathrm{MMD}_k^2 > \tau_{\mathrm{MMD}} MMD k 2 > τ MMD is admitted without a discrepancy evaluation (on the mode-collapse family of Sec. V D at n = 3 n = 3 n = 3 , W 1 = 0.03 \mathrm{W}_1 = 0.03 W 1 = 0.03 gives M M D k 2 = 1.5 × 10 − 4 \mathrm{MMD}_k^2 = 1.5\times10^{-4} MMD k 2 = 1.5 × 1 0 − 4 , which D 4 D_4 D 4 admits at η = 0.01 \eta= 0.01 η = 0.01 ). The tolerance-mismatch region of Secs. V D–V G is transport-good by construction, so rejecting it enforces the discrepancy tolerance on the window; it corrects no transport false pass. Every error rate below is stated against the transport-defective or the jointly-good population.
The protocol
Algorithm 1 summarizes the escalation-based certification protocol (D 4 D_4 D 4 ), which takes as input the target distribution p p p , the generator output q q q , the ground-cost matrix C C C , kernel function k k k , thresholds ( τ W 1 , τ M M D ) (\tau_{\mathrm{W1}},\tau_{\mathrm{MMD}}) ( τ W1 , τ MMD ) , and a margin parameter η > 0 \eta> 0 η > 0 . Outside the escalation window a single W 1 \mathrm{W}_1 W 1 evaluation suffices. On the PASS side of the window, τ W 1 − η < w ≤ τ W 1 \tau_{\mathrm{W1}}-\eta< w \le\tau_{\mathrm{W1}} τ W1 − η < w ≤ τ W1 , an MMD evaluation is added and the protocol declares PASS only if both metrics agree on PASS. This trades a small increase in false-fail rate for a quadratic suppression of the false-pass rate for defects with a margin (Sec. V B). The window is one-sided because a run on the FAIL side is rejected by the AND rule irrespective of the MMD vote. A symmetric window would only add MMD evaluations that cannot change the verdict, and letting an MMD=PASS override a W 1 \mathrm{W}_1 W 1 =FAIL would reintroduce the false passes the rule exists to exclude. When p p p and q q q are estimated from measurement shots, the protocol samples in two adaptive stages: the W 1 \mathrm{W}_1 W 1 stage draws its shots unconditionally, and the MMD stage draws a fresh, disjoint set only if the escalation window is entered. The second set is independent of the first regardless of when it is drawn. The sampling scheme itself enforces the independent-metric-errors hypothesis of the false-pass bound of Sec. V B, and the expected shot cost is m m m plus m m m times the escalation frequency, against 2 m 2m 2 m if both sets were always drawn. Section V C calibrates the residual dependence when both metrics are computed from the same data, as in the exact-statevector campaign below.
The MMD verdict v M M D v_{\mathrm{MMD}} v MMD thresholds the squared statistic against τ M M D \tau_{\mathrm{MMD}} τ MMD (Sec. II B).
Algorithm 1 Bi-metric escalation protocol (D 4 D_4 D 4 )
Require: Target p p p , generator output q q q (when sampled, the MMD sample set is drawn fresh and disjoint from the W 1 \mathrm{W}_1 W 1 set, and only upon escalation), cost matrix C C C , kernel k k k , thresholds ( τ W 1 , τ M M D ) (\tau_{\mathrm{W1}},\tau_{\mathrm{MMD}}) ( τ W1 , τ MMD ) , escalation window η \eta η
Ensure: Decision { P A S S , F A I L } \{\mathrm{PASS},\mathrm{FAIL}\} { PASS , FAIL }
Compute w ← W 1 ( p , q ; C ) w \leftarrow\mathrm{W}_1(p,q;C) w ← W 1 ( p , q ; C )
if w ≤ τ W 1 − η w \le\tau_{\mathrm{W1}}-\eta w ≤ τ W1 − η then — Clear PASS: no escalation
return PASS
else if w > τ W 1 w > \tau_{\mathrm{W1}} w > τ W1 then — Clear FAIL: W 1 \mathrm{W}_1 W 1 vote already fails
return FAIL
end if
Compute M M D k 2 ( p , q ) \mathrm{MMD}_k^2(p,q) MMD k 2 ( p , q ) — Escalation window τ W 1 − η < w ≤ τ W 1 \tau_{\mathrm{W1}}-\eta< w \le\tau_{\mathrm{W1}} τ W1 − η < w ≤ τ W1
v M M D ← 1 [ M M D k 2 ( p , q ) ≤ τ M M D ] v_{\mathrm{MMD}} \leftarrow\mathbf{1}[\mathrm{MMD}_k^2(p,q) \le\tau_{\mathrm{MMD}}] v MMD ← 1 [ MMD k 2 ( p , q ) ≤ τ MMD ]
if v M M D = 1 v_{\mathrm{MMD}} = 1 v MMD = 1 then — W 1 \mathrm{W}_1 W 1 already PASS; conservative AND
return PASS
else
return FAIL
end if
Combined false-pass bound
Both metrics concentrate at Hoeffding rates when estimated from m m m samples (the shot count S S S of the experiments below), and Appendix B1 states the two bounds as Propositions 6 and 8. The empirical transport distance deviates from its mean by more than t t t with probability at most 2 exp ( − 2 m t 2 / D max 2 ) 2\exp(-2mt^2/D_{\max}^2) 2 exp ( − 2 m t 2 / D m a x 2 ) when the target is known exactly and the generator is sampled, the observation model of Sec. V J and Appendix B4. Its bias is at most D max ∑ x q ( x ) / ( 2 m ) D_{\max}\sum_x\sqrt{q(x)}/(2\sqrt{m}) D m a x ∑ x q ( x ) / ( 2 m ) over the sampled generator output q q q , which is O ( 1 ) O(1) O ( 1 ) in n n n when that output is sparse, as for a near-exact fit of the clean B1 target, and at most 2 n / 2 2^n/2 2 n /2 for a full-support generator [30 ] ; the sparsity of the exact target does not enter. When both distributions are sampled, with m m m shots each, the exponent halves to − m t 2 / D max 2 -mt^2/D_{\max}^2 − m t 2 / D m a x 2 and the bias is the sum of the two one-sample terms (Proposition 6); the constants below are stated for the one-sample model, and Appendix B2 gives their two-sample counterparts. For our ground costs D max = O ( n ) D_{\max}=O(n) D m a x = O ( n ) , so the fluctuation complexity is O ( n 2 / t 2 ) O(n^2/t^2) O ( n 2 / t 2 ) in either model. The unbiased MMD
U U U -statistic exceeds its population value by more than t t t with probability at most exp ( − m t 2 / 16 K max 2 ) \exp(-mt^2/16K_{\max}^2) exp ( − m t 2 /16 K m a x 2 ) [7 ] , so with K max = 1 K_{\max}=1 K m a x = 1 the discrepancy needs m ≥ 16 ln ( 1 / δ ) / t 2 m \ge16\ln(1/\delta)/t^2 m ≥ 16 ln ( 1/ δ ) / t 2 shots, independent of n n n and of the support size N N N . Table IX fixes which observation model each part of the paper uses.
Table IX. Observation models. “Exact” means the statevector distribution; “one-sample” means the target is exact and the generator is sampled, with W1 and the MMD U-statistic on disjoint draws; “two-sample” means both are sampled.
To compose the two channels, we consider the escalation protocol D 4 D_4 D 4 (Algorithm 1) with escalation window η > 0 \eta> 0 η > 0 applied to the empirical estimate w ^ = W 1 ( p ^ , q ^ ; C ) \widehat{w} = W_1(\widehat{p},\widehat{q};C) w = W 1 ( p , q ; C ) , and we assume that W1 and MMD are computed from disjoint sample sets of the generator and, when the target is sampled, of the target as well, so that their estimation errors are independent. We write δ W 1 ( t ) = 2 exp ( − 2 m t 2 / D max 2 ) \delta_{W_1}(t) = 2\exp(-2mt^2/D_{\max}^2) δ W 1 ( t ) = 2 exp ( − 2 m t 2 / D m a x 2 ) and δ M M D ( t ) = exp ( − m t 2 / 16 K max 2 ) \delta_{\mathrm{MMD}}(t) = \exp(-mt^2/16K_{\max}^2) δ MMD ( t ) = exp ( − m t 2 /16 K m a x 2 ) for the tail probabilities of Propositions 6 and 8 at margin t t t , with δ W 1 ( t ) : = 1 \delta_{W_1}(t) := 1 δ W 1 ( t ) := 1 when the formula exceeds one, and we define the margin zone M η = { ( p , q ) : ∣ W 1 ( p , q ) − τ W 1 ∣ < η } \mathcal{M}_{\eta} = \{(p,q) : |W_1(p,q)-\tau_{W_1}| < \eta\} M η = {( p , q ) : ∣ W 1 ( p , q ) − τ W 1 ∣ < η } , the set of pairs whose true distance lies within one window of the threshold (the same η \eta η sets the window and the zone). The transport estimate needs no bias correction on the false-pass side: W 1 ( p , ⋅ ; C ) W_1(p,\cdot;C) W 1 ( p , ⋅ ; C ) is convex, so by Jensen’s inequality W ‾ m ≥ W 1 ( p , q ; C ) \overline{W}_m \ge W_1(p,q;C) W m ≥ W 1 ( p , q ; C ) , and a downward deviation from the true value is a downward deviation of at least the same size from the mean, which Proposition 6 bounds (the factor 2 in δ W 1 \delta_{W_1} δ W 1 is the two-sided tail and could be dropped). Outside the margin zone the escalation adds no protection to the transport vote, and for ( p , q ) ∉ M η (p,q) \notin\mathcal{M}_{\eta} ( p , q ) ∈ / M η the false-pass probability obeys
Pr ( D 4 false-pass ) ≤ δ W 1 ( η ) , (12) \Pr(D_4\ \text{false-pass}) \le\delta_{W_1}(\eta), \tag*{(12)} Pr ( D 4 false-pass ) ≤ δ W 1 ( η ) , (12)
which is identical to the bound for the W1-only rule D 1 D_1 D 1 . Inside the margin zone, in contrast, the escalation improves the bound: if ( p , q ) (p,q) ( p , q ) is defective with defect size Δ = W 1 ( p , q ; C ) − τ W 1 > 0 \Delta= W_1(p,q;C) - \tau_{W_1} > 0 Δ = W 1 ( p , q ; C ) − τ W 1 > 0 and jointly detectable, in the sense that M M D k 2 ( p , q ) ≥ τ M M D + ξ \mathrm{MMD}^2_k(p,q) \ge\tau_{\mathrm{MMD}} + \xi MMD k 2 ( p , q ) ≥ τ MMD + ξ for an MMD margin ξ > 0 \xi> 0 ξ > 0 , then
Pr ( D 4 false-pass ) ≤ δ W 1 ( η ) ⏟ escalation miss + δ W 1 ( Δ ) δ M M D ( ξ ) ⏟ joint misvote , (13) \Pr(D_4\ \text{false-pass}) \le\underbrace{\delta_{W_1}(\eta)}_{\text{escalation miss}} + \underbrace{\delta_{W_1}(\Delta)\,\delta_{\mathrm{MMD}}(\xi)}_{\text{joint misvote}}, \tag*{(13)} Pr ( D 4 false-pass ) ≤ escalation miss δ W 1 ( η ) + joint misvote δ W 1 ( Δ ) δ MMD ( ξ ) , (13)
as we prove in Appendix B. The escalation-miss term is present because D 4 D_4 D 4 triggers on the empirical w ^ \widehat{w} w : a defective pair whose estimate falls at w ^ ≤ τ W 1 − η \widehat{w} \le\tau_{W_1} - \eta w ≤ τ W 1 − η is certified without ever consulting MMD, an event of probability at most δ W 1 ( η ) \delta_{W_1}(\eta) δ W 1 ( η ) that enters the bound additively. For defects with margin Δ ≥ t \Delta\ge t Δ ≥ t the second term is the quadratic product δ W 1 ( t ) δ M M D ( ξ ) \delta_{W_1}(t)\,\delta_{\mathrm{MMD}}(\xi) δ W 1 ( t ) δ MMD ( ξ ) , and widening the window η \eta η , at the cost of more frequent escalation, suppresses the miss term below the product (Appendix B 2). The joint-detectability hypothesis is necessary, since a defect with M M D k 2 ( p , q ) < τ M M D \mathrm{MMD}^2_k(p,q) < \tau_{\mathrm{MMD}} MMD k 2 ( p , q ) < τ MMD is not caught by the escalation and the second term then loses its MMD factor. The constants of Eq. (13) are loose at realistic shot budgets, and Sec. V J reports the measured rates.
A D 4 D_4 D 4 PASS decomposes into two events with different structure: in the clear case a defective pair escapes on a single W1 error and the term enters additively; in the escalated case both votes must fail at once on disjoint samples and the term is a product. The union is the acceptance region of an intersection–union test [31 ] , and Appendix B carries out the bookkeeping.
The bound translates into an explicit quadratic improvement on false pass over single-metric certification. We fix a target false-pass level δ \delta δ for defects with margin Δ ≥ t \Delta\ge t Δ ≥ t , allocate δ W 1 ( t ) = δ M M D ( ξ ) = δ / 2 \delta_{W_1}(t) = \delta_{\mathrm{MMD}}(\xi) = \sqrt{\delta/2} δ W 1 ( t ) = δ MMD ( ξ ) = δ /2 to the joint-misvote term, and size the escalation window so that δ W 1 ( η ) = δ / 2 \delta_{W_1}(\eta) = \delta/2 δ W 1 ( η ) = δ /2 , which yields
η = t ln ( 4 / δ ) ln ( 2 2 / δ ) . (14) \eta= t\sqrt{\frac{\ln(4/\delta)}{\ln(2\sqrt{2}/\sqrt{\delta})}}. \tag*{(14)} η = t ln ( 2 2 / δ ) ln ( 4/ δ ) . (14)
With this choice Eq. (13) gives Pr ( D 4 false-pass ) ≤ δ / 2 + δ / 2 = δ \Pr(D_4\ \text{false-pass}) \le\delta/2 + \delta/2 = \delta Pr ( D 4 false-pass ) ≤ δ /2 + δ /2 = δ , while the W1-only rule D 1 D_1 D 1 at the same per-metric sample size achieves only δ W 1 ( t ) = δ / 2 \delta_{W_1}(t) = \sqrt{\delta/2} δ W 1 ( t ) = δ /2 . For δ = 0.05 \delta= 0.05 δ = 0.05 the window ratio (14) is η ≈ 1.31 t \eta\approx1.31\,t η ≈ 1.31 t , and the comparison reads D 4 ≤ 0.05 D_4 \le0.05 D 4 ≤ 0.05 vs. D 1 ≤ 0.158 D_1 \le0.158 D 1 ≤ 0.158 . This comparison uses upper bounds at equal sample size per metric, both conventions that flatter D 4 D_4 D 4 : the realized rates of both rules sit far below either bound at the budgets we test (Sec. V J), and Appendix B 2 restates the comparison at equal total budget, where the advantage survives with a stiffened precondition. A second baseline is the single-metric rule at the shifted threshold, D 1 ( τ W 1 − η ) D_1(\tau_{W_1}-\eta) D 1 ( τ W 1 − η ) , whose PASS region is contained in that of D 4 D_4 D 4 : it has the lower false-pass probability, at most δ W 1 ( η + Δ ) \delta_{W_1}(\eta+\Delta) δ W 1 ( η + Δ ) , and D 4 D_4 D 4 differs from it only by admitting, inside the window, the runs on which the discrepancy vote also passes. The escalation therefore sits between the two single-metric rules D 1 ( τ W 1 ) D_1(\tau_{W_1}) D 1 ( τ W 1 ) and D 1 ( τ W 1 − η ) D_1(\tau_{W_1}-\eta) D 1 ( τ W 1 − η ) , trading false passes against false fails, and Sec. V J reports both sides of the operating characteristic against both. In return, the AND rule incurs a false-fail probability on a jointly-good pair. For good-side margins, W 1 ( p , q ; C ) ≤ τ W 1 − t ′ W_1(p,q;C) \le\tau_{W_1}-t' W 1 ( p , q ; C ) ≤ τ W 1 − t ′ and M M D k 2 ( p , q ) ≤ τ M M D − ξ ′ \mathrm{MMD}^2_k(p,q) \le\tau_{\mathrm{MMD}}-\xi' MMD k 2 ( p , q ) ≤ τ MMD − ξ ′ , with t ′ > b m t' > b_m t ′ > b m , it is at most δ W 1 ( t ′ − b m ) + δ M M D ( ξ ′ ) \delta_{W_1}(t'-b_m)+\delta_{\mathrm{MMD}}(\xi') δ W 1 ( t ′ − b m ) + δ MMD ( ξ ′ ) , where b m b_m b m is the upward bias of Eq. (42) . The primed margins are distances to the thresholds from the good side, distinct from the defective-side margins Δ \Delta Δ and ξ \xi ξ . The bias pushes toward FAIL, so on this side it must be subtracted from the margin. The composition itself is the intersection–union test of Berger and Hsu [31 ] , and the contribution is its instantiation with explicit constants on { 0 , 1 } n \{0,1\}^n { 0 , 1 } n , the separation of the window from the analysis margins, and the cross-run agreement statistic of Sec. V C.
Same-data estimation and threshold calibration
The independence assumption holds by construction when the two metrics are estimated from disjoint sample sets, which is the protocol’s specification for sampled targets (§ 4 A). When both metrics are computed from the same data, or exactly as in the present statevector campaign, independence can fail, and the joint-misvote term is replaced by a bounded conditional error: if Pr ( v M M D = 1 ∣ v W 1 = 1 ) ≤ α C E < 1 \Pr(v_{\mathrm{MMD}}=1\mid v_{\mathrm{W1}}=1)\leq\alpha_{\mathrm{CE}}<1 Pr ( v MMD = 1 ∣ v W1 = 1 ) ≤ α CE < 1 on the defective pair, then
Pr ( D 4 false-pass ) ≤ δ W 1 ( η ) + α C E δ W 1 ( Δ ) . (15) \Pr(D_4\ \text{false-pass})\leq\delta_{\mathrm{W1}}(\eta)+\alpha_{\mathrm{CE}}\,\delta_{\mathrm{W1}}(\Delta). \tag*{(15)} Pr ( D 4 false-pass ) ≤ δ W1 ( η ) + α CE δ W1 ( Δ ) . (15)
The escalation-miss term δ W 1 ( η ) \delta_{\mathrm{W1}}(\eta) δ W1 ( η ) depends only on the single W1 estimate and is unaffected by α C E \alpha_{\mathrm{CE}} α CE . The campaign cannot estimate α C E \alpha_{\mathrm{CE}} α CE for a fixed pair, because its scores are exact and carry no estimation error, so we record instead a descriptive statistic a r u n a_{\mathrm{run}} a run : the fraction of the perturbed (ϵ > 0 \epsilon>0 ϵ > 0 ) B1 runs whose W1 vote passes on which the MMD vote also passes. It counts the co-occurrence of exact-value votes across different runs, some of them transport-good. A bound on the conditional misvote probability of a specified transport-defective pair would need repeated same-data sampling of that pair; the finite-shot measurements of § 4 J and V K bear on it. At the benchmark tolerance M M D k 2 ≤ 10 − 2 \mathrm{MMD}_k^2\leq10^{-2} MMD k 2 ≤ 1 0 − 2 , and still at 10 − 3 10^{-3} 1 0 − 3 , a r u n = 1.000 a_{\mathrm{run}}=1.000 a run = 1.000 at every scale and under both references: the escalation never overturns a W1 verdict, which is the conclusion § 4 J reaches independently by direct measurement and which Theorem 4 makes exact at 10 − 2 10^{-2} 1 0 − 2 for the population values of every pair. At the certification-grade M M D k 2 ≤ 10 − 4 \mathrm{MMD}_k^2\leq10^{-4} MMD k 2 ≤ 1 0 − 4 the statistic falls below one, by different amounts under the two references. Under the trained-on convention it reads 0.932, 0.931, 0.972, 0.977, 1.000, 1.000 at n = 3 n=3 n = 3 through 8, an improvement factor of at most 1.07, because a well-fitted defective run has a small residual against the target it was trained on and passes both votes. Against the clean reference it reads 0.634 at n = 3 n=3 n = 3 , 0.984 at n = 4 n=4 n = 4 , and 1.000 thereafter: the escalation’s overturns concentrate at the smallest scales, where the window analysis of § 4 D finds the window open, and vanish at the scales where it is closed. The statistic agrees in pattern with the finite-shot measurement and with the closed-form window. Realizing the margin-zone bound (13) thus requires calibrating ( τ W 1 , τ M M D ) (\tau_{\mathrm{W1}},\tau_{\mathrm{MMD}}) ( τ W1 , τ MMD ) at the operating point in use.
The tolerance-mismatch window on the bitstring space
On a finite outcome space W 1 \mathrm{W1} W1 and the MMD of a characteristic kernel induce the same topology, so the two agree on which sequences converge and differ in the shape of their acceptance regions and in the scale dependence of their sensitivity, the finite-space form of the comparisons drawn in the IPM hierarchy [32 ] and of the continuous-space witnesses of the GAN literature [33 ] . The inclusion R b i ⊇ R W 1 \mathcal{R}_{\mathrm{bi}}\supseteq\mathcal{R}_{\mathrm{W1}} R bi ⊇ R W1 between the rejection regions of the escalation and of the transport rule holds for any AND rule and carries no information by itself. Our contribution is the size of the difference at a stated operating point: for two defect families both metrics admit closed forms, so the tolerance-mismatch region, the set of generators that the transport threshold admits and the discrepancy threshold rejects, and the window of tolerance ratios in which it is non-empty can be computed exactly, and so can the single transport threshold that would reproduce the two-metric verdict on each family (§ 4 G).
Theorem 2 (Tolerance-mismatch window on { 0 , 1 } n \{0,1\}^n { 0 , 1 } n ). Let n ≥ 2 n\geq2 n ≥ 2 , let k ( x , y ) = exp ( − γ d H ( x , y ) ) k(x,y)=\exp(-\gamma d_{\mathrm{H}}(x,y)) k ( x , y ) = exp ( − γ d H ( x , y )) be the RBF–Hamming kernel with bandwidth γ > 0 \gamma>0 γ > 0 , and fix a mode x 0 ∈ { 0 , 1 } n x_0\in\{0,1\}^n x 0 ∈ { 0 , 1 } n . Take the uniform target p = Unif ( { 0 , 1 } n ) p=\operatorname{Unif}(\{0,1\}^n) p = Unif ({ 0 , 1 } n ) and the mode-contaminated generator family
q λ = ( 1 − λ ) Unif ( { 0 , 1 } n ) + λ δ x 0 , λ ∈ ( 0 , 1 ] , (16) q_\lambda=(1-\lambda)\operatorname{Unif}(\{0,1\}^n)+\lambda\delta_{x_0},\qquad\lambda\in(0,1], \tag*{(16)} q λ = ( 1 − λ ) Unif ({ 0 , 1 } n ) + λ δ x 0 , λ ∈ ( 0 , 1 ] , (16) which places an excess weight λ \lambda λ on the single mode x 0 x_0 x 0 and interpolates between the target at λ → 0 \lambda\to0 λ → 0 and complete collapse onto x 0 x_0 x 0 at λ = 1 \lambda=1 λ = 1 , the sense in which we call the family mode collapse. Then the following identities are exact for every λ \lambda λ , n ≥ 2 n\geq2 n ≥ 2 , and γ > 0 \gamma>0 γ > 0 :
(a) W 1 ( p , q λ ; C 1 ) = λ n / 2 \mathrm{W}_1(p,q_\lambda;C_1)=\lambda n/2 W 1 ( p , q λ ; C 1 ) = λn /2 under the Hamming cost;
(b) T V ( p , q λ ) = λ ( 1 − 2 − n ) \mathrm{TV}(p,q_\lambda)=\lambda(1-2^{-n}) TV ( p , q λ ) = λ ( 1 − 2 − n ) ;
(c) M M D k 2 ( p , q λ ) = λ 2 [ 1 − ( 1 + e − γ 2 ) n ] \mathrm{MMD}_k^2(p,q_\lambda)=\lambda^2\left[1-\left(\frac{1+e^{-\gamma}}{2}\right)^n\right] MMD k 2 ( p , q λ ) = λ 2 [ 1 − ( 2 1 + e − γ ) n ] .
Fix a W1 threshold τ W 1 > 0 \tau_{\mathrm{W1}}>0 τ W1 > 0 and an escalation window η ∈ ( 0 , τ W 1 ) \eta\in(0,\tau_{\mathrm{W1}}) η ∈ ( 0 , τ W1 ) . For every
λ ∈ ( 2 ( τ W 1 − η ) n , 2 τ W 1 n ] , (17) \lambda\in\left(\frac{2(\tau_{\mathrm{W1}}-\eta)}{n},\frac{2\tau_{\mathrm{W1}}}{n}\right], \tag*{(17)} λ ∈ ( n 2 ( τ W1 − η ) , n 2 τ W1 ] , (17) claim (a) gives W 1 ( p , q λ ; C 1 ) = λ n / 2 ∈ ( τ W 1 − η , τ W 1 ] \mathrm{W}_1(p,q_\lambda;C_1)=\lambda n/2\in(\tau_{\mathrm{W1}}-\eta,\tau_{\mathrm{W1}}] W 1 ( p , q λ ; C 1 ) = λn /2 ∈ ( τ W1 − η , τ W1 ] , so the witness is certified PASS by W1 yet lies in the one-sided escalation window of D 4 D_4 D 4 , triggering the MMD vote. Every such λ \lambda λ lies in the bi-metric rejection region but not the W1-only region provided the detectability condition, evaluated at the interval’s left endpoint,
τ M M D < ( 2 ( τ W 1 − η ) n ) 2 [ 1 − ( 1 + e − γ 2 ) n ] (18) \tau_{\mathrm{MMD}}<\left(\frac{2(\tau_{\mathrm{W1}}-\eta)}{n}\right)^2\left[1-\left(\frac{1+e^{-\gamma}}{2}\right)^n\right] \tag*{(18)} τ MMD < ( n 2 ( τ W1 − η ) ) 2 [ 1 − ( 2 1 + e − γ ) n ] (18) holds, in which case the tolerance-mismatch region R b i ∖ R W 1 \mathcal{R}_{\mathrm{bi}}\setminus\mathcal{R}_{\mathrm{W1}} R bi ∖ R W1 is non-empty,
R b i ⊃ R W 1 , (19) \mathcal{R}_{\mathrm{bi}}\supset\mathcal{R}_{\mathrm{W1}}, \tag*{(19)} R bi ⊃ R W1 , (19) with R W 1 = { q : W 1 ( p , q ; C 1 ) > τ W 1 } \mathcal{R}_{\mathrm{W1}}=\{q:\mathrm{W}_1(p,q;C_1)>\tau_{\mathrm{W1}}\} R W1 = { q : W 1 ( p , q ; C 1 ) > τ W1 } and R b i = { q : D 4 ( q ) = F A I L } \mathcal{R}_{\mathrm{bi}}=\{q:D_4(q)=\mathrm{FAIL}\} R bi = { q : D 4 ( q ) = FAIL } . Since M M D k 2 ( p , q λ ) \mathrm{MMD}_k^2(p,q_\lambda) MMD k 2 ( p , q λ ) is increasing in λ \lambda λ by claim (c), condition (18) at the left endpoint certifies the whole interval; it holds for every window η ≤ η max ( n ) \eta\leq\eta_{\max}(n) η ≤ η m a x ( n ) , with
η max ( n ) = τ W 1 − n 2 τ M M D 1 − ( ( 1 + e − γ ) / 2 ) n . (20) \eta_{\max}(n)=\tau_{\mathrm{W1}}-\frac{n}{2}\sqrt{\frac{\tau_{\mathrm{MMD}}}{1-\left((1+e^{-\gamma})/2\right)^n}}. \tag*{(20)} η m a x ( n ) = τ W1 − 2 n 1 − ( ( 1 + e − γ ) /2 ) n τ MMD . (20)
At the certification-grade tolerance τ M M D = 10 − 4 \tau_{\mathrm{MMD}} = 10^{-4} τ MMD = 1 0 − 4 , with τ W 1 = 0.05 \tau_{\mathrm{W}_1} = 0.05 τ W 1 = 0.05 and γ = 1 / n \gamma= 1/n γ = 1/ n , η max ( n ) = 0.025 , 0.017 , 0.009 , 0.001 \eta_{\max}(n) = 0.025, 0.017, 0.009, 0.001 η m a x ( n ) = 0.025 , 0.017 , 0.009 , 0.001 at n = 3 , 4 , 5 , 6 n = 3, 4, 5, 6 n = 3 , 4 , 5 , 6 , an open window of relative width 0.51 , 0.35 , 0.19 , 0.03 0.51, 0.35, 0.19, 0.03 0.51 , 0.35 , 0.19 , 0.03 times τ W 1 \tau_{\mathrm{W}_1} τ W 1 . The separation is robust for n ≤ 5 n \le5 n ≤ 5 and narrows at n = 6 n = 6 n = 6 to the knife-edge λ = 2 τ W 1 / n \lambda= 2\tau_{\mathrm{W}_1}/n λ = 2 τ W 1 / n , the limit η → 0 \eta\to0 η → 0 . For n ≥ 7 n \ge7 n ≥ 7 the window closes, η max ( n ) = − 0.007 , − 0.015 \eta_{\max}(n) = -0.007, -0.015 η m a x ( n ) = − 0.007 , − 0.015 at n = 7 , 8 n = 7, 8 n = 7 , 8 , because the right-hand side of (18) at η = 0 \eta= 0 η = 0 drops below τ M M D = 10 − 4 \tau_{\mathrm{MMD}} = 10^{-4} τ MMD = 1 0 − 4 (to 7.8 × 10 − 5 7.8 \times10^{-5} 7.8 × 1 0 − 5 and 6.0 × 10 − 5 6.0 \times10^{-5} 6.0 × 1 0 − 5 ). At fixed certification-grade thresholds the mode-collapse mismatch region is an n ≤ 6 n \le6 n ≤ 6 phenomenon, and reopening it at larger n n n requires lowering ρ = τ M M D / τ W 1 \rho= \sqrt{\tau_{\mathrm{MMD}}/\tau_{\mathrm{W}_1}} ρ = τ MMD / τ W 1 , by tightening τ M M D \tau_{\mathrm{MMD}} τ MMD or loosening τ W 1 \tau_{\mathrm{W}_1} τ W 1 .
The proof, which evaluates the transport plan, the total variation and the mean embedding of the witness in closed form, is given in Appendix B.
The mismatch originates in a linear-versus-quadratic asymmetry: the W1 _1 1 penalty is linear in the contamination while the MMD penalty is quadratic, so that at the margin the detectable defect scales as Θ ( τ W 1 2 / n 2 ) \Theta(\tau_{\mathrm{W}_1}^{2}/n^{2}) Θ ( τ W 1 2 / n 2 ) and the witness is detected only while τ W 1 ≳ ( n / 2 ) τ M M D \tau_{\mathrm{W}_1} \gtrsim(n/2)\sqrt{\tau_{\mathrm{MMD}}} τ W 1 ≳ ( n /2 ) τ MMD . At η = 0 \eta= 0 η = 0 the ratio of the right-hand side of Eq. (18) to τ M M D \tau_{\mathrm{MMD}} τ MMD falls from 4.1 4.1 4.1 at n = 3 n = 3 n = 3 to 1.1 1.1 1.1 at n = 6 n = 6 n = 6 and below 1 1 1 at n = 7 , 8 n = 7, 8 n = 7 , 8 .
Two defect families
The mode-collapse construction of Theorem 2 is the first of two defect families (Family 1 ), and the second exploits a different geometric mechanism.
Family 2 (correlation shifts) is constructed as follows: we let p p p be uniform on { 0 , 1 } n \{0,1\}^{n} { 0 , 1 } n and take
q ( x ) = 2 − n [ 1 + β ( − 1 ) x 0 ⊕ x 1 ] , β ∈ ( 0 , 1 ) , (21) q(x) = 2^{-n}\left[1 + \beta(-1)^{x_0 \oplus x_1}\right], \qquad\beta\in(0,1), \tag*{(21)} q ( x ) = 2 − n [ 1 + β ( − 1 ) x 0 ⊕ x 1 ] , β ∈ ( 0 , 1 ) , (21)
which matches every single-qubit marginal of p p p but introduces a parity correlation β \beta β between qubits 0 0 0 and 1 1 1 ; it is normalized because ∑ x ( − 1 ) x 0 ⊕ x 1 = 0 \sum_x (-1)^{x_0 \oplus x_1} = 0 ∑ x ( − 1 ) x 0 ⊕ x 1 = 0 . The discrepant mass has total-variation weight TV ( p , q ) = β / 2 \operatorname{TV}(p,q) = \beta/2 TV ( p , q ) = β /2 , and W 1 ( p , q ; C 1 ) = β / 2 \mathrm{W}_1(p,q;\mathcal{C}_1) = \beta/2 W 1 ( p , q ; C 1 ) = β /2 exactly. Both bounds are explicit: the excess sits on the 2 n − 1 2^{n-1} 2 n − 1 strings with x 0 = x 1 x_0 = x_1 x 0 = x 1 and the deficit on their complements, and flipping either x 0 x_0 x 0 or x 1 x_1 x 1 carries excess to deficit at Hamming distance 1 1 1 , so W 1 ≤ β / 2 \mathrm{W}_1 \le\beta/2 W 1 ≤ β /2 ; the potential f ( x ) = 1 2 ( − 1 ) x 0 ⊕ x 1 f(x) = \frac{1}{2}(-1)^{x_0 \oplus x_1} f ( x ) = 2 1 ( − 1 ) x 0 ⊕ x 1 is C 1 \mathcal{C}_1 C 1 -Lipschitz and attains ⟨ f , q − p ⟩ = β / 2 \langle f,q-p\rangle= \beta/2 ⟨ f , q − p ⟩ = β /2 , so W 1 ≥ β / 2 \mathrm{W}_1 \ge\beta/2 W 1 ≥ β /2 by Kantorovich duality (2) (both bounds are given in full in Appendix B). The RBF–Hamming MMD, in turn, reads exactly
MMD k 2 ( p , q ) = β 2 ( 1 − e − γ ) 2 ( 1 + e − γ ) n − 2 2 − n , (22) \operatorname{MMD}_{k}^{2}(p,q) = \beta^{2}(1-e^{-\gamma})^{2}(1+e^{-\gamma})^{n-2}2^{-n}, \tag*{(22)} MMD k 2 ( p , q ) = β 2 ( 1 − e − γ ) 2 ( 1 + e − γ ) n − 2 2 − n , (22)
a quadratic form in β \beta β with no higher-order correction, since MMD k 2 \operatorname{MMD}_{k}^{2} MMD k 2 is quadratic in p − q p-q p − q . Rudolph et al. [17 ] study the same low-order structure as a trainability question, the vanishing of the loss gradient when the kernel weights high-order correlations weakly; here the quantity compared with a fixed threshold is the population
discrepancy itself, and the question is which defects a verdict at that threshold covers. The transport metric does see the defect, since W 1 = β / 2 \mathrm{W}_1 = \beta/2 W 1 = β /2 with every single-qubit marginal matched, but its response is linear in β \beta β . The discrepancy’s response is quadratic in β \beta β and carries the bracket A n A_n A n , which at γ = 1 / n \gamma= 1/n γ = 1/ n decays as n − 2 n^{-2} n − 2 (Eq. (24) ). The mismatch is thus a threshold effect of the two scalings. Choosing β < 2 τ W 1 \beta< 2\tau_{\mathrm{W}_1} β < 2 τ W 1 but β 2 > τ M M D 2 n / [ ( 1 − e − γ ) 2 ( 1 + e − γ ) n − 2 ] \beta^2 > \tau_{\mathrm{MMD}}2^n/[(1-e^{-\gamma})^2(1+e^{-\gamma})^{n-2}] β 2 > τ MMD 2 n / [( 1 − e − γ ) 2 ( 1 + e − γ ) n − 2 ] yields a tolerance mismatch that MMD catches. Because both metrics are again exact, the window admits the same closed form as Eq. (20) , now reading η max ( n ) = τ W 1 − 1 2 τ M M D / A n \eta_{\max}(n) = \tau_{\mathrm{W}_1} - \frac{1}{2}\sqrt{\tau_{\mathrm{MMD}}/A_n} η m a x ( n ) = τ W 1 − 2 1 τ MMD / A n with A n = ( 1 − e − γ ) 2 ( 1 + e − γ ) n − 2 2 − n A_n = (1-e^{-\gamma})^2(1+e^{-\gamma})^{n-2}2^{-n} A n = ( 1 − e − γ ) 2 ( 1 + e − γ ) n − 2 2 − n . It closes far sooner than the mode-collapse window: at ( τ W 1 , τ M M D , γ ) = ( 0.05 , 10 − 4 , 1 / n ) (\tau_{\mathrm{W}_1},\tau_{\mathrm{MMD}},\gamma) = (0.05,10^{-4},1/n) ( τ W 1 , τ MMD , γ ) = ( 0.05 , 1 0 − 4 , 1/ n ) it is open only at n = 3 n = 3 n = 3 , with relative width 0.24 0.24 0.24 , and already negative at n = 4 n = 4 n = 4 . The reason is the size of the two brackets per unit transport distance. The squared discrepancy are 4 Φ n W 1 2 / n 2 4\Phi_n\mathrm{W}_1^2/n^2 4 Φ n W 1 2 / n 2 on mode collapse and 4 A n W 1 2 4A_n\mathrm{W}_1^2 4 A n W 1 2 on correlation shift, both of order n − 2 n^{-2} n − 2 at γ = 1 / n \gamma= 1/n γ = 1/ n since ( 1 + e − γ ) n − 2 2 − n (1+e^{-\gamma})^{n-2}2^{-n} ( 1 + e − γ ) n − 2 2 − n cancels the explicit 2 − n 2^{-n} 2 − n . The correlation signal is the smaller by the factor Φ n / ( n 2 A n ) ≈ 2.6 \Phi_n/(n^2 A_n) \approx2.6 Φ n / ( n 2 A n ) ≈ 2.6 , the square of the threshold ratio of Sec. V G. Correlation shifts are therefore the harder of the two families to certify against, and Sec. V L trains generators into both.
Both windows depend on the operating point through a single dimensionless ratio, so the scope claim holds for every pair ( τ W 1 , τ M M D ) (\tau_{\mathrm{W}_1},\tau_{\mathrm{MMD}}) ( τ W 1 , τ MMD ) with the same ratio. Writing ρ = τ M M D / τ W 1 \rho= \sqrt{\tau_{\mathrm{MMD}}/\tau_{\mathrm{W}_1}} ρ = τ MMD / τ W 1 , and Φ n = 1 − ( ( 1 + e − γ ) / 2 ) n \Phi_n = 1 - ((1+e^{-\gamma})/2)^n Φ n = 1 − (( 1 + e − γ ) /2 ) n for the mode-collapse bracket of Theorem 2, the two closed forms collapse to
η max τ W 1 = 1 − n 2 ρ Φ n and 1 − 1 2 ρ A n , (23) \frac{\eta_{\max}}{\tau_{\mathrm{W}_1}} = 1 - \frac{n}{2}\frac{\rho}{\sqrt{\Phi_n}}
\qquad\text{and}\qquad
1 - \frac{1}{2}\frac{\rho}{\sqrt{A_n}}, \tag*{(23)} τ W 1 η m a x = 1 − 2 n Φ n ρ and 1 − 2 1 A n ρ , (23)
so each family closes above a critical ratio, ρ 1 ∗ ( n ) = 2 Φ n / n \rho_1^*(n) = 2\sqrt{\Phi_n}/n ρ 1 ∗ ( n ) = 2 Φ n / n and ρ 2 ∗ ( n ) = 2 A n \rho_2^*(n) = 2\sqrt{A_n} ρ 2 ∗ ( n ) = 2 A n , however τ W 1 \tau_{\mathrm{W}_1} τ W 1 , τ M M D \tau_{\mathrm{MMD}} τ MMD are chosen separately. Figure 4 maps that plane, and the certification-grade point of Sec. V L sits at ρ = 0.2 \rho= 0.2 ρ = 0.2 .
Figure 4. Escalation window in the ( n , ρ ) (n,\rho) ( n , ρ ) plane, ρ = τ M M D / τ W 1 \rho=\sqrt{\tau_{\mathrm{MMD}}/\tau_{\mathrm{W1}}} ρ = τ MMD / τ W1 , the only combination through which the operating point enters either closed form. Shading: the deployment window, the minimum over the two families. Lines: closure boundaries for mode collapse (solid) and correlation shift (dashed). The star marks the certification-grade point, where the deployment window is open at n = 3 n=3 n = 3 alone.
A practitioner certifying a generator does not know which defect is present, so the escalation window must be sized against the minimum over the families considered plausible, a requirement stricter than either family alone. At ρ = 0.2 \rho= 0.2 ρ = 0.2 that minimum is + 0.24 +0.24 + 0.24 at n = 3 n = 3 n = 3 and negative from n = 4 n = 4 n = 4 onward, the correlation-shift family binding at every scale. The mode-collapse figures, open through n = 6 n = 6 n = 6 , therefore describe the best case. Against an unknown defect family at this operating point the usable window exists at n = 3 n = 3 n = 3 alone, and reopening it at larger n n n requires lowering ρ \rho ρ , that is, tightening τ M M D \tau_{\mathrm{MMD}} τ MMD faster than τ W 1 2 \tau_{\mathrm{W}_1}^{2} τ W 1 2 .
At the operational bandwidth γ = 1 / n \gamma= 1/n γ = 1/ n the correlation signal (22) decays with n n n , and the rate follows from Eq. (22) directly. Substituting γ = 1 / n \gamma= 1/n γ = 1/ n and expanding, ( 1 − e − 1 / n ) 2 = n − 2 + O ( n − 3 ) (1-e^{-1/n})^2 = n^{-2}+O(n^{-3}) ( 1 − e − 1/ n ) 2 = n − 2 + O ( n − 3 ) while ( 1 + e − 1 / n ) n − 2 2 − n → e − 1 / 2 / 4 (1+e^{-1/n})^{n-2}2^{-n} \to e^{-1/2}/4 ( 1 + e − 1/ n ) n − 2 2 − n → e − 1/2 /4 , so
MMD k 2 ( p , q ) = e − 1 / 2 4 β 2 n 2 ( 1 + O ( n − 1 ) ) , (24) \operatorname{MMD}_{k}^{2}(p,q) = \frac{e^{-1/2}}{4}\frac{\beta^{2}}{n^{2}}\left(1+O(n^{-1})\right), \tag*{(24)} MMD k 2 ( p , q ) = 4 e − 1/2 n 2 β 2 ( 1 + O ( n − 1 ) ) , (24)
that is Θ ( β 2 / n 2 ) \Theta(\beta^2/n^2) Θ ( β 2 / n 2 ) with the explicit constant e − 1 / 2 / 4 ≈ 0.1516 e^{-1/2}/4 \approx0.1516 e − 1/2 /4 ≈ 0.1516 . Detecting a fixed-β \beta β correlation defect therefore requires an MMD tolerance that tightens as n − 2 n^{-2} n − 2 , and since the sample complexity of Proposition 8 scales as t − 2 t^{-2} t − 2 at additive accuracy t t t , taking t = Θ ( β 2 / n 2 ) t=\Theta(\beta^2/n^2) t = Θ ( β 2 / n 2 ) gives a sufficient shot count m = O ( n 4 / β 4 ) m=O(n^4/\beta^4) m = O ( n 4 / β 4 ) under the Hoeffding-type certificate of Proposition 8. This is a range-based upper bound on what the certificate requires, which a variance-adaptive bound could tighten (§ 4 J). The dimension-free complexity of Proposition 8 is stated at fixed additive accuracy and does not bound the cost of resolving a defect whose signal itself shrinks with n n n .
One response to the n − 2 n^{-2} n − 2 decay is to repair the kernel, and the repair is exactly computable. Augment the RBF–Hamming kernel with the parity feature of the suspected pair, k ′ ( x , y ) = k ( x , y ) + φ ( x ) φ ( y ) k'(x,y)=k(x,y)+\varphi(x)\varphi(y) k ′ ( x , y ) = k ( x , y ) + φ ( x ) φ ( y ) with φ ( x ) = ( − 1 ) x 0 ⊕ x 1 \varphi(x)=(-1)^{x_0\oplus x_1} φ ( x ) = ( − 1 ) x 0 ⊕ x 1 . For the Family 2 witness, E q [ φ ] − E p [ φ ] = β \mathbb{E}_q[\varphi]-\mathbb{E}_p[\varphi]=\beta E q [ φ ] − E p [ φ ] = β exactly, so the augmentation adds β 2 \beta^2 β 2 to M M D k ′ 2 \mathrm{MMD}^2_{k'} MMD k ′ 2 at every n n n , with no 2 − n 2^{-n} 2 − n suppression, and the correlation-shift window reopens at every scale at the cost of widening the kernel range from [ 0 , 1 ] [0,1] [ 0 , 1 ] to [ − 1 , 2 ] [-1,2] [ − 1 , 2 ] . If the defect axis is unknown, the same construction summed over all ( n 2 ) \binom{n}{2} ( 2 n ) pairs still adds exactly β 2 \beta^2 β 2 , because only the true pair’s feature pairs with the discrepancy (any other parity is uncorrelated with it under p p p and q q q alike; we verified both statements numerically at n = 3 , 4 , 6 n=3,4,6 n = 3 , 4 , 6 ). The cost, however, does not vanish: the kernel range widens to [ 1 − ( n 2 ) , 1 + ( n 2 ) ] [1-\binom{n}{2},1+\binom{n}{2}] [ 1 − ( 2 n ) , 1 + ( 2 n ) ] , so the range width K max K_{\max} K m a x of Proposition 8 grows as 2 ( n 2 ) 2^{\binom{n}{2}} 2 ( 2 n ) , and the certificate scales as K max 2 / t 2 = Θ ( n 4 / β 4 ) K_{\max}^2/t^2=\Theta(n^4/\beta^4) K m a x 2 / t 2 = Θ ( n 4 / β 4 ) , the same scaling as Eq. (24) gives for the un-augmented kernel, and normalizing the sum instead returns the n − 2 n^{-2} n − 2 decay to the signal. Certification against a specified defect family thus costs O ( 1 ) O(1) O ( 1 ) in the kernel range; against an unspecified one the certified shot cost is O ( n 4 / β 4 ) O(n^4/\beta^4) O ( n 4 / β 4 ) under every repair of this kind, and § 4 F extends the limitation to every normalized kernel.
A third mechanism, ground-cost mismatch, has neither an exact witness nor a trained-generator demonstration: the transport acceptance regions { q : W 1 ( p , q ; C ) ≤ τ } \{q:W_1(p,q;C)\le\tau\} { q : W 1 ( p , q ; C ) ≤ τ } for two ground costs are generically different polytopes in Δ N \Delta_N Δ N , and the MMD acceptance region is an RKHS ball, sensitive to the kernel bandwidth where the polytope is sensitive to the cost. § 4 § 4.1.11 asks whether trained generators occupy the first two families.
Correlation order and the redundancy boundary
The two families are instances of a structure that the RBF–Hamming kernel makes exact. Write χ S ( x ) = ( − 1 ) ∑ i ∈ S x i \chi_S(x)=(-1)^{\sum_{i\in S}x_i} χ S ( x ) = ( − 1 ) ∑ i ∈ S x i for the parity of a subset S ⊆ { 1 , … , n } S\subseteq\{1,\ldots,n\} S ⊆ { 1 , … , n } , so that the 2 n 2^n 2 n parities are orthogonal under the uniform distribution u u u . The kernel factorizes over qubits, e − γ d H ( x , y ) = ∏ i [ c + b ( − 1 ) x i ⊕ y i ] e^{-\gamma d_H(x,y)}=\prod_i[c+b(-1)^{x_i\oplus y_i}] e − γ d H ( x , y ) = ∏ i [ c + b ( − 1 ) x i ⊕ y i ] with c = ( 1 + e − γ ) / 2 c=(1+e^{-\gamma})/2 c = ( 1 + e − γ ) /2 and b = ( 1 − e − γ ) / 2 b=(1-e^{-\gamma})/2 b = ( 1 − e − γ ) /2 , and expanding the product gives
k ( x , y ) = ∑ S c n − ∣ S ∣ b ∣ S ∣ χ S ( x ) χ S ( y ) , k(x,y)=\sum_S c^{n-|S|}b^{|S|}\chi_S(x)\chi_S(y), k ( x , y ) = S ∑ c n − ∣ S ∣ b ∣ S ∣ χ S ( x ) χ S ( y ) ,
M M D k 2 ( p , q ) = ∑ S ≠ ∅ c n − ∣ S ∣ b ∣ S ∣ ( E p χ S − E q χ S ) 2 . (25) \mathrm{MMD}^2_k(p,q)=\sum_{S\ne\varnothing}c^{n-|S|}b^{|S|}\left(\mathbb{E}_p\chi_S-\mathbb{E}_q\chi_S\right)^2. \tag*{(25)} MMD k 2 ( p , q ) = S = ∅ ∑ c n − ∣ S ∣ b ∣ S ∣ ( E p χ S − E q χ S ) 2 . (25)
The discrepancy is a weighted sum of squared parity-moment mismatches, with a weight that falls geometrically in the order ∣ S ∣ |S| ∣ S ∣ , and the expansion states exactly which correlations the diagnostic measures and how strongly. Theorem 2 and Family 2 read off from it: mode collapse excites every parity, λ 2 ∑ S ≠ ∅ c n − ∣ S ∣ b ∣ S ∣ = λ 2 ( 1 − c n ) = λ 2 Φ n \lambda^2\sum_{S\ne\varnothing}c^{n-|S|}b^{|S|}=\lambda^2(1-c^n)=\lambda^2\Phi_n λ 2 ∑ S = ∅ c n − ∣ S ∣ b ∣ S ∣ = λ 2 ( 1 − c n ) = λ 2 Φ n , and the correlation shift excites a single parity of order two, β 2 c n − 2 b 2 = β 2 A n \beta^2c^{n-2}b^2=\beta^2A_n β 2 c n − 2 b 2 = β 2 A n . The expansion also bounds, for any kernel, how evenly that sensitivity can be spread.
Theorem 3 (Parity defects). Let u u u be uniform on { 0 , 1 } n \{0,1\}^n { 0 , 1 } n and, for ∅ ≠ S ⊆ { 1 , … , n } \varnothing\ne S\subseteq\{1,\ldots,n\} ∅ = S ⊆ { 1 , … , n } and 0 < β < 1 0<\beta<1 0 < β < 1 , let q S , β ( x ) = 2 − n [ 1 + β χ S ( x ) ] q_{S,\beta}(x)=2^{-n}[1+\beta\chi_S(x)] q S , β ( x ) = 2 − n [ 1 + β χ S ( x )] . The distribution q S , β q_{S,\beta} q S , β has full support, matches every marginal of u u u on fewer than ∣ S ∣ |S| ∣ S ∣ qubits, and is the Born distribution of a circuit of n n n single-qubit rotations and ∣ S ∣ − 1 |S|-1 ∣ S ∣ − 1 CNOT gates.
(a) TV ( u , q S , β ) = W 1 ( u , q S , β ; C 1 ) = β / 2 \operatorname{TV}(u,q_{S,\beta})=W_1(u,q_{S,\beta};C_1)=\beta/2 TV ( u , q S , β ) = W 1 ( u , q S , β ; C 1 ) = β /2 for every S S S .
(b) For the RBF–Hamming kernel, M M D k 2 ( u , q S , β ) = β 2 c n − ∣ S ∣ b ∣ S ∣ \mathrm{MMD}^2_k(u,q_{S,\beta})=\beta^2c^{n-|S|}b^{|S|} MMD k 2 ( u , q S , β ) = β 2 c n − ∣ S ∣ b ∣ S ∣ , which at γ = 1 / n \gamma=1/n γ = 1/ n and fixed r = ∣ S ∣ r=|S| r = ∣ S ∣ equals e − 1 / 2 β 2 ( 2 n ) − r [ 1 + O ( n − 1 ) ] e^{-1/2}\beta^2(2n)^{-r}[1+O(n^{-1})] e − 1/2 β 2 ( 2 n ) − r [ 1 + O ( n − 1 )] .
(c) For every positive-semidefinite kernel with k ( x , x ) = 1 k(x,x)=1 k ( x , x ) = 1 ,
min S ≠ ∅ M M D k 2 ( u , q S , β ) ≤ β 2 2 n − 1 . (26) \min_{S\ne\varnothing}\mathrm{MMD}^2_k(u,q_{S,\beta})\le\frac{\beta^2}{2^n-1}. \tag*{(26)} S = ∅ min MMD k 2 ( u , q S , β ) ≤ 2 n − 1 β 2 . (26)
with equality for the centered kernel k ∗ ( x , y ) = ( 2 n 1 [ x = y ] − 1 ) / ( 2 n − 1 ) k_\ast(x,y)=(2^n\mathbb{1}[x=y]-1)/(2^n-1) k ∗ ( x , y ) = ( 2 n 1 [ x = y ] − 1 ) / ( 2 n − 1 ) , so that the supremum of the left-hand side over normalized kernels equals the right-hand side.
The proof is in Appendix B3; part (c) rests on the identity ∑ S a S = 1 \sum_S a_S=1 ∑ S a S = 1 for the parity weights a S = 2 − 2 n χ S K χ S a_S=2^{-2n}\chi_S K\chi_S a S = 2 − 2 n χ S K χ S of any unit-diagonal kernel, so that a kernel has one unit of sensitivity to distribute over 2 n − 1 2^n-1 2 n − 1 correlation directions. Figure 5 shows the hierarchy of part (b): at the same transport distance β / 2 \beta/2 β /2 , a single-qubit bias costs the discrepancy Θ ( β 2 / n ) \Theta(\beta^2/n) Θ ( β 2 / n ) , a two-qubit correlation Θ ( β 2 / n 2 ) \Theta(\beta^2/n^2) Θ ( β 2 / n 2 ) , an r r r -qubit correlation Θ ( β 2 / n r ) \Theta(\beta^2/n^r) Θ ( β 2 / n r ) , and the full parity β 2 b n \beta^2 b^n β 2 b n . Part (c) says that no bandwidth, and no other normalized kernel, gives a scale-independent signal on all of these alternatives: the centered kernel k ∗ k_\ast k ∗ flattens the hierarchy, responding equally to every nonempty parity, but only at the level β 2 / ( 2 n − 1 ) \beta^2/(2^n-1) β 2 / ( 2 n − 1 ) , and for every fixed kernel some parity defect of transport size β / 2 \beta/2 β /2 has squared discrepancy at most that value, which for β = 0.1 \beta=0.1 β = 0.1 lies below the certification-grade tolerance from n = 7 n=7 n = 7 on. Two qualifications bound the statement. It concerns the population signal, and it fixes the kernel before the defect direction is chosen. A test that knows the plausible correlations can add their parity features, as in § 4 E; a test restricted to the ( n 2 ) \binom{n}{2} ( 2 n ) pairs needs only O ( β − 2 log ( n / δ ) ) O(\beta^{-2}\log(n/\delta)) O ( β − 2 log ( n / δ )) samples by a union of Hoeffding bounds; and adaptive kernel selection [34 ] is the general tool. The certification consequence is that the correlation orders a verdict is meant to cover must be named.
Figure 5. Squared discrepancy of the parity defect q S , β q_{S,\beta} q S , β against the uniform target at fixed transport distance β / 2 \beta/2 β /2 , by correlation order r = ∣ S ∣ r=|S| r = ∣ S ∣ , for the RBF–Hamming kernel at γ = 1 / n \gamma=1/n γ = 1/ n (Theorem 3 (b)). Dotted: the minimax reference β 2 / ( 2 n − 1 ) \beta^2/(2^n-1) β 2 / ( 2 n − 1 ) of Eq. (26) , the largest worst-parity signal any normalized kernel attains; it bounds the minimum over S S S from above. Dashed: the Hamming-cost ceiling 2 ( 1 − e − γ ) W 1 2 2(1-e^{-\gamma})W_1^2 2 ( 1 − e − γ ) W 1 2 of Eq. (27) .
Closely related statements exist for kernel training losses, and the differences fix what part (c) adds. Kurkin et al. [18 ] write the discrepancy of any stationary bounded kernel on { 0 , 1 } n \{0,1\}^n { 0 , 1 } n as an average over the kernel’s spectral measure of squared parity-moment mismatches, of which Eq. (25) is the RBF–Hamming instance. They show that the Gaussian-kernel discrepancy of a particular pair is O ( 2 − n ) O(2^{-n}) O ( 2 − n ) at total variation of order one. They also show that for every sufficiently large n n n some pair at total variation one, built from a random split of the cube, has discrepancy at most 2 − b n 2^{-bn} 2 − bn , with an unspecified constant b > 0 b>0 b > 0 , for every stationary characteristic kernel whose spectral measure is a probability measure. This last condition is the unit-diagonal normalization, under which the bound is meaningful since MMD C k 2 = C MMD k 2 \operatorname{MMD}^2_{Ck}=C\operatorname{MMD}^2_k MMD C k 2 = C MMD k 2 . Part (c) differs in three respects. The kernel class is every positive-semidefinite kernel with unit diagonal, stationary or not, so that nonstationary kernels and the normalization of any sum of kernels are covered. The alternatives form the explicit family q S , β q_{S,\beta} q S , β , each member at the same transport and total-variation distance β / 2 \beta/2 β /2 from the target and prepared by a shallow circuit carrying a single parity. The bound is an exact value with an attaining kernel. Its quantifier order is, in addition, the one certification needs: the kernel is fixed before the defect is chosen. In Ref. [18 ] the exponentially small pair is fixed first and then defeats every kernel in its class, so neither statement contains the other.
In the other direction, the kernel bounds the discrepancy by the transport distance, which says exactly when the second channel can contribute nothing.
Theorem 4 (Population redundancy boundary). For the Hamming cost C 1 C_1 C 1 and the RBF–Hamming kernel, every pair p , q ∈ Δ N p,q\in\Delta_N p , q ∈ Δ N satisfies
MMD k 2 ( p , q ) ≤ 2 ( 1 − e − γ ) W 1 ( p , q ; C 1 ) 2 , (27) \operatorname{MMD}^2_k(p,q)\le2(1-e^{-\gamma})W_1(p,q;C_1)^2, \tag*{(27)} MMD k 2 ( p , q ) ≤ 2 ( 1 − e − γ ) W 1 ( p , q ; C 1 ) 2 , (27) with equality for p = δ x p=\delta_x p = δ x and q = ( 1 − w ) δ x + w δ y q=(1-w)\delta_x+w\delta_y q = ( 1 − w ) δ x + w δ y at Hamming-adjacent x , y x,y x , y and 0 ≤ w ≤ 1 0\le w\le1 0 ≤ w ≤ 1 . Consequently, for 0 < τ W 1 ≤ 1 0<\tau_{W_1}\le1 0 < τ W 1 ≤ 1 , the discrepancy vote on the population values is redundant on the whole transport acceptance region, so that D 4 D_4 D 4 and D 1 ( τ W 1 ) D_1(\tau_{W_1}) D 1 ( τ W 1 ) evaluated on the same pair ( p , q ) (p,q) ( p , q ) , exactly agree for every window η > 0 \eta>0 η > 0 , if and only if τ M M D ≥ 2 ( 1 − e − γ ) τ W 1 2 \tau_{\mathrm{MMD}}\ge2(1-e^{-\gamma})\tau_{W_1}^2 τ MMD ≥ 2 ( 1 − e − γ ) τ W 1 2 .
The inequality is an instance of the general bound MMD k ( p , q ) ≤ W 1 ( p , q ; ρ k ) \operatorname{MMD}_k(p,q)\le W_1(p,q;\rho_k) MMD k ( p , q ) ≤ W 1 ( p , q ; ρ k ) of Theorem 21(i) of Sripe-rumbudur et al. [32 ] , in which the ground metric is the kernel-induced distance ρ k ( x , y ) = ∥ k ( ⋅ , x ) − k ( ⋅ , y ) ∥ H = 2 ( 1 − e − γ d H ( x , y ) ) \rho_k(x,y)=\lVert k(\cdot,x)-k(\cdot,y)\rVert_{\mathcal H}=\sqrt{2(1-e^{-\gamma d_H(x,y)})} ρ k ( x , y ) = ∥ k ( ⋅ , x ) − k ( ⋅ , y ) ∥ H = 2 ( 1 − e − γ d H ( x , y ) ) : for integer d ≥ 1 d\ge1 d ≥ 1 , 1 − e − γ ≤ d ( 1 − e − γ ) 1-e^{-\gamma}\le d(1-e^{-\gamma}) 1 − e − γ ≤ d ( 1 − e − γ ) and d ≤ d \sqrt d\le d d ≤ d give ρ k ≤ 2 ( 1 − e − γ ) d H \rho_k\le\sqrt{2(1-e^{-\gamma})}d_H ρ k ≤ 2 ( 1 − e − γ ) d H , and Eq. (27) follows. The proof in Appendix B3 takes the same route, a shortest Hamming path in feature space, where adjacent strings are 2 ( 1 − e − γ ) \sqrt{2(1-e^{-\gamma})} 2 ( 1 − e − γ ) apart, through an optimal coupling. What the theorem adds is the sharpness, since the adjacent pair attains the constant, and the decision statement, the equivalence on the tolerance ratio; it is the elementary converse, on the Hamming cube, of the results that control transport distances by kernel norms on R d \mathbb{R}^d R d [35 , 36 ] . At τ W 1 = 0.05 \tau_{W_1}=0.05 τ W 1 = 0.05 and γ = 1 / n \gamma=1/n γ = 1/ n the boundary is 1.4 × 10 − 3 1.4\times10^{-3} 1.4 × 1 0 − 3 , 1.1 × 10 − 3 1.1\times10^{-3} 1.1 × 1 0 − 3 , 9.1 × 10 − 4 9.1 \times10^{-4} 9.1 × 1 0 − 4 , 7.7 × 10 − 4 7.7 \times10^{-4} 7.7 × 1 0 − 4 , 6.7 × 10 − 4 6.7 \times10^{-4} 6.7 × 1 0 − 4 and 5.9 × 10 − 4 5.9 \times10^{-4} 5.9 × 1 0 − 4 at n = 3 n = 3 n = 3 –8 8 8 . The loose tolerance τ M M D = 10 − 2 \tau_{\mathrm{MMD}} = 10^{-2} τ MMD = 1 0 − 2 makes the discrepancy redundant for every pair of distributions at every scale we study, the intermediate 10 − 3 10^{-3} 1 0 − 3 from n = 5 n = 5 n = 5 on, and the certification-grade 10 − 4 10^{-4} 1 0 − 4 at none. This is the population form of the finding of § 4 C and V J that the escalation changes no verdict at the loose tolerance. The theorem does not extend by itself to the sampled protocol, whose discrepancy vote comes from an independent draw. It guarantees every pair whose population values lie in the window a population margin τ M M D − 2 ( 1 − e − γ ) τ W 1 2 ≥ 0.0086 \tau_{\mathrm{MMD}} - 2(1 - e^{-\gamma})\tau_{\mathrm{W1}}^2 \ge0.0086 τ MMD − 2 ( 1 − e − γ ) τ W1 2 ≥ 0.0086 below the loose tolerance, but a draw that lands in the window does not certify that premise for the pair it came from. For a pair that does meet it, the Hoeffding tail of Proposition 8 at that margin is 0.63 at S = 10 5 S = 10^5 S = 1 0 5 and falls to 0.05 only near S = 6 × 10 5 S = 6 \times10^5 S = 6 × 1 0 5 . The finite-shot agreement is therefore a measurement: on the 41,771 in-window draws of the fixed-threshold test of § 4 K, taken on a fixed panel of 600 fits, the independent discrepancy vote failed in none (one-sided 95% upper limit 7 × 10 − 5 7 \times10^{-5} 7 × 1 0 − 5 per in-window draw). The adjacent pair is also the extremal defect for the window analysis: its discrepancy per unit transport distance is the largest any pair attains, so its window, η max a d j = τ W 1 − τ M M D / ( 2 ( 1 − e − γ ) ) = 0.037 \eta_{\max}^{\mathrm{adj}} = \tau_{\mathrm{W1}} - \sqrt{\tau_{\mathrm{MMD}}/(2(1-e^{-\gamma}))} = 0.037 η m a x adj = τ W1 − τ MMD / ( 2 ( 1 − e − γ )) = 0.037 to 0.029 at n = 3 n = 3 n = 3 –8 8 8 for the certification-grade point, contains the window of every defect family. Both statements are for the Hamming cost; under C 2 C_2 C 2 and C 3 C_3 C 3 the constant in Eq. (27) changes with the cost of an adjacent flip.
Matched tolerances and the equivalent transport threshold
Both windows are statements about a fixed pair ( τ W 1 , τ M M D ) (\tau_{\mathrm{W1}}, \tau_{\mathrm{MMD}}) ( τ W1 , τ MMD ) , so we ask which single transport threshold reproduces the escalation verdict on each family. On each family the escalation acts as a transport-only rule with an effective threshold: by Theorem 2 it rejects exactly the λ \lambda λ with λ n / 2 > T 1 \lambda n/2 > T_1 λn /2 > T 1 , and by Eq. (22) exactly the β \beta β with β / 2 > T 2 \beta/2 > T_2 β /2 > T 2 , where
T i = min { τ W 1 , max { τ W 1 − η , b i } } , i = 1 , 2 , T_i = \min\{\tau_{\mathrm{W1}}, \max\{\tau_{\mathrm{W1}}-\eta, b_i\}\}, \qquad i=1,2, T i = min { τ W1 , max { τ W1 − η , b i }} , i = 1 , 2 ,
b 1 = n 2 τ M M D Φ n , b 2 = 1 2 τ M M D A n , (28) b_1 = \frac{n}{2}\sqrt{\frac{\tau_{\mathrm{MMD}}}{\Phi_n}}, \qquad b_2 = \frac{1}{2}\sqrt{\frac{\tau_{\mathrm{MMD}}}{A_n}}, \tag*{(28)} b 1 = 2 n Φ n τ MMD , b 2 = 2 1 A n τ MMD , (28)
and b i b_i b i is the transport distance at which the family’s discrepancy crosses τ M M D \tau_{\mathrm{MMD}} τ MMD , the MMD-implied transport threshold. The outer minimum records that a clear transport FAIL receives no discrepancy vote, and the inner maximum that a clear transport PASS receives none either. Since η max , i = τ W 1 − b i \eta_{\max,i} = \tau_{\mathrm{W1}} - b_i η m a x , i = τ W1 − b i , the window of family i i i is open at η \eta η exactly when b i ≤ τ W 1 − η b_i \le\tau_{\mathrm{W1}}-\eta b i ≤ τ W1 − η , and then T i = τ W 1 − η T_i = \tau_{\mathrm{W1}}-\eta T i = τ W1 − η : inside its own window the escalation is, on that family, the shifted single-metric rule D 1 ( τ W 1 − η ) D_1(\tau_{\mathrm{W1}}-\eta) D 1 ( τ W1 − η ) , and inside the deployment window of § 4 E, η ≤ min i η max , i \eta\le\min_i\eta_{\max,i} η ≤ min i η m a x , i , one transport threshold reproduces the two-metric verdict on both families. At n = 3 n = 3 n = 3 and η = 0.005 \eta= 0.005 η = 0.005 , for instance, b 1 = 0.025 b_1 = 0.025 b 1 = 0.025 and b 2 = 0.038 b_2 = 0.038 b 2 = 0.038 both lie below τ W 1 − η = 0.045 \tau_{\mathrm{W1}}-\eta= 0.045 τ W1 − η = 0.045 , and T 1 = T 2 = 0.045 T_1 = T_2 = 0.045 T 1 = T 2 = 0.045 . Two notions must be kept apart here. The window of family i i i is open when b i ≤ τ W 1 − η b_i \le\tau_{\mathrm{W1}}-\eta b i ≤ τ W1 − η , so that every in-window member of the family is rejected and T i = τ W 1 − η T_i = \tau_{\mathrm{W1}}-\eta T i = τ W1 − η . The mismatch interval [ max { τ W 1 − η , b i } , τ W 1 ] [\max\{\tau_{\mathrm{W1}}-\eta,b_i\},\tau_{\mathrm{W1}}] [ max { τ W1 − η , b i } , τ W1 ] is non-empty under the weaker condition b i < τ W 1 b_i < \tau_{\mathrm{W1}} b i < τ W1 , which also holds for windows wider than η max , i \eta_{\max,i} η m a x , i . We use “open” in the first sense throughout, and the shifted-threshold statement holds in that sense. When both windows are closed, b i > τ W 1 b_i > \tau_{\mathrm{W1}} b i > τ W1 and T 1 = T 2 = τ W 1 T_1 = T_2 = \tau_{\mathrm{W1}} T 1 = T 2 = τ W1 , so the escalation is D 1 ( τ W 1 ) D_1(\tau_{\mathrm{W1}}) D 1 ( τ W1 ) on both. The two effective thresholds differ only for windows wider than the narrower family’s, η > η max , 2 \eta> \eta_{\max,2} η > η m a x , 2 , and then at most by the ratio of the MMD-implied thresholds,
b 2 b 1 = Φ n n 2 A n = ρ 1 ∗ ( n ) ρ 2 ∗ ( n ) = 1.54 , 1.56 , 1.57 , 1.57 , 1.58 , 1.58 (29) \begin{aligned}
\frac{b_2}{b_1} = \sqrt{\frac{\Phi_n}{n^2 A_n}} = \frac{\rho_1^*(n)}{\rho_2^*(n)} \\
&= 1.54,\ 1.56,\ 1.57,\ 1.57,\ 1.58,\ 1.58 \tag*{(29)}
\end{aligned} b 1 b 2 = n 2 A n Φ n = ρ 2 ∗ ( n ) ρ 1 ∗ ( n ) = 1.54 , 1.56 , 1.57 , 1.57 , 1.58 , 1.58 (29)
at n = 3 n = 3 n = 3 –8 8 8 , with limit 2 e 1 / 2 − 1 ≈ 1.61 2\sqrt{e^{1/2}-1} \approx1.61 2 e 1/2 − 1 ≈ 1.61 : relative to W 1 \mathrm{W}_1 W 1 , the discrepancy is more sensitive to mode collapse than to correlation shift by this factor at every n n n , and one discrepancy tolerance is equivalent to transport tolerances that differ between the families by it. The factor enters through the two transport responses, W 1 \mathrm{W}_1 W 1 per unit λ \lambda λ on Family 1 and per unit β \beta β on Family 2 (n / 2 n/2 n /2 and 1 / 2 1/2 1/2 under C 1 C_1 C 1 ), and so depends on the ground cost. Under C 2 C_2 C 2 with the campaign weights it is 1.52–1.58. Under C 3 C_3 C 3 with parity weight 0.5, where the responses are n / 2 + ( n − 1 ) / 4 n/2 + (n-1)/4 n /2 + ( n − 1 ) /4 and 3 / 4 3/4 3/4 (verified by exact LP), it is 1.73 at n = 3 n = 3 n = 3 falling to 1.65 at n = 8 n = 8 n = 8 . The factor is thus of the same size under every ground cost we use. Matching the tolerances on Family 1, so that both metrics reject the same λ \lambda λ , means τ M M D = ( 2 τ W 1 / n ) 2 Φ n \tau_{\mathrm{MMD}} = (2\tau_{\mathrm{W1}}/n)^2\Phi_n τ MMD = ( 2 τ W1 / n ) 2 Φ n , the closing value of Eq. (20) ; the Family-1 window then has zero width by construction, and the Family-2 window is closed at every n n n , with relative width − 0.54 -0.54 − 0.54 to − 0.58 -0.58 − 0.58 for n = 3 n = 3 n = 3 –8 8 8 . Matching on Family 2 instead, τ M M D = 4 τ W 1 2 A n \tau_{\mathrm{MMD}} = 4\tau_{\mathrm{W1}}^2 A_n τ MMD = 4 τ W1 2 A n , closes that window and leaves the Family-1 window open at every n n n with the n n n -independent relative width 1 − n 2 A n / Φ n = 0.35 1-\sqrt{n^2A_n/\Phi_n}=0.35 1 − n 2 A n / Φ n = 0.35 –0.37 0.37 0.37 . The open window at the certification-grade point is therefore a calibration statement: at ρ = 0.2 \rho= 0.2 ρ = 0.2 the discrepancy threshold is tighter than the transport threshold on mode collapse and looser on correlation shift. The factor (29) between the two MMD-implied thresholds does not depend on the calibration: one discrepancy tolerance assigns it to both families at once, so that a certifier who wants a given transport tolerance enforced on both must name the family.
Bi-metric disagreement
At n ≤ 6 n \le6 n ≤ 6 the disagreement rate is essentially constant at 3.8%–5.2% (Table X , Figure 6 ) and dominated by W 1 = P A S S \mathrm{W}_1=\mathrm{PASS} W 1 = PASS / M M D = F A I L \mathrm{MMD}=\mathrm{FAIL} MMD = FAIL : a population that W 1 \mathrm{W}_1 W 1 -only certification would admit and that does not shrink with scale. For n = 7 , 8 n = 7,8 n = 7 , 8 the total rises to 19.8% and 46.7%, entirely in the opposite pattern (82 and 196 W₁F/MMDP runs, in which the transport fit fails while the discrepancy passes), and the W₁P/MMDF count falls to 1 and 0. In contrast, against the clean reference the joint-failure population dominates at every scale, because a run trained on a perturbed target stays far from the clean one however well it fits. The same transport threshold is used for benchmarks B1, B2, and B3, whose attainable fits differ, so that PASS/FAIL rates are comparable across them. A persistent flagged population at the certification-grade tolerance is what the window analysis of Sec. V D predicts. The counts pool all three benchmarks, including the synthetic near-uniform B3.
Figure 6. W₁/MMD disagreement vs. qubit count at τ W 1 = 0.05 \tau_{\mathrm{W}_1}=0.05 τ W 1 = 0.05 and M M D k 2 ≤ 10 − 4 \mathrm{MMD}^{2}_{k}\leq10^{-4} MMD k 2 ≤ 1 0 − 4 . Bars: the two patterns against the trained-on reference; lines: totals for both references. The rise at n = 7 , 8 n=7,8 n = 7 , 8 is entirely W₁=FAIL / MMD=PASS.
Table X. W₁/MMD disagreement at τ W 1 = 0.05 \tau_{\mathrm{W}_1}=0.05 τ W 1 = 0.05 and the certification-grade M M D k 2 ≤ 10 − 4 \mathrm{MMD}^{2}_{k}\leq10^{-4} MMD k 2 ≤ 1 0 − 4 , trained-on reference. “Dis.” is the disagreement rate in percent of the 420 runs at each n n n , the next three columns are run counts, and “Dis.cl” repeats the rate against the clean reference. W₁P/MMDF dominates through n = 6 n=6 n = 6 and inverts at n = 7 , 8 n=7,8 n = 7 , 8 .
The W₁P/MMDF column, the population the escalation is designed to catch, holds 16, 19, 17, 11 runs at n = 3 n=3 n = 3 –6; at n = 3 n=3 n = 3 these are all B1 runs at the two largest perturbations. A per-run Fourier decomposition finds them no more correlation-dominated than agreeing runs, so they cannot be attributed to a single defect family of Sec. V E.
family of Sec. V E. The complementary direction appears below n = 7 n=7 n = 7 only at n = 6 n=6 n = 6 (11 runs), and the joint-failure column grows with scale (5 to 92 runs), tracking the loss of trainability.
Rank deficiency among certified runs
Among the runs the bi-metric protocol certifies as W₁-PASS and MMD-PASS at the operating point of Sec. V H, the count with rank 10 − 6 ( F c l ) < d \operatorname{rank}_{10^{-6}}(F_{\mathrm{cl}})<d rank 1 0 − 6 ( F cl ) < d is 67 of 399, 52 of 370, 58 of 341 and 107 of 332 at n = 3 , 4 , 5 , 6 n=3,4,5,6 n = 3 , 4 , 5 , 6 respectively, and 171 of 249 and 101 of 132 at n = 7 , 8 n=7,8 n = 7 , 8 . These are runs in which the converged distribution matches the target but the parameter regime concentrates its sensitivity in few directions. A plausible hypothesis is that they are fragile certifications , runs whose verdict a small parameter perturbation could break. We test it directly. Perturbing the converged parameters of certified runs (30 rank-deficient and 30 full-rank per scale, 24 random unit directions per run at ∥ δ θ ∥ ∈ { 0.02 , 0.05 , 0.1 , 0.2 } \lVert\delta\theta\rVert\in\{0.02,0.05,0.1,0.2\} ∥ δ θ ∥ ∈ { 0.02 , 0.05 , 0.1 , 0.2 } , and margin-matched between groups so that proximity to the threshold does not confound the comparison) shows the opposite of generic fragility: at n = 4 n=4 n = 4 –6 and ∥ δ θ ∥ = 0.1 \lVert\delta\theta\rVert=0.1 ∥ δ θ ∥ = 0.1 , a rank-deficient certified verdict breaks in 17–50% of random directions against 65–89% for full-rank certified runs. With Tr F c l = d \operatorname{Tr}F_{\mathrm{cl}}=d Tr F cl = d fixed by the identity of Sec. III, the isotropic mean-square response is the same for every run, so rank deficiency concentrates the sensitivity in few directions (median λ max = 19.2 \lambda_{\max}=19.2 λ m a x = 19.2 against 1.1 at n = 6 n=6 n = 6 ) and a random perturbation lands mostly in the insensitive subspace. Adversarial fragility does hold: along its top Fisher eigendirection the same rank-deficient run breaks essentially always (100% at n = 6 n=6 n = 6 , ∥ δ θ ∥ = 0.1 \lVert\delta\theta\rVert=0.1 ∥ δ θ ∥ = 0.1 , against 24% in random directions). The rank thus flags an anisotropic verdict robustness: a run it flags is vulnerable to structured drift along a few sensitive directions and robust to generic drift, and the flag complements the bi-metric verdict without overriding it.
Finite-shot behavior of the bi-metric false-pass bound
The bound of Eq. (13) is stated in the finite-sample regime with W₁ and MMD estimated from disjoint sample sets. The campaign of Sec. II H is exact-state-vector, so it never exercises the independence mechanism that makes the bound quadratic, and the calibration of Sec. V C is computed in the same-data regime the bound excludes. We therefore exercise the protocol in the regime that the bound describes. For a defective pair we draw two disjoint sets of S S S shots from the generator, evaluate W₁ on one and MMD on the other, and run both the escalation rule D 4 D_4 D 4 and the W₁-only rule D 1 D_1 D 1 , repeating 2,000 times per configuration at S = 10 2 S=10^2 S = 1 0 2 through 10 5 10^5 1 0 5 . The MMD vote uses the unbiased U U U -statistic of Proposition 8, in its one-sample form against the exact tar get. On the same draws we also evaluate the plug-in V V V -statistic, whose positive bias (Appendix B § 5.8.4 ) pushes the vote toward FAIL. Defective pairs are of two kinds: members of the mode-collapse family of Theorem 2 placed a controlled distance Δ ∈ { 0.005 , 0.02 } \Delta\in\{0.005, 0.02\} Δ ∈ { 0.005 , 0.02 } above τ W 1 \tau_{W_1} τ W 1 , and trained generators from the campaign scored against the clean benchmark target, retained when W 1 W_1 W 1 exceeds τ W 1 \tau_{W_1} τ W 1 (nine fits, three of which are the warm-restarted versions of Sec. IV B; the pair is defined by its stored distribution, so the choice does not enter the rates).
The practical case for the protocol rests on the measured false-pass rates, which separate the two threshold regimes. At the benchmark-comparison thresholds the escalation contributes essentially nothing: D 4 D_4 D 4 and D 1 D_1 D 1 return identical false-pass rates in 55 of 60 configurations, because the benchmark tolerance M M D k 2 ≤ 10 − 2 \mathrm{MMD}_k^2 \le10^{-2} MMD k 2 ≤ 1 0 − 2 is loose enough that the MMD vote passes whenever the W 1 W_1 W 1 vote does. The five exceptions are all campaign-fit pairs at S = 100 S=100 S = 100 , where the shot noise of the discrepancy estimate exceeds the margin to the loose tolerance and D 4 D_4 D 4 falls below D 1 D_1 D 1 by 0.002 to 0.009; none occurs at S ≥ 10 3 S \ge10^3 S ≥ 1 0 3 . The calibration of § 4 C agrees. At the certification-grade tolerance, on the other hand, the escalation helps: among the 29 configurations in which D 1 D_1 D 1 falsely certifies at a nonzero rate, D 4 D_4 D 4 drives the rate to zero in 2,000 repetitions in 10 (one-sided 95% upper limit 0.0015 each), and where both are nonzero the reduction is a median factor of 2.0 (range 1.2 to 55). The largest realized D 1 D_1 D 1 false-pass rate is 0.357, on a configuration where D 4 D_4 D 4 realizes 0.20, the largest D 4 D_4 D 4 rate in the regime. The median realized improvement is smaller than the factor of 3.2 that the comparison of upper bounds below Eq. (14) obtains. Against the shifted single-metric rule D 1 ( τ W 1 − η ) D_1(\tau_{W_1}-\eta) D 1 ( τ W 1 − η ) , evaluated on the same draws, the ordering on this side reverses, as the inclusion of § 4 B requires. At S = 10 3 S=10^3 S = 1 0 3 its mean false-pass rate over the certification-grade configurations is 0.006, against 0.033 for D 4 D_4 D 4 and 0.071 for D 1 ( τ W 1 ) D_1(\tau_{W_1}) D 1 ( τ W 1 ) . Among the 29 configurations with a nonzero D 1 ( τ W 1 ) D_1(\tau_{W_1}) D 1 ( τ W 1 ) rate D 4 D_4 D 4 matches D 1 ( τ W 1 − η ) D_1(\tau_{W_1}-\eta) D 1 ( τ W 1 − η ) in 11 and exceeds it in 18, by a median factor of 2.3. At S = 10 5 S=10^5 S = 1 0 5 all three rates are zero in 2,000 repetitions (upper limit 0.0015 per configuration). On the same draws the plug-in estimator reports 13 configurations driven to zero, a median factor of 9.5, and a largest D 4 D_4 D 4 rate of 0.062. The difference is the plug-in bias, which pushes defective and good generators alike toward FAIL; it returns below as a higher false-fail rate.
Across all 120 configurations these realized rates stay below the bound of Eq. (13) , with no violations and by a wide margin: the bound is non-vacuous in only 22 of the 120 configurations, because δ W 1 ( η ) \delta_{W_1}(\eta) δ W 1 ( η ) with a window of one concentration margin does not fall below unity until S ≈ 10 5 S \approx10^5 S ≈ 1 0 5 at these defect sizes, while the realized rates are already at the percent level by S = 10 3 S=10^3 S = 1 0 3 . The guarantee of Eq. (13) is correct but loose at the shot budgets a hardware experiment would plausibly use. The looseness comes from the tail bounds, since Propositions 6 and 8 are Hoeffding-type and use only the range of the statistic, while the MMD estimator’s variance at these budgets is smaller than its range by orders of magnitude. Measuring the realized standard deviation of M M D k 2 \mathrm{MMD}_k^2 MMD k 2 on the mode-collapse witness and comparing an empirical-Bernstein tail against the Hoeffding tail at matched confidence gives a ratio of 2.6 × 10 3 2.6 \times10^3 2.6 × 1 0 3 at S = 10 3 S=10^3 S = 1 0 3 , rising to 7.5 × 10 4 7.5 \times10^4 7.5 × 1 0 4 at S = 10 5 S=10^5 S = 1 0 5 for n = 3 n=3 n = 3 and to 2.0 × 10 5 2.0 \times10^5 2.0 × 1 0 5 at n = 6 n=6 n = 6 . A variance-adaptive certificate would be non-vacuous in the regime where the realized rates already sit. We have not proved one: a Bernstein bound for the degenerate U U U -statistic, with the empirical variance estimated from the same shots, is a separate technical problem.
The other half of the operating characteristic is the false-fail rate on good generators, together with the escalation frequency that sets the adaptive scheme’s extra shot cost, and we measure both on pairs that are good under both exact channels at the regime’s tolerances. In the clear region both rules are clean and cheap: false fails are zero in 2,000 repetitions by S = 10 5 S=10^5 S = 1 0 5 (upper limit 0.0015) and the escalation fires in at most a few percent of repetitions, so the expected cost of D 4 D_4 D 4 is one sample set. The cost of the conservative AND rule appears on jointly-good pairs placed in the window. At S = 10 5 S=10^5 S = 1 0 5 and certification grade the escalation fires in 59–100% of repetitions on these pairs. The realized false-fail rate of D 4 D_4 D 4 is at most 0.11, against 0.053 for D 1 ( τ W 1 ) D_1(\tau_{W_1}) D 1 ( τ W 1 ) , and where both are nonzero it exceeds D 1 D_1 D 1 ’s by a factor of 1.2 to 2.0 (2.6 with the plug-in estimator). The shifted rule D 1 ( τ W 1 − η ) D_1(\tau_{W_1}-\eta) D 1 ( τ W 1 − η ) obtains its lower false-pass rate at this cost: on the same window pairs at S = 10 5 S=10^5 S = 1 0 5 it rejects 59% to 100% of repetitions (mean 0.87), against a mean of 0.02 for D 4 D_4 D 4 and 0.01 for D 1 ( τ W 1 ) D_1(\tau_{W_1}) D 1 ( τ W 1 ) , while in the clear region the three rules agree at S = 10 5 S=10^5 S = 1 0 5 . The escalation therefore keeps the false-fail behavior of D 1 ( τ W 1 ) D_1(\tau_{W_1}) D 1 ( τ W 1 ) to within a factor of two and takes about half of the false-pass gap toward D 1 ( τ W 1 − η ) D_1(\tau_{W_1}-\eta) D 1 ( τ W 1 − η ) ; the shifted rule closes the rest of that gap with a false-fail rate forty times larger inside the window. Pairs placed in the window are rejected by the shifted rule by construction, so the deployment cost is set by how often ordinary fits land there: of the 776 unperturbed campaign fits that the transport threshold admits against the clean reference, 70 (9.0%; 8, 17, 15 and 30 at n = 5 n=5 n = 5 –8, none below n = 5 n=5 n = 5 ) have W 1 ∈ ( τ W 1 − η , τ W 1 ] W_1 \in(\tau_{W_1}-\eta,\tau_{W_1}] W 1 ∈ ( τ W 1 − η , τ W 1 ] and would be rejected outright by D 1 ( τ W 1 − η ) D_1(\tau_{W_1}-\eta) D 1 ( τ W 1 − η ) . The escalation evaluates them and, on exact values, keeps 62 and rejects 8 (1.0% of the admitted fits). At smaller budgets both rules reject borderline good runs frequently, through the same W 1 W_1 W 1 bias mechanism recorded below. On defective pairs the escalation frequency is small (median 0.00, maximum 0.32 across the 120 configurations), because most estimates fall on a clear side of the window. Four in-window witness pairs are excluded from the good-pair accounting at certification grade because their exact M M D k 2 \mathrm{MMD}_k^2 MMD k 2 exceeds the tolerance: they are the mismatch-region members of Theorem 2, and D 4 D_4 D 4 rejects them correctly.
The false-pass rate is not monotone in S S S : it is small at S = 10 2 S=10^2 S = 1 0 2 , peaks near S = 10 3 S=10^3 S = 1 0 3 or 10 4 10^4 1 0 4 , and then decays. The small-S S S suppression comes from bias: the W1 _1 1 bias term of Eq. (42) inflates w ^ \hat{w} w ^ at low shot counts and pushes borderline defective pairs across τ W 1 \tau_{\mathrm{W1}} τ W1 into a FAIL verdict reached for the wrong reason. A protocol tuned on low-shot data would inherit that bias as if it were discrimination.
What certification-grade tolerances cost in shots is quantified in Appendix B § 5.8.4 .
Fixed-threshold validation on the full-support benchmark
Every rate above was measured on pairs chosen for the purpose, exact witnesses and campaign fits selected by their defect, and the controls of Sec. IV B condition on scores. We therefore test the protocol on targets never used to set a threshold. The 600 fits of the perturbed full-support benchmark (Sec. IV C) are certified against the clean B3 target at the thresholds, kernel, window and budgets fixed above (τ W 1 = 0.05 \tau_{\mathrm{W1}} = 0.05 τ W1 = 0.05 , η = 2 × 0.01 \eta= \sqrt{2} \times0.01 η = 2 × 0.01 , γ = 1 / n \gamma= 1/n γ = 1/ n , S = 10 3 S = 10^{3} S = 1 0 3 –10 5 10^{5} 1 0 5 , 200 repetitions per fit). The matching rule was written down before the run, and the held-out calibration below was fixed in a dated addendum (see Data availability). Each fit is classified by its exact values as transport-defective (W 1 > τ W 1 W_1 > \tau_{\mathrm{W1}} W 1 > τ W1 ), jointly good (W 1 ≤ τ W 1 W_1 \le\tau_{\mathrm{W1}} W 1 ≤ τ W1 and M M D k 2 ≤ τ M M D \mathrm{MMD}^{2}_{k} \le\tau_{\mathrm{MMD}} MMD k 2 ≤ τ MMD ) or mismatch (W 1 ≤ τ W 1 W_1 \le\tau_{\mathrm{W1}} W 1 ≤ τ W1 and M M D k 2 > τ M M D \mathrm{MMD}^{2}_{k} > \tau_{\mathrm{MMD}} MMD k 2 > τ MMD ), and D 4 D_4 D 4 is compared with D 1 ( τ W 1 ) D_1(\tau_{\mathrm{W1}}) D 1 ( τ W1 ) , with D 1 ( τ W 1 − η ) D_1(\tau_{\mathrm{W1}}-\eta) D 1 ( τ W1 − η ) and with a recalibrated single-metric rule D 1 ( τ ∗ ) D_1(\tau^{*}) D 1 ( τ ∗ ) . The recalibration is cross-fitted by seed cluster. For each of the 20 training seeds of a scale, τ ∗ \tau^{*} τ ∗ is the smallest threshold whose false-fail rate on the jointly-good fits of the other 19 seeds does not exceed that of D 4 D_4 D 4 on the same fits; this is the tightest matching threshold and the choice least favorable to the escalation. D 1 ( τ ∗ ) D_1(\tau^{*}) D 1 ( τ ∗ ) is then evaluated on the held-out seed’s fits. Its realized false-fail rates differ from the calibration target by up to 0.009 (Table XI ), and τ ∗ \tau^{*} τ ∗ varies by less than 0.003 across the folds in every cell. A cost-matched variant repeats the calibration and the evaluation on draws of S ′ = ⌈ S ( 1 + e ˉ ) ⌉ S' = \lceil S(1+\bar{e})\rceil S ′ = ⌈ S ( 1 + e ˉ )⌉ shots, where e ˉ \bar{e} e ˉ is the escalation rate over all fits of the scale. That rate is zero where every rule rejects every fit and otherwise 0.09 to 0.36 (over the jointly-good fits alone, 0.01 to 0.57). Table XI reports the certification-grade regime.
TABLE XI. Fixed-threshold certification of the 600 perturbed full-support fits at certification grade (τ M M D = 10 − 4 \tau_{\mathrm{MMD}} = 10^{-4} τ MMD = 1 0 − 4 ), 200 repetitions per fit, all 18 cells. Counts: jointly good / transport-defective / mismatch fits. FF: false-fail rate on the good fits, printed as 1.000 when it rounds there (the smallest such value is 0.9996); τ ∗ \tau^* τ ∗ : mean of the cross-fitted threshold over the 20 folds; FF D 1 ( τ ∗ ) D_1(\tau^*) D 1 ( τ ∗ ) : realized false-fail rate of the cross-fitted rule on the held-out fits; S ′ / S S'/S S ′ / S : shot budget of the cost-matched variant, S ′ = ⌈ S ( 1 + e ˉ ) ⌉ S' = \lceil S(1+\bar{e})\rceil S ′ = ⌈ S ( 1 + e ˉ )⌉ , relative to S S S , and “cost”: the realized false-fail rate of the rule recalibrated and evaluated at S ′ S' S ′ ; FP: false-pass rate on the transport-defective fits, for D 1 ( τ ∗ ) D_1(\tau^*) D 1 ( τ ∗ ) cross-fitted at S S S and, under “cost”, at S ′ S' S ′ ; MR: rejection rate on the mismatch fits. An entry 0 is zero in every repetition (one-sided 95% upper limits 3 / ( 200 N d e f ) 3/(200N_{\mathrm{def}}) 3/ ( 200 N def ) for FP on this fixed panel, 0.0002–0.0009 for the 67 to 17 defective fits per scale), and < 0.001 <0.001 < 0.001 is a nonzero rate below 0.0005.
Table XI.
At the loose tolerance no fit is a mismatch pair at any n n n , as Theorem 4 requires of the population values, and D 4 D_4 D 4 returns the rates of D 1 ( τ W 1 ) D_1(\tau_{\mathrm{W1}}) D 1 ( τ W1 ) to the last digit in every cell, with no disagreement on any of the 41,771 in-window draws (§ 4 F). At certification grade the mismatch fits number 22, 11, 5 and 1 at n = 3 n = 3 n = 3 –6 and none beyond, and the escalation rejects at 43–98% at S = 10 5 S = 10^{5} S = 1 0 5 , against 0–61% for the recalibrated rule. On the transport-defective fits we compare the rules through the paired difference F P ( D 4 ) − F P ( D 1 ( τ ∗ ) ) \mathrm{FP}(D_4)-\mathrm{FP}(D_1(\tau^{*})) FP ( D 4 ) − FP ( D 1 ( τ ∗ )) , with a seed-cluster interval and a shot-repetition interval at fixed fits. At n = 3 n = 3 n = 3 , S = 10 3 S = 10^{3} S = 1 0 3 the difference is − 0.014 -0.014 − 0.014 , with cluster interval [ − 0.022 , − 0.008 ] [-0.022,-0.008] [ − 0.022 , − 0.008 ] over the 16 seeds that carry the 18 defective fits and repetition interval [ − 0.020 , − 0.009 ] [-0.020,-0.009] [ − 0.020 , − 0.009 ] . This is the only cell in which the cluster interval excludes zero in the escalation’s favor, and the advantage disappears when the single-metric rule is granted the escalation’s expected shot cost, with the threshold recalibrated on the larger draws (0.017 against 0.017; cost-matched difference 0.000, cluster interval [ − 0.004 , 0.004 ] [-0.004,0.004] [ − 0.004 , 0.004 ] ). At n = 3 n = 3 n = 3 , S = 10 4 S = 10^{4} S = 1 0 4 the cost-matched difference is − 0.003 -0.003 − 0.003 , with cluster interval [ − 0.007 , 0.000 ] [-0.007,0.000] [ − 0.007 , 0.000 ] and repetition interval [ − 0.005 , − 0.001 ] [-0.005,-0.001] [ − 0.005 , − 0.001 ] . At n = 4 n = 4 n = 4 and at n = 5 n = 5 n = 5 , S = 10 4 S = 10^{4} S = 1 0 4 it is 0.000 to 0.003 with cluster intervals that reach zero. At n = 5 n = 5 n = 5 , S = 10 5 S = 10^{5} S = 1 0 5 the recalibrated rule is the better one, 0.015 against 0, with cluster interval [ 0.000 , 0.046 ] [0.000,0.046] [ 0.000 , 0.046 ] and repetition interval $[0.013,0.019]; the false fails that D 4 D_4 D 4 incurs on in-window good fits, when spent on a lower single threshold instead, yield a cleaner false-pass rate than the discrepancy vote returns. From n = 6 n = 6 n = 6 every rule’s false-pass rate is zero or below 0.001. The cluster intervals describe variation across the fitted generators of this benchmark, and the repetition intervals the shot noise on a fixed panel of fits. Both are conditional on the fitted cross-validation decisions, since the calibration is not repeated inside the resamples and the folds overlap, and we pre-specified no equivalence margin. At S ≤ 10 4 S \le10^{4} S ≤ 1 0 4 the upward bias of w ^ \hat{w} w ^ rejects all but a few draws of the good fits under every rule (false-fail above 0.999 at n = 5 n = 5 n = 5 , S = 10 3 S = 10^{3} S = 1 0 3 , and at n ≥ 7 n \ge7 n ≥ 7 , S ≤ 10 4 S \le10^{4} S ≤ 1 0 4 ), the bias mechanism of the preceding subsection on targets whose W 1 W_1 W 1 sits near the threshold. On targets never used to tune it, the escalation enforces the discrepancy tolerance on in-window fits at small n n n . Its false-pass rate shows no advantage over a single transport threshold calibrated on held-out seeds to its false-fail rate and given its shot cost: the two agree to the third decimal in six of the nine cells with false passes, the recalibrated rule is better in the two cells at n = 5 n = 5 n = 5 , and the escalation is better in one cell, at n = 3 n = 3 n = 3 and S = 10 4 S = 10^{4} S = 1 0 4 , by 0.003.
Trained generators in the tolerance-mismatch region
On the pre-registered stress families at loose tolerances, trained-model W 1 W_1 W 1 and MMD AUROCs are both low (0.40–0.77 across cells) and neither approaches its convergence-conditioned performance, because the defects these generators produce do not fall into the mismatch regions here. Here we ask the same question at certification grade.
Theorem 2 and the Family 2 construction of § 4 E establish the mismatch regions as differences of acceptance regions. To confirm that trained generators can occupy them, we instantiate both constructions directly at n = 3 n = 3 n = 3 through 6. On the uniform reference p = u p = u p = u we build the defective target of each family, placing its parameter at the midpoint of the escalation window so that W 1 ( u , q ) ≤ τ W 1 W_1(u,q) \le\tau_{\mathrm{W1}} W 1 ( u , q ) ≤ τ W1 by construction. We then train the Mötönen QCBM to fit that target under the C 1 C_1 C 1 cost, twelve seeds per cell, and certify the trained output q ^ \hat{q} q ^ against the clean reference u u u , exactly as the protocol would in deployment.
Both families admit an exact window, so the construction fixes the thresholds. At the certification-grade operating point the mode-collapse window is open through n = 6 n = 6 n = 6 and the correlation-shift window only at n = 3 n = 3 n = 3 (Secs. VD and VE). Where the window is too narrow to train into reliably, we tighten τ M M D \tau_{\mathrm{MMD}} τ MMD to a quarter of the value at which the window closes, which pins η max \eta_{\max} η m a x to τ W 1 / 2 \tau_{\mathrm{W1}}/2 τ W1 /2 exactly, and keep the cell. Table XII marks those cells; their thresholds are 2.6 × 10 − 5 2.6 \times10^{-5} 2.6 × 1 0 − 5 for mode collapse at n = 6 n = 6 n = 6 , and 2.4, 1.5, and 1.1 × 10 − 5 \times10^{-5} × 1 0 − 5 for correlation shift at n = 4 , 5 , 6 n = 4,5,6 n = 4 , 5 , 6 .
TABLE XII. Trained generators in the two mismatch regions (12 seeds per cell), certified against the clean uniform reference at τ W 1 = 0.05 \tau_{\mathrm{W1}} = 0.05 τ W1 = 0.05 . Cells marked † use the tightened τ M M D \tau_{\mathrm{MMD}} τ MMD given in the text. Seeds absent from the D 1 D_1 D 1 PASS column overshot τ W 1 \tau_{\mathrm{W1}} τ W1 and are rejected by W1 alone, so they lie outside the mismatch region.
Table XII.
Table XII reports the outcome: across the eight cells 78 of the 96 trained generators land inside the W1-only acceptance region, and the escalation rejects 77 of those 78. The single escape is a mode-collapse run at n = 5 n = 5 n = 5 whose trained output drifted to M M D k 2 = 9.6 × 10 − 5 \mathrm{MMD}_k^2 = 9.6 \times10^{-5} MMD k 2 = 9.6 × 1 0 − 5 , just under the 10 − 4 10^{-4} 1 0 − 4 tolerance, in the cell whose window is narrowest at certification grade (η max / τ W 1 = 0.19 \eta_{\max}/\tau_{\mathrm{W1}} = 0.19 η m a x / τ W1 = 0.19 ). The eighteen generators absent from the D 1 D_1 D 1 PASS column overshot τ W 1 \tau_{\mathrm{W1}} τ W1 during training and are rejected by W1 alone, so they lie outside the mismatch region. They cluster at n = 5 n = 5 n = 5 for mode collapse and n = 6 n = 6 n = 6 for correlation shift, the cells where η max \eta_{\max} η m a x is smallest and the training target hardest to hit.
The targets were built at the window midpoint and the generators trained to fit them, so the experiment
shows that a QCBM can be driven into the mismatch region. To see whether ordinary training toward a legitimate target lands there, we searched the main campaign. Among the 1,052 runs of the 2,520-run campaign that W1-only certification admits against the clean reference, 56 carry M M D k 2 > 10 − 4 \mathrm{MMD}_k^2 > 10^{-4} MMD k 2 > 1 0 − 4 and are rejected by the escalation: 34, 4, 11, 6 and 1 at n = 3 n = 3 n = 3 through 7, and none at n = 8 n = 8 n = 8 . These are the clean-reference counterparts of the W1P/MMDP counts of Table X , which reports the trained-on decomposition at the same tolerance. The 56 split into two populations. Thirty-four are B1 runs at n = 3 n = 3 n = 3 and ϵ = 0.05 \epsilon= 0.05 ϵ = 0.05 , where the injected defect itself has clean-reference W 1 ≈ 0.043 < τ W 1 W_1 \approx0.043 < \tau_{W_1} W 1 ≈ 0.043 < τ W 1 and median M M D k 2 = 2.0 × 10 − 4 \mathrm{MMD}_k^2 = 2.0 \times10^{-4} MMD k 2 = 2.0 × 1 0 − 4 ; one more is a B1 run at n = 4 n = 4 n = 4 , ϵ = 0.05 \epsilon= 0.05 ϵ = 0.05 . The transport threshold admits them because the defect lies inside its declared tolerance, so these runs are the tolerance mismatch of Sec. V D on a trained generator and do not count toward the ordinary route. The remaining 21 are unperturbed B2 and B3 runs at n = 4 n = 4 n = 4 –7, generators fitted to a target carrying no injected defect at all, admitted by transport and rejected by discrepancy; they are 21 of the 776 admitted unperturbed fits, or 2.7%. Across the 56 the transport residuals sit at a median W 1 = 0.039 W_1 = 0.039 W 1 = 0.039 against a median M M D k 2 = 1.7 × 10 − 4 \mathrm{MMD}_k^2 = 1.7 \times10^{-4} MMD k 2 = 1.7 × 1 0 − 4 , and 36 of the 56 meet the convergence criterion of Sec. IV B. The population depends on the tolerance: at τ M M D = 10 − 3 \tau_{\mathrm{MMD}} = 10^{-3} τ MMD = 1 0 − 3 and 10 − 2 10^{-2} 1 0 − 2 it is empty at every scale, and at 10 − 5 10^{-5} 1 0 − 5 it grows to 448 of the 1,052. The ordinary route into the mismatch region is therefore real at certification grade, at a rate of about 3% of admitted unperturbed fits, and it is empty at the loose thresholds, where Theorem 4 makes the discrepancy vote redundant. Whether rejecting these fits improves downstream utility is a question only an external application criterion can settle.
Four of the eight cells required tightening τ M M D \tau_{\mathrm{MMD}} τ MMD below certification grade, as the closed-form window dictates. At τ W 1 = 0.05 \tau_{W_1} = 0.05 τ W 1 = 0.05 and τ M M D = 10 − 4 \tau_{\mathrm{MMD}} = 10^{-4} τ MMD = 1 0 − 4 the window is closed at n ≥ 7 n \ge7 n ≥ 7 for mode collapse and at n ≥ 4 n \ge4 n ≥ 4 for correlation shift. It reopens at any n n n once ρ \rho ρ falls below ρ ∗ ( n ) \rho^*(n) ρ ∗ ( n ) , that is, once τ M M D \tau_{\mathrm{MMD}} τ MMD drops below its closing value (7.8 × 10 − 5 7.8 \times10^{-5} 7.8 × 1 0 − 5 and 6.0 × 10 − 5 6.0 \times10^{-5} 6.0 × 1 0 − 5 for mode collapse at n = 7 , 8 n = 7, 8 n = 7 , 8 ; 9.7 × 10 − 5 9.7 \times10^{-5} 9.7 × 1 0 − 5 to 2.4 × 10 − 5 2.4 \times10^{-5} 2.4 × 1 0 − 5 for correlation shift at n = 4 n = 4 n = 4 through 8), at the shot cost of Appendix B 4.
Discussion
Why a bi-metric protocol, and when it is unnecessary
The Fisher-trace identity of Sec. III removes one channel, but it does not by itself select the bi-metric stack. With the classical Fisher trace pinned (Eq. (10) ), the W 1 W_1 W 1 /MMD escalation of Algorithm 1, with the rank as an anisotropy flag, is one choice among the distributional channels that remain. Total variation, the Sinkhorn divergence (which tracks W 1 W_1 W 1 closely across our campaign and interpolates between the two endpoints at n = 8 n = 8 n = 8 ; Sec. IV B), and spectral alternatives to the trace are all admissible in its place. The case for this particular pair rests on the finite-sample guarantee of Eq. (13) and the constructive separations of Sec. V E. Among spectral alternatives, the effective dimension of Abbas et al. [8 ] and the entropy-based effective rank of Roy and Vetterli [28 ] interpolate between the rank and the trace, and a parameter-shift estimate of the quantum Fisher matrix
on the full density operator would capture coherence contributions invisible to the diagonal classical estimator.
The escalation is useful only where the generator is approximately correct but may harbor a defect one metric cannot see, since far from the thresholds every metric agrees. Theorem 4 makes the redundancy exact on the Hamming cost: once τ M M D ≥ 2 ( 1 − e − γ ) τ W 1 2 \tau_{\mathrm{MMD}} \ge2(1-e^{-\gamma})\tau_{W_1}^2 τ MMD ≥ 2 ( 1 − e − γ ) τ W 1 2 the discrepancy vote on population values never differs from the transport vote on any pair of distributions, which covers the loose tolerance at every scale we study. With an independent discrepancy draw the agreement is a measured fact (Sec. V F), and Sec. V J shows the intermediate regime is non-empty only at tight tolerances.
Diagnostics that recover the experimental design
The headline score of a diagnostic on a constructed benchmark does not show whether it measured the distributional property or recovered the control variable (Sec. I). Sections III–IV B give three instances of this failure: the trace is pinned by construction, the rank counts the support of the converged distribution, and the pooled scores move with the scoring convention. Five inexpensive controls expose them: a severity-stratified restriction to the perturbed runs, a convergence restriction of the negative class, a stress test across independently constructed defect families, a replication under further optimizers and objectives, and a labeled benchmark whose clean target has full support. We recommend them as default guards whenever a diagnostic is scored by AUROC against a constructed benchmark.
Among the runs the bi-metric protocol certifies, the rank-deficient subset isolates verdicts with anisotropic robustness (Sec. VI): these verdicts withstand random drift better than full-rank ones while remaining fragile along a few sensitive directions. We therefore use the rank as a flag for that anisotropy, triggering manual inspection where structured drift is plausible; the bi-metric decision stands. A full-rank certified run at n ≥ 6 n \ge6 n ≥ 6 is one whose fit did not reach exact sparsity (Sec. IV A), so the flag separates converged-sparse from residual-background fits. Replacing the Fisher trace in Algorithm 1 by the eigenvalue rank would change the certification question from “does the output distribution match the target?” to “does it match, and is the converged parameter regime robust?”, which is a different certification problem.
Ground-cost sensitivity and quantum optimal transport
The paired seed-cluster difference between the ground costs C 1 C_1 C 1 (Hamming) and C 2 C_2 C 2 (weighted Hamming) excludes zero at n = 5 n = 5 n = 5 alone, where C 2 C_2 C 2 scores 0.24 higher on the clean-reference W 1 W_1 W 1 AUROC, and covers zero at every other scale (Sec. IV B). No fixed cost is therefore best at every scale. The ground-cost mismatch noted in § 4 E needs only that the acceptance polytopes differ.
The sensitivity can be framed inside the order-p p p quantum-Wasserstein construction of De Palma, Marvian, Trevisan, and Lloyd [37 ] and the reformulation of Beatty and Stilck França [38 ] . In the classical diagonal limit, the order-1 quantum-Wasserstein distance of these constructions reduces to the classical W 1 W_1 W 1 on { 0 , 1 } n \{0,1\}^n { 0 , 1 } n with a ground cost determined by an operator-norm choice on the underlying observable algebra. Within that framework, C 1 C_1 C 1 is the canonical choice corresponding to the single-qubit Pauli norm; C 2 C_2 C 2 and C 3 C_3 C 3 (Hamming with a parity-correlation correction) are heuristic refinements motivated by the order-p p p machinery without being restrictions of it. The quantum transport metric on diagonal states is therefore not unique, which motivates deriving an operationally optimal cost from the underlying construction.
The identity under noise
The proof of Proposition 1 uses reality of the state vector to remove the phase term and the involutive generator to fix the tangent norm (Appendix A). Neither survives a mixed state, so Tr F Q \operatorname{Tr}F_Q Tr F Q is no longer pinned to d d d and the trace regains information about the fit. The pinning fails already at small noise rates. In the noiseless case each outcome contributes 4 ∑ k ( ∂ k ψ θ ( x ) ) 2 4\sum_k(\partial_k\psi_\theta(x))^2 4 ∑ k ( ∂ k ψ θ ( x ) ) 2 regardless of how small p θ ( x ) p_\theta(x) p θ ( x ) is, which is why the identity holds; under a depolarizing channel of rate p d e p p_{\mathrm{dep}} p dep the same contribution is divided by ( 1 − p d e p ) p θ ( x ) + p d e p 2 − n (1-p_{\mathrm{dep}})p_\theta(x)+p_{\mathrm{dep}}2^{-n} ( 1 − p dep ) p θ ( x ) + p dep 2 − n and is suppressed once the outcome falls below the noise floor p d e p 2 − n p_{\mathrm{dep}}2^{-n} p dep 2 − n , and sparse fits put nearly all of their outcomes below that floor at once. Explicitly, the channel maps J → ( 1 − p d e p ) J J\to(1-p_{\mathrm{dep}})J J → ( 1 − p dep ) J and p θ → ( 1 − p d e p ) p θ + p d e p 2 − n p_\theta\to(1-p_{\mathrm{dep}})p_\theta+p_{\mathrm{dep}}2^{-n} p θ → ( 1 − p dep ) p θ + p dep 2 − n , so that
Tr F c l = ∑ x ( 1 − p d e p ) 2 ∥ J x ∥ 2 ( 1 − p d e p ) p θ ( x ) + p d e p 2 − n , (30) \operatorname{Tr}F_{\mathrm{cl}}=\sum_x\frac{(1-p_{\mathrm{dep}})^2\lVert J_x\rVert^2}{(1-p_{\mathrm{dep}})p_\theta(x)+p_{\mathrm{dep}}2^{-n}}, \tag*{(30)} Tr F cl = x ∑ ( 1 − p dep ) p θ ( x ) + p dep 2 − n ( 1 − p dep ) 2 ∥ J x ∥ 2 , (30)
and post-processing the stored Jacobians of the campaign with Eq. (30) gives the values of Table XIII .
A depolarizing rate of one part in a thousand thus removes the identity on the sparse targets of the deployment regime and leaves diffuse targets untouched. The trace regains a dependence on the entropy of the target, which is unrelated to fit quality, so Table XIII shows the fragility of the identity and no route to certification under noise. This sweep applies a distribution-level noise model to stored Jacobians; without circuit-level simulation it cannot show how gate and readout error propagate into the two protocol metrics. A second preliminary finding compounds the first: the naive certification AUROC inverts under depolarizing and readout error, because noise pushes outputs toward uniform and so penalizes concentrated targets more than diffuse ones. A noise-aware protocol therefore cannot certify against the noiseless target, and the false-pass guarantee of Eq. (13) has not been established under device noise. Settling the correct reference and extending the bound, through noise-aware coupling arguments for W 1 W_1 W 1 and an adaptation for the MMD U U U -statistic, requires circuit-level simulation.
Limitations
The campaign covers n = 3 n=3 n = 3 –8 8 8 , bounded by the exact-statevector requirement and the trainability frontier; verification at n ≥ 9 n\geq9 n ≥ 9 with converged training would strengthen the extrapolation. The near-coincidence of the escalation-window closure (n ≥ 7 n\geq7 n ≥ 7 for mode collapse at the certification-grade operating point) with the trainability frontier (n = 9 n=9 n = 9 ) is accidental: the closure depends on the threshold ratio ρ \rho ρ alone, and Figure 4 shows an open window at any n n n for sufficiently small ρ \rho ρ , at the shot cost of Appendix B4, while the frontier depends on the optimizer budget. Between the two sits an untested regime, n = 7 , 8 n=7,8 n = 7 , 8 at operating points below ρ ∗ ( n ) \rho^*(n) ρ ∗ ( n ) , where models train and the window is open but the tolerances cost more shots than any budget we exercise. All experiments use exact noiseless statevector simulation, which is the sharpest limitation of the paper (Sec. VI D): the protocol carries no false-pass guarantee under device noise.
The Fisher-trace identity of Proposition 1 applies to ansätze satisfying conditions (cond1 _1 1 ) and (cond2 _2 2 ) of § 2 A. Our empirical specialization covers only the Möttönen real-amplitude state-preparation ansatz of § 2 D, and PQC families that violate either condition require a separate analysis. The campaign trains the QCBM by CMA-ES, with the rank control of Sec. IV A replicated at n ≤ 5 n\leq5 n ≤ 5 under Powell and Adam. Other generative architectures, such as QGANs with adversarial training or QGDMs with diffusion schedules, may occupy the mismatch regions differently: the theoretical results (Proposition 1, the false-pass bound (13) and Theorem 2) are model-agnostic on the appropriate ansatz class, but the disagreement rates we report are model-specific.
The escalation’s advantage is confined to certification-grade tolerances and to the scales at which η max ( n ) \eta_{\max}(n) η m a x ( n ) is positive; both limits are exact, since the window is a closed form. At the loose benchmark-comparison thresholds the trained-model W 1 W_1 W 1 and MMD AUROCs are both low on the pre-specified stress families (§ 4 § 4.1.11 ). At certification grade, four of the eight cells in § 4 § 4.1.11 required a τ M M D \tau_{\mathrm{MMD}} τ MMD tightened below 10 − 4 10^{-4} 1 0 − 4 to keep the window open at all, the ground-cost mismatch has no trained-generator demonstration, and the tolerances at which the demonstrated families separate are expensive to resolve on hardware (Appendix B § 5.8.4 ). We demonstrate that two distinct defect mechanisms, realizable by a trained QCBM, sit inside the W 1 W_1 W 1 -only acceptance region and are caught by the escalation; deployment scale and deployable shot budgets remain untested. On the fixed-threshold test of § 4 K, a transport threshold calibrated on held-out seeds to the escalation’s false-fail rate, and given its shot cost, matches its false-pass rate in six of the nine cells with false passes and betters it in two, so that test records no advantage of the escalation over a single recalibrated threshold on independent targets. Theorems 3 ( c ) 3(\mathrm{c}) 3 ( c ) and 4 4 4 are exact for the Hamming cost and the RBF–Hamming kernel; the minimax bound holds for every normalized kernel, but the redundancy constant is cost-specific.
The principal AUROC analysis rests on B1, whose perturbation label coincides with the support of the target. The perturbed full-support benchmark of Sec. IV C covers n = 3 n = 3 n = 3 –8 8 8 , but its clean target is unconverged within the publication budget at n ≥ 7 n \ge7 n ≥ 7 , where the conditioned comparison is empty. It shows that the perfect conditioned scores are a property of the sparse benchmark, so the convergence-conditioned claims are scoped to B1 and to defects that exceed the fit residual. Benchmark B3 is a synthetic near-uniform, high-entropy distribution, which limits the external validity of B3-specific conclusions. Finally, the zero-amplitude locus Z \mathcal{Z} Z is explicit for the Möttönen ansatz (Appendix A 2), but its relation to the set of parameters the optimizer actually reaches is open.
Conclusion
To conclude, we have asked whether the diagnostics used to certify a quantum generative model measure the quality of its output distribution or recover how its benchmark was built and how it was trained. Our starting point was structural: for any circuit that prepares real amplitudes through half-angle rotation generators, classical and quantum Fisher information coincide away from the zero-amplitude set, so that the trace is pinned to the number of parameters (Proposition 1 and Eq. (10) ). The trace carries no signal about how well a model fits. The sub-cap values that a finite-precision estimator returns originate in the clamp on the inverse-probability weight.
The two remaining candidates recover the experimental design in different ways, which our controls separate. The Fisher eigenvalue rank separates the benchmark classes because it reads the support of the converged distribution at the resolution of its eigenvalue threshold: on the two-peak target an exact fit leaves
2 n − 1 2n - 1 2 n − 1 live nodes of the amplitude tree, the rank of the quantum Fisher matrix there. It therefore falls to chance on a severity split and on full-support defect families, and the same count appears under Powell and under Adam on the same objective. The pooled distributional scores respond to optimizer convergence, to the coverage of a re-optimization policy and to the reference against which they are taken; conditioning the negative class on convergence makes them perfect on the sparse benchmark, and on a full-support benchmark, where the rank scores at chance, they resolve a defect to the extent that it exceeds the fit residual. Certification rests on the two distributional channels. For these the escalation rule admits a finite-sample false-pass bound with a quadratic margin-zone improvement on the single-metric rule at the same threshold under disjoint-sample estimation. The improvement survives at equal total sample budget for a larger discrepancy margin, and it is paid for in false fails inside the window relative to the single-metric rule at the shifted threshold.
For a practitioner, the closed-form window η max ( n ) \eta_{\max}(n) η m a x ( n ) of two exactly evaluable defect families states at which scales and tolerance ratios escalation is worth its extra shots. Inside that window the escalation acts on both families as the single shifted transport threshold, and the discrepancy tolerance is equivalent to transport tolerances that differ between the families by a fixed factor of about 1.6 1.6 1.6 . On the Hamming cost, the discrepancy is redundant for every pair of distributions above the tolerance ratio 2 ( 1 − e − τ ) τ 2 W 1 2(1-e^{-\tau})\tau^2 W_1 2 ( 1 − e − τ ) τ 2 W 1 , which covers the loose tolerance at every scale we study. For any fixed normalized kernel, some parity defect of transport size β / 2 \beta/2 β /2 has squared discrepancy at most β 2 / ( 2 n − 1 ) \beta^2/(2^n-1) β 2 / ( 2 n − 1 ) , so that certification must name the correlation orders it is meant to detect. At the loose tolerance the discrepancy vote is redundant on population values and changed no verdict in any in-window draw of the fixed-threshold panel at a thousand shots or more; the only exceptions in our measurements are five hundred-shot configurations in which the shot noise of the discrepancy estimate exceeds the margin. At certification grade it rejects about 3 % 3\% 3% of the admitted unperturbed campaign fits, the remaining rejections being defects placed inside the transport tolerance by construction. The window closes with scale and persists against an unknown defect family only at n = 3 n = 3 n = 3 , and the tolerances it requires are expensive to resolve on hardware (Appendix B § 5.8.4 ). Three extensions follow from these results, namely an operationally optimal ground cost derived from the order-1 quantum-Wasserstein construction [37 , 38 ] , certification under device noise, where the pure-state identity no longer holds (Sec. VI D), and a variance-adaptive concentration bound for the discrepancy channel tight enough to set shot budgets. We expect the last of these to be the most immediate step, since the measured variance of the discrepancy estimator already lies orders of magnitude below its range, so that a Bernstein-type certificate should become non-vacuous at the shot budgets a hardware experiment would plausibly use.
Data availability
The data that support the findings of this article are deposited at Zenodo on submission [39 ] ; until the record is public they are available from the authors on request. The code is released under the MIT license and the data under the Creative Commons Attribution 4.0 license. The archive contains the raw per-run results and merged tables of the 2,520-run campaign at n = 3 n = 3 n = 3 –8 8 8 , as run and warm-restarted, and of the excluded 420-run batch at n = 9 n = 9 n = 9 ; the data of the finite-shot, shot-budget, mismatch-region, depolarizing, optimizer-replication and full-support-benchmark studies; the dated pre-registration documents and their addenda; the Möttönen QCBM implementation; the analysis scripts that reproduce every table and figure; and the \LaTeX source of this article.
Acknowledgments
This material is based upon work supported by the U.S. Department of Energy, Office of Science, Accelerated Research in Quantum Computing, Fundamental Algorithmic Research toward Quantum Utility (FAR-Qu).
Proof of the structural Fisher-trace identity
This appendix gives the full proof of Proposition 1 and the explicit zero-amplitude locus of the Möttönen ansatz.
Proof of Proposition 1. The argument proceeds in five steps via the quantum–classical Cramér–Rao matrix inequality combined with the half-angle structure of the generators.
Step 1 (Cramér–Rao matrix bound). For any pure parameterized state ∣ ψ ( θ ) ⟩ |\psi(\theta)\rangle ∣ ψ ( θ )⟩ and any fixed positive operator-valued measure (POVM), in particular the projective measurement in the computational basis, the classical Fisher information matrix of the resulting outcome distribution is dominated by the quantum Fisher information matrix in the Löwner (PSD) partial order [26 , 27 , 40 ] :
F c l ( θ ) ⪯ F Q ( θ ) . (31) F_{\mathrm{cl}}(\theta) \preceq F_{\mathrm{Q}}(\theta). \tag*{(31)} F cl ( θ ) ⪯ F Q ( θ ) . (31)
Taking traces preserves the inequality: Tr F c l ( θ ) ≤ Tr F Q ( θ ) \operatorname{Tr}F_{\mathrm{cl}}(\theta) \le\operatorname{Tr}F_{\mathrm{Q}}(\theta) Tr F cl ( θ ) ≤ Tr F Q ( θ ) .
Step 2 (Quantum FIM diagonal for pure states). For a pure state ∣ ψ ( θ ) ⟩ |\psi(\theta)\rangle ∣ ψ ( θ )⟩ depending smoothly on θ \theta θ , the diagonal entries of F Q F_{\mathrm{Q}} F Q are [12 ]
F Q k k ( θ ) = 4 ( ∥ ∂ k ψ ∥ 2 − ∣ ⟨ ψ ∣ ∂ k ψ ⟩ ∣ 2 ) . (32) F_{\mathrm{Q}}^{kk}(\theta) = 4\left(\lVert\partial_k\psi\rVert^2 - \lvert\langle\psi\mid\partial_k\psi\rangle\rvert^2\right). \tag*{(32)} F Q k k ( θ ) = 4 ( ∥ ∂ k ψ ∥ 2 − ∣⟨ ψ ∣ ∂ k ψ ⟩ ∣ 2 ) . (32)
By condition (cond1 _1 1 ), ∣ ψ ( θ ) ⟩ |\psi(\theta)\rangle ∣ ψ ( θ )⟩ is a real vector for every θ \theta θ , so its derivative ∂ k ∣ ψ ⟩ \partial_k|\psi\rangle ∂ k ∣ ψ ⟩ is real as well, and the overlap ⟨ ψ ∣ ∂ k ψ ⟩ \langle\psi|\partial_k\psi\rangle ⟨ ψ ∣ ∂ k ψ ⟩ is therefore real. Differentiating the normalization ⟨ ψ ∣ ψ ⟩ = 1 \langle\psi|\psi\rangle= 1 ⟨ ψ ∣ ψ ⟩ = 1 gives ⟨ ψ ∣ ∂ k ψ ⟩ + ⟨ ∂ k ψ ∣ ψ ⟩ = 2 ⟨ ψ ∣ ∂ k ψ ⟩ = 0 \langle\psi|\partial_k\psi\rangle+ \langle\partial_k\psi|\psi\rangle= 2\langle\psi|\partial_k\psi\rangle= 0 ⟨ ψ ∣ ∂ k ψ ⟩ + ⟨ ∂ k ψ ∣ ψ ⟩ = 2 ⟨ ψ ∣ ∂ k ψ ⟩ = 0 , whence
⟨ ψ ∣ ∂ k ψ ⟩ = 0 for all θ and all k . (33) \langle\psi|\partial_k\psi\rangle= 0 \qquad\text{for all }\theta\text{ and all }k. \tag*{(33)} ⟨ ψ ∣ ∂ k ψ ⟩ = 0 for all θ and all k . (33)
(For a general pure state this overlap is only purely imaginary; the reality condition (cond1 _1 1 ) forces it to vanish outright.) This is the computational-basis specialization of the classical–quantum Fisher equality condition for real-amplitude pure states established by Facchi et al. [11 ] . Substituting Eq. (A3) into Eq. (A2) gives
F Q k k ( θ ) = 4 ∥ ∂ k ψ ∥ 2 . (34) F_{\mathrm{Q}}^{kk}(\theta)=4\lVert\partial_k\psi\rVert^2. \tag*{(34)} F Q k k ( θ ) = 4 ∥ ∂ k ψ ∥ 2 . (34)
Step 3 (Tangent norm from a half-angle generator). Condition (cond2 _2 2 ) gives ∂ k ∣ ψ ⟩ = − i 2 G ~ ( k ) ∣ ψ ⟩ \partial_k|\psi\rangle=-\frac{i}{2}\widetilde{G}_{(k)}|\psi\rangle ∂ k ∣ ψ ⟩ = − 2 i G ( k ) ∣ ψ ⟩ with G ~ ( k ) \widetilde{G}_{(k)} G ( k ) unitary. Therefore
∥ ∂ k ψ ∥ 2 = 1 4 ∥ G ~ ( k ) ψ ∥ 2 = 1 4 ∥ ψ ∥ 2 = 1 4 . (35) \lVert\partial_k\psi\rVert^2=\frac{1}{4}\lVert\widetilde{G}_{(k)}\psi\rVert^2=\frac{1}{4}\lVert\psi\rVert^2=\frac{1}{4}. \tag*{(35)} ∥ ∂ k ψ ∥ 2 = 4 1 ∥ G ( k ) ψ ∥ 2 = 4 1 ∥ ψ ∥ 2 = 4 1 . (35)
Step 4 (Combining). Equations (A4) and (A5) give F Q k k ( θ ) = 1 F_{\mathrm{Q}}^{kk}(\theta)=1 F Q k k ( θ ) = 1 for every k ∈ { 1 , … , d } k\in\{1,\ldots,d\} k ∈ { 1 , … , d } and every θ \theta θ , so Tr F Q ( θ ) = d \operatorname{Tr}F_{\mathrm{Q}}(\theta)=d Tr F Q ( θ ) = d exactly. Combining with Step 1:
Tr F c l ( θ ) ≤ Tr F Q ( θ ) = d . (36) \operatorname{Tr}F_{\mathrm{cl}}(\theta)\leq\operatorname{Tr}F_{\mathrm{Q}}(\theta)=d. \tag*{(36)} Tr F cl ( θ ) ≤ Tr F Q ( θ ) = d . (36)
This is part (a).
Step 5 (Equality off the zero-amplitude set). For a real state, Eq. (A3) reduces the full quantum Fisher matrix to F Q j k = 4 ⟨ ∂ j ψ ∣ ∂ k ψ ⟩ = 4 ∑ x ∂ j ψ θ ( x ) ∂ k ψ θ ( x ) F_{\mathrm{Q}}^{jk}=4\langle\partial_j\psi|\partial_k\psi\rangle=4\sum_x\partial_j\psi_\theta(x)\partial_k\psi_\theta(x) F Q j k = 4 ⟨ ∂ j ψ ∣ ∂ k ψ ⟩ = 4 ∑ x ∂ j ψ θ ( x ) ∂ k ψ θ ( x ) , the sum running over all outcomes. On an outcome with p θ ( x ) = ψ θ ( x ) 2 > 0 p_\theta(x)=\psi_\theta(x)^2>0 p θ ( x ) = ψ θ ( x ) 2 > 0 the classical summand is ∂ j p θ ∂ k p θ / p = 4 ∂ j ψ θ ( x ) ∂ k ψ θ ( x ) \partial_jp_\theta\partial_kp_\theta/p=4\partial_j\psi_\theta(x)\partial_k\psi_\theta(x) ∂ j p θ ∂ k p θ / p = 4 ∂ j ψ θ ( x ) ∂ k ψ θ ( x ) , the same as the quantum one, so
F c l ( θ ) = F Q ( θ ) − 4 ∑ x : ψ θ ( x ) = 0 ( ∇ ψ θ ( x ) ) ( ∇ ψ θ ( x ) ) T . (37) F_{\mathrm{cl}}(\theta)=F_{\mathrm{Q}}(\theta)-4\sum_{x:\,\psi_\theta(x)=0}\left(\nabla\psi_\theta(x)\right)\left(\nabla\psi_\theta(x)\right)^{\mathsf{T}}. \tag*{(37)} F cl ( θ ) = F Q ( θ ) − 4 x : ψ θ ( x ) = 0 ∑ ( ∇ ψ θ ( x ) ) ( ∇ ψ θ ( x ) ) T . (37)
If θ ∉ Z \theta\notin\mathcal{Z} θ ∈ / Z , every outcome with ψ θ ( x ) = 0 \psi_\theta(x)=0 ψ θ ( x ) = 0 also has ∇ ψ θ ( x ) = 0 \nabla\psi_\theta(x)=0 ∇ ψ θ ( x ) = 0 , the correction vanishes, and F c l = F Q F_{\mathrm{cl}}=F_{\mathrm{Q}} F cl = F Q with trace d d d by Step 4. If θ ∈ Z \theta\in\mathcal{Z} θ ∈ Z , the correction has trace 4 ∑ x : ψ θ ( x ) = 0 ∥ ∇ ψ θ ( x ) ∥ 2 > 0 4\sum_{x:\,\psi_\theta(x)=0}\lVert\nabla\psi_\theta(x)\rVert^2>0 4 ∑ x : ψ θ ( x ) = 0 ∥ ∇ ψ θ ( x ) ∥ 2 > 0 and the inequality is strict. This is part (b). Steps 1–5 use only conditions (cond1 _1 1 ) and (cond2 _2 2 ); when (cond3 _3 3 ) holds as well, d = 2 n − 1 d=2^n-1 d = 2 n − 1 .
The cost of observing the identity at finite shots
The identity and its saturation are statements about the exact output distribution p θ p_\theta p θ . A hardware experimenter, however, estimates p θ p_\theta p θ from a finite number of measurement shots S S S , and the plug-in classical Fisher trace inherits the resulting estimation error through the inverse-probability weighting in F c l = J T diag ( 1 / p ) J F_{\mathrm{cl}}=J^{\mathsf{T}}\operatorname{diag}(1/p)J F cl = J T diag ( 1/ p ) J .
The saturation is supported in part on rare outcomes, whose probability decreases with n n n : the median minimum basis probability on saturated parameters falls from 4.8 × 10 − 4 4.8 \times10^{-4} 4.8 × 1 0 − 4 at n = 3 n=3 n = 3 to 2.8 × 10 − 8 2.8 \times10^{-8} 2.8 × 1 0 − 8 at n = 6 n=6 n = 6 , and further to 1.6 × 10 − 11 1.6 \times10^{-11} 1.6 × 1 0 − 11 at n = 7 n=7 n = 7 and 2.4 × 10 − 13 2.4 \times10^{-13} 2.4 × 1 0 − 13 at n = 8 n=8 n = 8 . The naive plug-in estimator 1 / p ^ 1/\hat{p} 1/ p ^ is dominated by zero-count outcomes and overshoots the cap by several orders of magnitude even at S = 10 6 S=10^{6} S = 1 0 6 . A smoothed estimator removes this pathology: with the Krichevsky–Trofimov estimate p ^ x = ( c x + 1 2 ) / ( S + 2 n − 1 ) \hat{p}_x=(c_x+\frac{1}{2})/(S+2^{n-1}) p ^ x = ( c x + 2 1 ) / ( S + 2 n − 1 ) , where c x c_x c x is the shot count of outcome x x x , the plug-in trace concentrates onto the cap, as shown in Figure 7 .
The run-to-run spread sets the budget: the interquartile width reaches ± 1 % \pm1\% ± 1% only at S ≈ 3 × 10 4 S \approx3 \times10^{4} S ≈ 3 × 1 0 4 for n = 3 n=3 n = 3 and S ≈ 10 5 S \approx10^{5} S ≈ 1 0 5 –3 × 10 5 3 \times10^{5} 3 × 1 0 5 beyond, and at S = 10 2 S=10^{2} S = 1 0 2 the median overshoots the cap by 78% at n = 8 n=8 n = 8 . A formal finite-shot bound accounting for the rare-outcome tail is left to future work.
The zero-amplitude locus of the Möttönen ansatz
For the Möttönen ansatz the set Z \mathcal{Z} Z of Proposition 1 (b) is explicit. Each stage of the circuit is a uniformly controlled R Y R_Y R Y rotation on qubit k k k conditioned on qubits 0 , … , k − 1 0,\ldots,k-1 0 , … , k − 1 [13 , 14 ] , so every basis amplitude factorizes along a binary tree into n n n half-angle trigonometric factors [41 Lemma 3.2] ,
ψ θ ( x ) = ∏ k = 0 n − 1 κ x k ( 1 2 ϑ x < k ) , κ 0 = cos , κ 1 = sin , (38) \psi_{\theta}(x)=\prod_{k=0}^{n-1}\kappa_{x_k}\left(\frac{1}{2}\vartheta_{x_{<k}}\right),\qquad\kappa_0=\cos,\ \kappa_1=\sin, \tag*{(38)} ψ θ ( x ) = k = 0 ∏ n − 1 κ x k ( 2 1 ϑ x < k ) , κ 0 = cos , κ 1 = sin , (38)
where x < k = ( x 0 , … , x k − 1 ) x_{<k}=(x_0,\ldots,x_{k-1}) x < k = ( x 0 , … , x k − 1 ) labels the node of the tree at depth k k k and the 2 n − 1 2^n-1 2 n − 1 node angles are ϑ = Λ θ \vartheta=\Lambda\theta ϑ = Λ θ , with Λ \Lambda Λ an invertible matrix of entries ± 1 \pm1 ± 1 fixed by the Gray-code compilation of each stage. Our compilation omits the closing CNOT of each Gray-code ladder, which only relabels the computational basis and leaves Eq. (38) intact up to that relabeling; stage k k k then uses 2 k − 1 2^{k-1} 2 k − 1 rotations and 2 k − 1 − 1 2^{k-1}-1 2 k − 1 − 1 CNOTs, so the circuit has ∑ k = 1 n 2 k − 1 = 2 n − 1 \sum_{k=1}^{n}2^{k-1}=2^n-1 ∑ k = 1 n 2 k − 1 = 2 n − 1 rotations and ∑ k = 2 n ( 2 k − 1 − 1 ) = 2 n − n − 1 \sum_{k=2}^{n}(2^{k-1}-1)=2^n-n-1 ∑ k = 2 n ( 2 k − 1 − 1 ) = 2 n − n − 1 CNOTs, the counts of Sec. II D. We verified the factorization and the invertibility of Λ \Lambda Λ numerically for n = 2 n=2 n = 2 to 7.
If two or more factors of ψ θ ( x ) \psi_{\theta}(x) ψ θ ( x ) vanish, every first derivative of the product vanishes, whereas if exactly one vanishes its ϑ \vartheta ϑ -derivative is nonzero and, Λ \Lambda Λ being invertible, so is some ∂ k ψ θ ( x ) \partial_k\psi_{\theta}(x) ∂ k ψ θ ( x ) . Hence Z \mathcal{Z} Z is the set of θ \theta θ at which, for some outcome x x x , exactly one factor in Eq. (38) vanishes. Its closure is the union over the 2 n − 1 2^n-1 2 n − 1 nodes of the hyperplane families { ϑ j ∈ π Z } \{\vartheta_j\in\pi\mathbb{Z}\} { ϑ j ∈ π Z } , flat in the node angles and oblique in the circuit angles, and Z \mathcal{Z} Z omits from this union only the codimension-two intersections on which a factor further up the same branch also vanishes. The same structure places exact sparse fits on Z \mathcal{Z} Z . An exact representation of a target with vanishing probabilities prunes whole subtrees of Eq. (38) , and since the angles inside a pruned subtree are unconstrained, each pruned subtree is generically cut by a single vanishing factor.
A per-run certificate for the clamped rank
The clamp of Sec. II H changes the Fisher matrix wherever an outcome lies below the floor, and Sec. IV A reports how often the numerical rank survives the change. The following statement turns the trace deficit of Eq. (11) into a sufficient condition, checkable run by run, for the clamped and the quantum Fisher ranks to agree.
Proposition 5 (Rank certificate for the clamped estimator). Let A = F ^ c l ( θ ) A=\hat{F}_{\mathrm{cl}}(\theta) A = F ^ cl ( θ ) be the clamped estimator with floor p f l o o r p_{\mathrm{floor}} p floor and B = F Q ( θ ) B=F_Q(\theta) B = F Q ( θ ) the quantum Fisher matrix at the same real-amplitude parameters, and write g x = ∇ θ ψ θ ( x ) g_x=\nabla_\theta\psi_\theta(x) g x = ∇ θ ψ θ ( x ) . Then
B − A = 4 ∑ x : p θ ( x ) < p f l o o r ( 1 − p θ ( x ) p f l o o r ) g x g x T ⪰ 0 , (39) B-A=4\sum_{x:p_\theta(x)<p_{\mathrm{floor}}}\left(1-\frac{p_\theta(x)}{p_{\mathrm{floor}}}\right)g_xg_x^{\mathsf{T}}\succeq0, \tag*{(39)} B − A = 4 x : p θ ( x ) < p floor ∑ ( 1 − p floor p θ ( x ) ) g x g x T ⪰ 0 , (39)
so that with ε = Tr ( B − A ) \varepsilon=\operatorname{Tr}(B-A) ε = Tr ( B − A ) one has ∥ B − A ∥ o p ≤ ε \lVert B-A\rVert_{\mathrm{op}}\leq\varepsilon ∥ B − A ∥ op ≤ ε and 0 ≤ λ i ( B ) − λ i ( A ) ≤ ε 0\leq\lambda_i(B)-\lambda_i(A)\leq\varepsilon 0 ≤ λ i ( B ) − λ i ( A ) ≤ ε for the eigenvalues in decreasing order. For a relative cutoff ρ ≥ 0 \rho\geq0 ρ ≥ 0 let m i = λ i ( A ) − ρ λ max ( A ) m_i=\lambda_i(A)-\rho\lambda_{\max}(A) m i = λ i ( A ) − ρ λ m a x ( A ) . Then
− ρ ε ≤ [ λ i ( B ) − ρ λ max ( B ) ] − m i ≤ ε , (40) -\rho\varepsilon\leq\left[\lambda_i(B)-\rho\lambda_{\max}(B)\right]-m_i\leq\varepsilon, \tag*{(40)} − ρε ≤ [ λ i ( B ) − ρ λ m a x ( B ) ] − m i ≤ ε , (40)
every eigenvalue with m i > ρ ϵ m_i > \rho\epsilon m i > ρ ϵ lies above the cutoff in both matrices, every eigenvalue with m i < − ϵ m_i < -\epsilon m i < − ϵ lies below it in both, and if each eigenvalue satisfies one of the two alternatives then rank ρ ( A ) = rank ρ ( B ) \operatorname{rank}_{\rho}(A) = \operatorname{rank}_{\rho}(B) rank ρ ( A ) = rank ρ ( B ) .
Proof. Wherever ψ θ ( x ) ≠ 0 \psi_{\theta}(x) \ne0 ψ θ ( x ) = 0 the classical summand is ( ∂ k p ) ( ∂ l p ) / p = 4 ∂ k ψ ∂ l ψ (\partial_k p)(\partial_l p)/p = 4\partial_k\psi\,\partial_l\psi ( ∂ k p ) ( ∂ l p ) / p = 4 ∂ k ψ ∂ l ψ , which is also its continuous extension at ψ θ ( x ) = 0 \psi_{\theta}(x) = 0 ψ θ ( x ) = 0 (Proposition 1), so B = 4 ∑ x g x g x T B = 4\sum_x g_xg_x^{\mathsf{T}} B = 4 ∑ x g x g x T , whereas the clamp multiplies the summand of a sub-floor outcome by p θ ( x ) / p f l o o r p_{\theta}(x)/p_{\mathrm{floor}} p θ ( x ) / p floor ; the difference is Eq. (39) , a sum of positive semidefinite rank-one terms, whose trace ϵ \epsilon ϵ bounds its operator norm. Weyl’s inequality for a positive semidefinite perturbation gives 0 ≤ λ i ( B ) − λ i ( A ) ≤ ϵ 0 \le\lambda_i(B) - \lambda_i(A) \le\epsilon 0 ≤ λ i ( B ) − λ i ( A ) ≤ ϵ for every i i i , in particular for the largest eigenvalue, and subtracting ρ \rho ρ times the bound on λ max \lambda_{\max} λ m a x from the bound on λ i \lambda_i λ i gives Eq. (40) . An eigenvalue counts toward rank ρ \operatorname{rank}_{\rho} rank ρ when it exceeds ρ λ max \rho\lambda_{\max} ρ λ m a x , so m i > ρ ϵ ≥ 0 m_i > \rho\epsilon\ge0 m i > ρ ϵ ≥ 0 places it above the cutoff in A A A and, by the lower bound of Eq. (40) , in B B B , while m i < − ϵ ≤ 0 m_i < -\epsilon\le0 m i < − ϵ ≤ 0 places it below in A A A and, by the upper bound, in B B B ; when no margin m i m_i m i falls in [ − ϵ , ρ ϵ ] [-\epsilon,\rho\epsilon] [ − ϵ , ρ ϵ ] the two counts agree.
The certificate says nothing about eigenvectors, which require spectral-gap information in addition, and it is silent on a run with a margin in the ambiguous band, where the two ranks may or may not agree. It is evaluated in floating point, with ϵ \epsilon ϵ taken as the trace difference and the margins from the computed spectra, and is not a rounding-rigorous certificate; Sec. IV A reports how far the margins of the certified runs stay from the two decision boundaries.
Concentration bounds and proofs for the protocol
Concentration of the two metrics
Proposition 6 (W 1 W_1 W 1 concentration). Let p , q ∈ Δ N p,q \in\Delta_N p , q ∈ Δ N , let p p p be known exactly, and let q ^ m \hat{q}_m q ^ m be the empirical distribution from m m m i.i.d. samples from the sampled distribution q q q . Let C C C be a ground-cost matrix with D max = max x , y C ( x , y ) D_{\max} = \max_{x,y} C(x,y) D m a x = max x , y C ( x , y ) . Then, writing W ‾ m = E W 1 ( p , q ^ m ; C ) \overline{W}_m = \mathbb{E}W_1(p,\hat{q}_m;C) W m = E W 1 ( p , q ^ m ; C ) for the mean of the empirical estimate,
Pr ( ∣ W 1 ( p , q ^ m ; C ) − W ‾ m ∣ > t ) ≤ 2 exp ( − 2 m t 2 D max 2 ) (41) \Pr\left(\left|W_1(p,\hat{q}_m;C)-\overline{W}_m\right|>t\right) \le2\exp\left(-\frac{2mt^2}{D_{\max}^2}\right) \tag*{(41)} Pr ( W 1 ( p , q ^ m ; C ) − W m > t ) ≤ 2 exp ( − D m a x 2 2 m t 2 ) (41) and the bias of the estimate is bounded by
∣ W ‾ m − W 1 ( p , q ; C ) ∣ ≤ D max 2 m ∑ x q ( x ) ≤ D max N 2 m , (42) \left|\overline{W}_m-W_1(p,q;C)\right| \le\frac{D_{\max}}{2\sqrt{m}}\sum_x\sqrt{q(x)} \le\frac{D_{\max}\sqrt{N}}{2\sqrt{m}}, \tag*{(42)} W m − W 1 ( p , q ; C ) ≤ 2 m D m a x x ∑ q ( x ) ≤ 2 m D m a x N , (42) in which the sum runs over the distribution that is sampled. If both distributions are sampled, with m p m_p m p and m q m_q m q i.i.d. samples, the same bounded-differences argument, in which one sample of p p p moves the statistic by at most D max / m p D_{\max}/m_p D m a x / m p and one sample of q q q by at most D max / m q D_{\max}/m_q D m a x / m q , gives
Pr ( ∣ W 1 ( p ^ m p , q ^ m q ; C ) − E W 1 ( p ^ m p , q ^ m q ; C ) ∣ > t ) ≤ 2 exp ( − 2 t 2 D max 2 ( 1 / m p + 1 / m q ) ) (43) \Pr\left(\left|W_1(\hat{p}_{m_p},\hat{q}_{m_q};C)-\mathbb{E}W_1(\hat{p}_{m_p},\hat{q}_{m_q};C)\right|>t\right)
\le2\exp\left(-\frac{2t^2}{D_{\max}^2(1/m_p+1/m_q)}\right) \tag*{(43)} Pr ( W 1 ( p ^ m p , q ^ m q ; C ) − E W 1 ( p ^ m p , q ^ m q ; C ) > t ) ≤ 2 exp ( − D m a x 2 ( 1/ m p + 1/ m q ) 2 t 2 ) (43) whose exponent at m p = m q = m m_p=m_q=m m p = m q = m is − m t 2 / D max 2 -mt^2/D_{\max}^2 − m t 2 / D m a x 2 , half the one-sample rate; the bias is at most
1 2 D max ( m p − 1 / 2 ∑ x p ( x ) + m q − 1 / 2 ∑ x q ( x ) ) , \frac{1}{2}D_{\max}\left(m_p^{-1/2}\sum_x\sqrt{p(x)}+m_q^{-1/2}\sum_x\sqrt{q(x)}\right), 2 1 D m a x ( m p − 1/2 x ∑ p ( x ) + m q − 1/2 x ∑ q ( x ) ) , the sum of the two one-sample terms, by the triangle inequality, and since W 1 W_1 W 1 is jointly convex in ( p , q ) (p,q) ( p , q ) , Jensen’s inequality still gives E W 1 ( p ^ m p , q ^ m q ; C ) ≥ W 1 ( p , q ; C ) \mathbb{E}W_1(\hat{p}_{m_p},\hat{q}_{m_q};C)\ge W_1(p,q;C) E W 1 ( p ^ m p , q ^ m q ; C ) ≥ W 1 ( p , q ; C ) .
Corollary 7 (W 1 W_1 W 1 sample complexity). To achieve concentration about the mean, ∣ W 1 ( p , q ^ m ) − W ‾ m ∣ ≤ t \left|W_1(p,\hat{q}_m)-\overline{W}_m\right|\le t W 1 ( p , q ^ m ) − W m ≤ t with probability ≥ 1 − δ \ge1-\delta ≥ 1 − δ :
m ≥ D max 2 ln ( 2 / δ ) 2 t 2 . (44) m \ge\frac{D_{\max}^2\ln(2/\delta)}{2t^2}. \tag*{(44)} m ≥ 2 t 2 D m a x 2 ln ( 2/ δ ) . (44) For the total error ∣ W 1 ( p , q ^ m ) − W 1 ( p , q ) ∣ ≤ t \left|W_1(p,\hat{q}_m)-W_1(p,q)\right|\le t ∣ W 1 ( p , q ^ m ) − W 1 ( p , q ) ∣ ≤ t the bias bound (B2) must be added: splitting t t t evenly between fluctuation and bias gives
m ≥ max { 2 D max 2 ln ( 2 / δ ) t 2 , D max 2 ( ∑ x q ( x ) ) 2 t 2 } . (45) m \ge\max\left\{\frac{2D_{\max}^2\ln(2/\delta)}{t^2},\frac{D_{\max}^2\left(\sum_x\sqrt{q(x)}\right)^2}{t^2}\right\}. \tag*{(45)} m ≥ max ⎩ ⎨ ⎧ t 2 2 D m a x 2 ln ( 2/ δ ) , t 2 D m a x 2 ( ∑ x q ( x ) ) 2 ⎭ ⎬ ⎫ . (45) For our ground costs D max = O ( n ) D_{\max}=O(n) D m a x = O ( n ) , so the fluctuation complexity is O ( n 2 / t 2 ) O(n^2/t^2) O ( n 2 / t 2 ) ; this bias term carries ∑ x q ( x ) \sum_x\sqrt{q(x)} ∑ x q ( x ) of the sampled distribution, which is O ( 1 ) O(1) O ( 1 ) when the generator’s output is sparse, as for a near-exact fit of the two-peak B 1 B1 B 1 target, and at most 2 n / 2 2^{n/2} 2 n /2 for a full-support generator [30 ] ; the sparsity of an exact target does not enter, because the target is not sampled.
Proposition 8 (MMD concentration). Let k : X × X → R k:\mathcal{X}\times\mathcal{X}\to\mathbb{R} k : X × X → R be a bounded kernel whose values lie in an interval of width K max K_{\max} K m a x ; for the RBF–Hamming kernel K max = 1 K_{\max}=1 K m a x = 1 . Let p ^ m , q ^ m \hat{p}_m,\hat{q}_m p ^ m , q ^ m be empirical distributions from m m m i.i.d. samples each. The unbiased estimator MMD ^ u 2 \widehat{\operatorname{MMD}}_u^2 MMD u 2 satisfies
Pr ( MMD ^ u 2 − MMD k 2 ( p , q ) > t ) ≤ exp ( − m t 2 16 K max 2 ) . (46) \Pr\left(\widehat{\operatorname{MMD}}_u^2-\operatorname{MMD}_k^2(p,q)>t\right)\le\exp\left(-\frac{mt^2}{16K_{\max}^2}\right). \tag*{(46)} Pr ( MMD u 2 − MMD k 2 ( p , q ) > t ) ≤ exp ( − 16 K m a x 2 m t 2 ) . (46) The MMD sample complexity is m ≥ 16 K max 2 ln ( 1 / δ ) / t 2 m\ge16K_{\max}^2\ln(1/\delta)/t^2 m ≥ 16 K m a x 2 ln ( 1/ δ ) / t 2 . With K max = 1 K_{\max}=1 K m a x = 1 (RBF kernel), this is O ( ln ( 1 / δ ) / t 2 ) O(\ln(1/\delta)/t^2) O ( ln ( 1/ δ ) / t 2 ) , independent of qubit count n n n and support size N N N .
Equal-budget comparison and window sizing
Fix a total budget of M M M samples per sampled distribution, counted across both stages. The W 1 W_1 W 1 -only rule spends all of it on one statistic; D 4 D_4 D 4 splits it disjointly and gives each metric M / 2 M/2 M /2 (Table XIV ). Substituting into Propositions 6 and 8, the three exponential rates, meaning the coefficient of − M -M − M in each exponent, are
Table XIV. Sample allocation behind Eqs. (B7) and (B9), per sampled distribution and across both stages. In the one-sample model only the generator is sampled; in the two-sample model the target is sampled as well, with the same counts.
r D 1 = 2 t 2 D max 2 , r m i s s = η 2 D max 2 , r_{D_1}=\frac{2t^2}{D_{\max}^2},\qquad r_{\mathrm{miss}}=\frac{\eta^2}{D_{\max}^2}, r D 1 = D m a x 2 2 t 2 , r miss = D m a x 2 η 2 ,
r j o i n t = t 2 D max 2 + ξ 2 32 K max 2 . (47) r_{\mathrm{joint}}=\frac{t^2}{D_{\max}^2}+\frac{\xi^2}{32K_{\max}^2}. \tag*{(47)} r joint = D m a x 2 t 2 + 32 K m a x 2 ξ 2 . (47)
in the one-sample model (exact target, sampled generator). Halving the sample size costs the joint term exactly half of D 1 D_1 D 1 ’s rate, and the MMD margin has to make that back. The joint term therefore has the larger rate at equal total budget if and only if
ξ > 4 2 K max t D max , (48) \xi>\frac{4\sqrt{2}K_{\max}t}{D_{\max}}, \tag*{(48)} ξ > D m a x 4 2 K m a x t , (48)
and the rate of the full bound of Eq. (13) exceeds the single-metric rate when in addition the escalation-miss term is subordinate, r m i s s ≥ 2 r D 1 r_{\mathrm{miss}}\ge2r_{D_1} r miss ≥ 2 r D 1 , that is η ≥ 2 t \eta\ge2t η ≥ 2 t ; the joint term retains the full quadratic rate, r j o i n t ≥ 2 r D 1 r_{\mathrm{joint}}\ge2r_{D_1} r joint ≥ 2 r D 1 , when ξ ≥ 4 6 K max t / D max \xi\ge4\sqrt{6}K_{\max}t/D_{\max} ξ ≥ 4 6 K m a x t / D m a x . The corresponding per-metric-budget requirement is ξ ≥ 4 2 K max t / D max \xi\ge4\sqrt{2}K_{\max}t/D_{\max} ξ ≥ 4 2 K m a x t / D m a x , so the fair normalization stiffens the requirement on the MMD margin by exactly 3 \sqrt{3} 3 and the requirement on the window from 2 t \sqrt{2}t 2 t to 2 t 2t 2 t ; nothing else changes. At K max = 1 K_{\max}=1 K m a x = 1 and the Hamming cost D max = n D_{\max}=n D m a x = n , Eq. (48) reads ξ > 5.66 t / n \xi>5.66\,t/n ξ > 5.66 t / n for break-even and ξ ≥ 9.80 t / n \xi\ge9.80\,t/n ξ ≥ 9.80 t / n for the full quadratic, both of which relax as n n n grows. When both distributions are sampled, with M M M draws per distribution, the two transport rates halve by Eq. (43) while the discrepancy rate, for which Proposition 8 is already a two-sample statement, does not:
r D 1 = t 2 D max 2 , r m i s s = η 2 2 D max 2 , r_{D_1}=\frac{t^2}{D_{\max}^2},\qquad r_{\mathrm{miss}}=\frac{\eta^2}{2D_{\max}^2}, r D 1 = D m a x 2 t 2 , r miss = 2 D m a x 2 η 2 ,
r j o i n t = t 2 2 D max 2 + ξ 2 32 K max 2 . (49) r_{\mathrm{joint}}=\frac{t^2}{2D_{\max}^2}+\frac{\xi^2}{32K_{\max}^2}. \tag*{(49)} r joint = 2 D m a x 2 t 2 + 32 K m a x 2 ξ 2 . (49)
so the conditions become ξ > 4 K max t / D max \xi>4K_{\max}t/D_{\max} ξ > 4 K m a x t / D m a x for break-even and ξ ≥ 4 3 K max t / D max \xi\ge4\sqrt{3}K_{\max}t/D_{\max} ξ ≥ 4 3 K m a x t / D m a x for the full quadratic rate (4.00 t / n t/n t / n and 6.93 t / n t/n t / n at the Hamming cost), with η ≥ 2 t \eta\ge2t η ≥ 2 t unchanged; if M M M instead counts the target and generator draws together, every rate in Eq. (49) is halved and the three conditions are the same. These are comparisons of exponential rates, which order the guarantees asymptotically in M M M : at finite M M M the prefactors and the additive miss term remain, so a larger rate does not by itself make the sum of the two D 4 D_4 D 4 terms smaller than the D 1 D_1 D 1 bound at every budget, and § 4 J measures the realized rates. At equal cost, then, the quadratic advantage holds only for defects whose discrepancy margin is large relative to the transport accuracy demanded. Appendix B 4 measures what that margin costs to resolve.
The escalation-miss term δ W 1 ( η ) \delta_{W_1}(\eta) δ W 1 ( η ) decays in η \eta η at the same Gaussian rate at which δ W 1 ( t ) \delta_{W_1}(t) δ W 1 ( t ) decays in t t t , so a fixed multiplicative widening of the window suppresses the miss term below any fixed multiple of the product
term. Taking η ≥ 2 t \eta\ge\sqrt{2}t η ≥ 2 t already gives δ W 1 ( η ) ≤ δ W 1 ( t ) 2 / 2 \delta_{W_1}(\eta)\le\delta_{W_1}(t)^2/2 δ W 1 ( η ) ≤ δ W 1 ( t ) 2 /2 , so the additive term is at most the square of the single-metric error and the quadratic character of Eq. (13) is preserved independently of δ M M D \delta_{\mathrm{MMD}} δ MMD . Unlike the sample size, the window can therefore be set generously, at one to two concentration margins, at no cost in the bound and with no effect on the in-zone false-fail bound of § 4 B; the realized false-fail rate does grow with the window, because a wider window escalates more good pairs and subjects them to the AND rule (§ 4 J).
Proofs of Theorems 3 and 4
Theorem 3. Write N = 2 n N=2^n N = 2 n and d = q S , β − u = β N − 1 χ S d=q_{S,\beta}-u=\beta N^{-1}\chi_S d = q S , β − u = β N − 1 χ S . (a) The excess of q S , β q_{S,\beta} q S , β over u u u sits on the N / 2 N/2 N /2 strings with χ S = + 1 \chi_S=+1 χ S = + 1 and the deficit on their complements, each of total mass β / 2 \beta/2 β /2 , so q = β / 2 q=\beta/2 q = β /2 . Flipping any one bit of S S S is a bijection from the excess set onto the deficit set at Hamming cost 1, so the coupling that moves the excess β N − 1 \beta N^{-1} β N − 1 of each excess string to its lip has cost β / 2 \beta/2 β /2 and W 1 ≤ β / 2 W_1\le\beta/2 W 1 ≤ β /2 ; the potential f = χ S / 2 f=\chi_S/2 f = χ S /2 changes by at most 1 across any single bit flip, hence is 1-Lipschitz for C 1 C_1 C 1 , and ⟨ f , q S , β − u ⟩ = β / 2 \langle f,q_{S,\beta}-u\rangle=\beta/2 ⟨ f , q S , β − u ⟩ = β /2 , so W 1 ≥ β / 2 W_1\ge\beta/2 W 1 ≥ β /2 by Eq. (2) . (b) By Eq. (25) and the orthonormality of the parities, E q S , β χ S − E u χ S = β 1 [ S = S ′ ] E_{q_{S,\beta}}\chi_S-E_u\chi_S=\beta1[S=S'] E q S , β χ S − E u χ S = β 1 [ S = S ′ ] , so the sum collapses to the single term β 2 c n − ∣ S ∣ b ∣ S ∣ \beta^2c^{n-|S|}b^{|S|} β 2 c n − ∣ S ∣ b ∣ S ∣ ; at y = 1 / n y=1/n y = 1/ n , b = ( 2 n ) − 1 [ 1 + O ( n − 1 ) ] b=(2n)^{-1}[1+O(n^{-1})] b = ( 2 n ) − 1 [ 1 + O ( n − 1 )] and c n − r = e − 1 / 2 [ 1 + O ( n − 1 ) ] c^{n-r}=e^{-1/2}[1+O(n^{-1})] c n − r = e − 1/2 [ 1 + O ( n − 1 )] . The circuit prepares every qubit outside S S S and all but one qubit of S S S in ∣ + ⟩ |+\rangle ∣ + ⟩ , the remaining qubit t ∈ S t\in S t ∈ S in ( 1 + β ) / 2 ∣ 0 ⟩ + ( 1 − β ) / 2 ∣ 1 ⟩ \sqrt{(1+\beta)/2}|0\rangle+\sqrt{(1-\beta)/2}|1\rangle ( 1 + β ) /2 ∣0 ⟩ + ( 1 − β ) /2 ∣1 ⟩ , and applies a CNOT with control any other qubit of S S S into t t t : the Born distribution is uniform on every qubit but t t t , the parity χ S \chi_S χ S equals ( − 1 ) t 0 (-1)^{t_0} ( − 1 ) t 0 for the pre-CNOT value t 0 t_0 t 0 of the target, and E χ S = β E\chi_S=\beta E χ S = β (we verified the identity ∣ ⟨ x ∣ U ∣ 0 ⟩ ∣ 2 = q S , β ( x ) | \langle x|U|0\rangle|^2=q_{S,\beta}(x) ∣ ⟨ x ∣ U ∣0 ⟩ ∣ 2 = q S , β ( x ) for every S S S at n ≤ 6 n\le6 n ≤ 6 ). (c) For a positive-semidefinite kernel matrix K K K , a S = N − 2 χ S T K χ S ≥ 0 a_S=N^{-2}\chi_S^{\mathsf T}K\chi_S\ge0 a S = N − 2 χ S T K χ S ≥ 0 and MMD K 2 ( u , q S , β ) = d T K d = β 2 a S \operatorname{MMD}_K^2(u,q_{S,\beta})=d^{\mathsf T}Kd=\beta^2a_S MMD K 2 ( u , q S , β ) = d T K d = β 2 a S . Since ∑ S χ S ( x ) χ S ( y ) = N 1 [ x = y ] \sum_S\chi_S(x)\chi_S(y)=N1[x=y] ∑ S χ S ( x ) χ S ( y ) = N 1 [ x = y ] , ∑ S a S = N − 1 Tr ( K ∑ S χ S χ S T ) = N − 1 Tr K = 1 \sum_Sa_S=N^{-1}\operatorname{Tr}(K\sum_S\chi_S\chi_S^{\mathsf T})=N^{-1}\operatorname{Tr}K=1 ∑ S a S = N − 1 Tr ( K ∑ S χ S χ S T ) = N − 1 Tr K = 1 when k ( x , x ) = 1 k(x,x)=1 k ( x , x ) = 1 , so the N − 1 N-1 N − 1 nonempty parities share at most one unit of weight and the smallest receives at most 1 / ( N − 1 ) 1/(N-1) 1/ ( N − 1 ) . For k ∗ , K ∗ = ( N I − 11 T ) / ( N − 1 ) k_*,K_*=(NI-11^{\mathsf T})/(N-1) k ∗ , K ∗ = ( N I − 1 1 T ) / ( N − 1 ) is positive semidefinite with unit diagonal, and χ S K ∗ χ S = N 2 / ( N − 1 ) \chi_SK_*\chi_S=N^2/(N-1) χ S K ∗ χ S = N 2 / ( N − 1 ) for every S ≠ ∅ S\ne\varnothing S = ∅ , so every nonempty parity receives exactly 1 / ( N − 1 ) 1/(N-1) 1/ ( N − 1 ) and the bound is attained.
Theorem 4. Let φ \varphi φ be the feature map of k k k , so that
∥ φ ( x ) − φ ( y ) ∥ 2 = 2 − 2 k ( x , y ) = 2 ( 1 − e − γ d H ( x , y ) ) \|\varphi(x)-\varphi(y)\|^2=2-2k(x,y)=2(1-e^{-\gamma d_H(x,y)}) ∥ φ ( x ) − φ ( y ) ∥ 2 = 2 − 2 k ( x , y ) = 2 ( 1 − e − γ d H ( x , y ) ) and adjacent strings are 2 ( 1 − e − γ ) \sqrt{2(1-e^{-\gamma})} 2 ( 1 − e − γ ) apart. Along a shortest Hamming path from x x x to y y y the triangle inequality gives ∥ φ ( x ) − φ ( y ) ∥ ≤ d H ( x , y ) 2 ( 1 − e − γ ) \|\varphi(x)-\varphi(y)\|\le d_H(x,y)\sqrt{2(1-e^{-\gamma})} ∥ φ ( x ) − φ ( y ) ∥ ≤ d H ( x , y ) 2 ( 1 − e − γ ) . For an optimal coupling π \pi π of ( p , q ) (p,q) ( p , q ) , MMD k ( p , q ) = ∥ ∑ x , y π ( x , y ) [ φ ( x ) − φ ( y ) ] ∥ ≤ ∑ x , y π ( x , y ) ∥ φ ( x ) − φ ( y ) ∥ ≤ 2 ( 1 − e − γ ) W 1 ( p , q ; C 1 ) \operatorname{MMD}_k(p,q)=\|\sum_{x,y}\pi(x,y)[\varphi(x)-\varphi(y)]\|\le\sum_{x,y}\pi(x,y)\|\varphi(x)-\varphi(y)\|\le\sqrt{2(1-e^{-\gamma})}W_1(p,q;C_1) MMD k ( p , q ) = ∥ ∑ x , y π ( x , y ) [ φ ( x ) − φ ( y )] ∥ ≤ ∑ x , y π ( x , y ) ∥ φ ( x ) − φ ( y ) ∥ ≤ 2 ( 1 − e − γ ) W 1 ( p , q ; C 1 ) , which is Eq. (27) after squaring. For p = δ x p=\delta_x p = δ x and q = ( 1 − w ) δ x + w δ y q=(1-w)\delta_x+w\delta_y q = ( 1 − w ) δ x + w δ y with d H ( x , y ) = 1 d_H(x,y)=1 d H ( x , y ) = 1 , W 1 = w W_1=w W 1 = w and MMD k 2 = w 2 ∥ φ ( x ) − φ ( y ) ∥ 2 = 2 ( 1 − e − γ ) w 2 \operatorname{MMD}_k^2=w^2\|\varphi(x)-\varphi(y)\|^2=2(1-e^{-\gamma})w^2 MMD k 2 = w 2 ∥ φ ( x ) − φ ( y ) ∥ 2 = 2 ( 1 − e − γ ) w 2 , so the bound is attained. If τ M M D ≥ 2 ( 1 − e − γ ) τ W 1 2 \tau_{\mathrm{MMD}}\ge2(1-e^{-\gamma})\tau_{W_1}^2 τ MMD ≥ 2 ( 1 − e − γ ) τ W 1 2 , every pair with W 1 ≤ τ W 1 W_1\le\tau_{W_1} W 1 ≤ τ W 1 has MMD k 2 ≤ τ M M D \operatorname{MMD}_k^2\le\tau_{\mathrm{MMD}} MMD k 2 ≤ τ MMD by Eq. (27) , so the discrepancy vote passes whenever the transport vote does and D 4 = D 1 ( τ W 1 ) D_4=D_1(\tau_{W_1}) D 4 = D 1 ( τ W 1 ) for every η \eta η ; if τ M M D < 2 ( 1 − e − γ ) τ W 1 2 \tau_{\mathrm{MMD}}<2(1-e^{-\gamma})\tau_{W_1}^2 τ MMD < 2 ( 1 − e − γ ) τ W 1 2 , the adjacent pair at w = τ W 1 ≤ 1 w=\tau_{W_1}\le1 w = τ W 1 ≤ 1 has W 1 = τ W 1 W_1=\tau_{W_1} W 1 = τ W 1 , which lies in every window ( τ W 1 − η , τ W 1 ] (\tau_{W_1}-\eta,\tau_{W_1}] ( τ W 1 − η , τ W 1 ] , and MMD k 2 > τ M M D \operatorname{MMD}_k^2>\tau_{\mathrm{MMD}} MMD k 2 > τ MMD , so D 4 D_4 D 4 rejects a pair that D 1 ( τ W 1 ) D_1(\tau_{W_1}) D 1 ( τ W 1 ) admits.
What certification-grade thresholds cost in shots
The escalation’s advantage appears only at τ M M D = 10 − 4 \tau_{\mathrm{MMD}}=10^{-4} τ MMD = 1 0 − 4 , and resolving a tolerance that tight costs measurements.
The certificate of Proposition 8 demands m ≥ 16 K max 2 ln ( 1 / δ ) / t 2 m\ge16K_{\max}^2\ln(1/\delta)/t^2 m ≥ 16 K m a x 2 ln ( 1/ δ ) / t 2 , i.e. m ≥ 4.8 × 10 9 m\ge4.8\times10^9 m ≥ 4.8 × 1 0 9 at t = 10 − 4 t=10^{-4} t = 1 0 − 4 ; conversely S = 10 5 S=10^5 S = 1 0 5 shots, the largest we test, certify MMD k 2 \operatorname{MMD}_k^2 MMD k 2 only to ± 2.2 × 10 − 2 \pm2.2\times10^{-2} ± 2.2 × 1 0 − 2 . The transport side (Corollary 7) is milder by three orders, so the discrepancy channel binds.
The certificate is loose, so we also measure the requirement directly, on the exact witnesses of both defect families placed at their window midpoints. A useful vote must return FAIL on the defect and PASS on a perfect generator. A biased estimator meets the first condition at any budget, so we require both. At power 0.95 and size 0.05, the mode-collapse witness needs S ≈ 10 5 S\approx10^5 S ≈ 1 0 5 shots at n = 3 , 4 n=3,4 n = 3 , 4 and n = 6 n=6 n = 6 –8 8 8 and 10 6 10^6 1 0 6 at n = 5 n=5 n = 5 , and the correlation-shift witness 10 5 10^5 1 0 5 at every scale through n = 7 n=7 n = 7 and 10 6 10^6 1 0 6 at n = 8 n=8 n = 8 . The measured requirement is four orders of magnitude below the certificate. It is still at or above the top of the range § 4 J explores, and a full order above it wherever the window is narrowest.
The second condition matters because the plug-in (V-statistic) estimator, which § 4 J reports for comparison, carries a positive bias. For a uniform reference the bias is ( 1 − E y , y ′ ∼ q [ k ( y , y ′ ) ] ) / m (1-\mathbb{E}_{y,y'\sim q}[k(y,y')])/m ( 1 − E y , y ′ ∼ q [ k ( y , y ′ )]) / m , which equals Φ n / m \Phi_n/m Φ n / m for a perfect generator and Φ n ( 1 − λ 2 ) / m \Phi_n(1-\lambda^2)/m Φ n ( 1 − λ 2 ) / m for the mode-collapse witness, with Φ n \Phi_n Φ n the bracket of Eq. (23) , the same quantity that carries the separation in Theorem 2. Since the witness has MMD k 2 = λ 2 Φ n \operatorname{MMD}_k^2=\lambda^2\Phi_n MMD k 2 = λ 2 Φ n , plug-in bias equals signal at m = ( 1 − λ 2 ) / λ 2 ≈ λ − 2 m=(1-\lambda^2)/\lambda^2\approx\lambda^{-2} m = ( 1 − λ 2 ) / λ 2 ≈ λ − 2 , independently of n n n and of the bandwidth, which is 1.6 × 10 3 1.6\times10^3 1.6 × 1 0 3 shots at n = 3 n=3 n = 3 rising to 1.1 × 10 4 1.1\times10^4 1.1 × 1 0 4 at n = 8 n=8 n = 8 . Below that budget the estimator exceeds τ M M D = 10 − 4 \tau_{\mathrm{MMD}}=10^{-4} τ MMD = 1 0 − 4 on a defective and a perfect generator alike, its size reaching 1.00 at S = 10 2 S=10^2 S = 1 0 2 and still 0.92 at S = 10 3 S=10^3 S = 1 0 3 for n = 3 n=3 n = 3 . This is the discrepancy-side counterpart of the transport bias of § 4 J. The unbiased U-statistic removes the term at the cost of a larger variance and reaches a usable operating point at S ≈ 10 5 S\approx10^5 S ≈ 1 0 5 .
For correlation defects the budget carries the additional Θ ( n 4 / β 4 ) \Theta(n^4/\beta^4) Θ ( n 4 / β 4 ) scaling of § 4 E.
Proofs
Proof of Proposition 6. We first show that W 1 ( r , q ; C ) W_1(r,q;C) W 1 ( r , q ; C ) is D max D_{\max} D m a x -Lipschitz with respect to total variation. Let r , r ′ ∈ Δ N r,r'\in\Delta_N r , r ′ ∈ Δ N and let Γ ∗ \Gamma^* Γ ∗ be an optimal coupling for ( r , q ) (r,q) ( r , q ) . Define the signed measure δ = r ′ − r \delta=r'-r δ = r ′ − r . Then any coupling Γ ′ \Gamma' Γ ′ for ( r ′ , q ) (r',q) ( r ′ , q ) can be written as Γ ′ = Γ ∗ + Δ \Gamma'=\Gamma^*+\Delta Γ ′ = Γ ∗ + Δ where Δ \Delta Δ has marginals ( δ , 0 ) (\delta,0) ( δ , 0 ) plus a correction. More precisely, for any r , r ′ r,r' r , r ′ :
∣ W 1 ( r , q ; C ) − W 1 ( r ′ , q ; C ) ∣ ≤ D max ⋅ TV ( r , r ′ ) , (50) |W_1(r,q;C)-W_1(r',q;C)|\le D_{\max}\cdot\operatorname{TV}(r,r'), \tag*{(50)} ∣ W 1 ( r , q ; C ) − W 1 ( r ′ , q ; C ) ∣ ≤ D m a x ⋅ TV ( r , r ′ ) , (50)
because the optimal transport plan can be modified to absorb the mass difference r ′ − r r'-r r ′ − r at cost at most D max D_{\max} D m a x per unit of total variation. This follows from the dual formulation (2) . Any C C C -Lipschitz potential f f f has oscillation max x f ( x ) − min x f ( x ) ≤ D max \max_x f(x)-\min_x f(x)\le D_{\max} max x f ( x ) − min x f ( x ) ≤ D m a x , and because the signed measure r − r ′ r-r' r − r ′ has zero total mass, ⟨ f , r − r ′ ⟩ \langle f,r-r'\rangle ⟨ f , r − r ′ ⟩ is unchanged by shifting f f f by a constant; centering f f f at c = 1 2 ( max x f + min x f ) c=\frac{1}{2}(\max_x f+\min_x f) c = 2 1 ( max x f + min x f ) gives ∥ f − c ∥ ∞ ≤ D max / 2 \|f-c\|_\infty\le D_{\max}/2 ∥ f − c ∥ ∞ ≤ D m a x /2 , so
∣ ⟨ f , r − q ⟩ − ⟨ f , r ′ − q ⟩ ∣ = ∣ ⟨ f , r − r ′ ⟩ ∣ = ∣ ⟨ f − c , r − r ′ ⟩ ∣ ≤ D max 2 ⋅ ∥ r − r ′ ∥ 1 = D max ⋅ TV ( r , r ′ ) . |\langle f,r-q\rangle-\langle f,r'-q\rangle|=|\langle f,r-r'\rangle|=|\langle f-c,r-r'\rangle|\le\frac{D_{\max}}{2}\cdot\|r-r'\|_1=D_{\max}\cdot\operatorname{TV}(r,r'). ∣ ⟨ f , r − q ⟩ − ⟨ f , r ′ − q ⟩ ∣ = ∣ ⟨ f , r − r ′ ⟩ ∣ = ∣ ⟨ f − c , r − r ′ ⟩ ∣ ≤ 2 D m a x ⋅ ∥ r − r ′ ∥ 1 = D m a x ⋅ TV ( r , r ′ ) .
Taking the supremum over C C C -Lipschitz f f f gives the result, and by the symmetry of W 1 W_1 W 1 in its two arguments the same bound holds in the second argument, which is the one the observation model uses.
Now apply McDiarmid’s inequality [42 ] . Let X 1 , … , X m X_1,\ldots,X_m X 1 , … , X m be i.i.d. draws from q q q , and let q ^ m = 1 m ∑ i δ X i \hat q_m=\frac{1}{m}\sum_i\delta_{X_i} q ^ m = m 1 ∑ i δ X i . Replacing X i X_i X i by X i ′ X'_i X i ′ moves mass 1 / m 1/m 1/ m from one outcome to another, so q ^ m \hat q_m q ^ m changes by 1 / m 1/m 1/ m in total variation under the convention TV = 1 2 ∥ ⋅ ∥ 1 \operatorname{TV}=\frac{1}{2}\|\cdot\|_1 TV = 2 1 ∥ ⋅ ∥ 1 (a change of 2 / m 2/m 2/ m in L 1 L^1 L 1 ), and by the Lipschitz bound it changes W 1 ( p , q ^ m ; C ) W_1(p,\hat q_m;C) W 1 ( p , q ^ m ; C ) by at most D max / m D_{\max}/m D m a x / m . By McDiarmid’s inequality with bounded differences c i = D max / m c_i=D_{\max}/m c i = D m a x / m :
Pr ( ∣ W 1 ( p , q ^ m ; C ) − E [ W 1 ( p , q ^ m ; C ) ] ∣ > t ) ≤ 2 exp ( − 2 t 2 ∑ i c i 2 ) = 2 exp ( − 2 m t 2 D max 2 ) . \Pr\left(|W_1(p,\hat q_m;C)-\mathbb{E}[W_1(p,\hat q_m;C)]|>t\right)\le2\exp\left(-\frac{2t^2}{\sum_i c_i^2}\right)=2\exp\left(-\frac{2mt^2}{D_{\max}^2}\right). Pr ( ∣ W 1 ( p , q ^ m ; C ) − E [ W 1 ( p , q ^ m ; C )] ∣ > t ) ≤ 2 exp ( − ∑ i c i 2 2 t 2 ) = 2 exp ( − D m a x 2 2 m t 2 ) .
This establishes Eq. (41) (concentration about the mean). For the bias Eq. (42) , the Lipschitz bound (2) gives ∣ W ‾ m − W 1 ( p , q ; C ) ∣ ≤ D max E [ TV ( q ^ m , q ) ] |\overline W_m-W_1(p,q;C)|\le D_{\max}\mathbb{E}[\operatorname{TV}(\hat q_m,q)] ∣ W m − W 1 ( p , q ; C ) ∣ ≤ D m a x E [ TV ( q ^ m , q )] , and E [ TV ( q ^ m , q ) ] = 1 2 ∑ x E ∣ q ^ m ( x ) − q ( x ) ∣ ≤ 1 2 m ∑ x q ( x ) \mathbb{E}[\operatorname{TV}(\hat q_m,q)]=\frac{1}{2}\sum_x\mathbb{E}|\hat q_m(x)-q(x)|\le\frac{1}{2\sqrt{m}}\sum_x\sqrt{q(x)} E [ TV ( q ^ m , q )] = 2 1 ∑ x E ∣ q ^ m ( x ) − q ( x ) ∣ ≤ 2 m 1 ∑ x q ( x ) ; Cauchy–Schwarz gives the worst-case N \sqrt{N} N factor (cf. Fourier and Guillemin [43 ] , Theorem 1 specialized to finite support). □ \square □
Proof of Proposition 8. This is Theorem 10 of Gretton et al. [7 ] . The unbiased estimator MMD ^ 2 \widehat{\operatorname{MMD}}^2 MMD 2 is a one-sample U U U -statistic of the pooled sample of size 2 m 2m 2 m . Each observation enters O ( m ) O(m) O ( m ) terms, and each term varies by at most K max K_{\max} K m a x . Replacing one observation changes MMD ^ u 2 \widehat{\operatorname{MMD}}_{u}^{2} MMD u 2 by at most 4 K max / m 4K_{\max}/m 4 K m a x / m (since each observation appears in at most 2 ( 2 m − 1 ) 2(2m-1) 2 ( 2 m − 1 ) of the ( 2 m 2 ) \binom{2m}{2} ( 2 2 m ) kernel evaluations, contributing O ( K max / m ) O(K_{\max}/m) O ( K m a x / m ) each). McDiarmid’s inequality with 2 m 2m 2 m variables and bounded differences c i = 4 K max / m c_i=4K_{\max}/m c i = 4 K m a x / m gives the stated exponential tail. The same argument covers the one-sample estimator of Sec. V J, in which one distribution is known exactly and m m m samples are drawn from the other: replacing one sample changes the U U U -statistic by at most 2 K max / m 2K_{\max}/m 2 K m a x / m through the pairwise kernel average and by at most 2 K max / m 2K_{\max}/m 2 K m a x / m through the cross term, so McDiarmid’s inequality with m m m variables and c i = 4 K max / m c_i=4K_{\max}/m c i = 4 K m a x / m gives exp ( − m t 2 / 8 K max 2 ) \exp(-mt^{2}/8K_{\max}^{2}) exp ( − m t 2 /8 K m a x 2 ) , which implies Eq. (46) . McDiarmid’s inequality is two-sided, so the same tail bounds the downward deviation that a false pass in Eq. (13) requires.
Proof of the false-pass bounds (13) and (14). Throughout, a false pass requires w ^ \hat{w} w ^ to fall below the true value, and since W ‾ m ≥ W 1 ( p , q ; C ) \overline{W}_{m} \ge W_{1}(p,q;C) W m ≥ W 1 ( p , q ; C ) by convexity of W 1 ( p , ⋅ ; C ) W_{1}(p,\cdot;C) W 1 ( p , ⋅ ; C ) and Jensen’s inequality, such a deviation from the true value is at least as large a deviation from the mean, so Proposition 6 applies without the bias term of Eq. (42) .
Outside the margin zone. Outside M η \mathcal{M}_{\eta} M η the true W 1 W_{1} W 1 value differs from the threshold by at least η \eta η . A D 4 D_{4} D 4 PASS always requires the W 1 W_{1} W 1 vote to pass, so a false pass requires the empirical W 1 W_{1} W 1 verdict to be wrong, an estimation error exceeding η \eta η ; by Proposition 6 this has probability at most δ W 1 ( η ) \delta_{W_{1}}(\eta) δ W 1 ( η ) , whether or not escalation fires, which is Eq. (12) .
Inside the margin zone. A D 4 D_{4} D 4 PASS falls into one of two disjoint events determined by w ^ \hat{w} w ^ : the clear case P 1 = { w ^ ≤ τ W 1 − η } P_{1}=\{\hat{w}\le\tau_{W_{1}}-\eta\} P 1 = { w ^ ≤ τ W 1 − η } (no escalation), and the escalated case P 2 = { τ W 1 − η < w ^ ≤ τ W 1 } ∩ { v M M D = 1 } P_{2}=\{\tau_{W_{1}}-\eta<\hat{w}\le\tau_{W_{1}}\}\cap\{v_{\mathrm{MMD}}=1\} P 2 = { τ W 1 − η < w ^ ≤ τ W 1 } ∩ { v MMD = 1 } (both votes PASS); empirical values w ^ > τ W 1 \hat{w}>\tau_{W_{1}} w ^ > τ W 1 return FAIL on either branch, since a PASS requires v W 1 = 1 v_{W_{1}}=1 v W 1 = 1 . On P 1 P_{1} P 1 the pair is defective with W 1 ( p , q ; C ) = τ W 1 + Δ W_{1}(p,q;C)=\tau_{W_{1}+\Delta} W 1 ( p , q ; C ) = τ W 1 + Δ , so w ^ ≤ τ W 1 − η \hat{w}\le\tau_{W_{1}}-\eta w ^ ≤ τ W 1 − η forces an estimation error exceeding η + Δ > η \eta+\Delta>\eta η + Δ > η ; by Proposition 6, Pr ( P 1 ) ≤ δ W 1 ( η ) \Pr(P_{1})\le\delta_{W_{1}}(\eta) Pr ( P 1 ) ≤ δ W 1 ( η ) , and the protocol certifies without evaluating MMD, so no product structure is available. On P 2 P_{2} P 2 the W 1 W_{1} W 1 vote passes only if w ^ ≤ τ W 1 \hat{w}\le\tau_{W_{1}} w ^ ≤ τ W 1 , an error exceeding Δ \Delta Δ , giving Pr ( v W 1 = 1 ) ≤ δ W 1 ( Δ ) \Pr(v_{W_{1}}=1)\le\delta_{W_{1}}(\Delta) Pr ( v W 1 = 1 ) ≤ δ W 1 ( Δ ) ; the MMD vote passes only if its empirical statistic falls below τ M M D \tau_{\mathrm{MMD}} τ MMD while the population value sits ξ \xi ξ above threshold, giving Pr ( v M M D = 1 ) ≤ δ M M D ( ξ ) \Pr(v_{\mathrm{MMD}}=1)\le\delta_{\mathrm{MMD}}(\xi) Pr ( v MMD = 1 ) ≤ δ MMD ( ξ ) by Proposition 8. The two statistics use disjoint samples, so the misvotes are independent and Pr ( P 2 ) ≤ δ W 1 ( Δ ) δ M M D ( ξ ) \Pr(P_{2})\le\delta_{W_{1}}(\Delta)\delta_{\mathrm{MMD}}(\xi) Pr ( P 2 ) ≤ δ W 1 ( Δ ) δ MMD ( ξ ) . This is the acceptance region of an intersection–union test [31 ] . The union bound over P 1 ∪ P 2 P_{1}\cup P_{2} P 1 ∪ P 2 gives Eq. (13) .
Proof of Theorem 2. Write q λ − p = λ ( δ x 0 − p ) q_{\lambda}-p=\lambda(\delta_{x_{0}}-p) q λ − p = λ ( δ x 0 − p ) from Eq. (16) .
Step 1: W1 _1 1 (claim (a)). The signed measure q λ − p q_{\lambda}-p q λ − p has an excess λ ( 1 − 2 − n ) \lambda(1-2^{-n}) λ ( 1 − 2 − n ) at x 0 x_{0} x 0 and a deficit λ 2 − n \lambda2^{-n} λ 2 − n at each other point. It has a single source, x 0 x_{0} x 0 , so the plan is forced: it moves mass λ 2 − n \lambda2^{-n} λ 2 − n from x 0 x_{0} x 0 to each x x x at Hamming cost d H ( x 0 , x ) d_{H}(x_{0},x) d H ( x 0 , x ) . Hence
W 1 ( p , q λ ; C 1 ) = λ ∑ x 2 − n d H ( x 0 , x ) = λ n / 2 , (51) W_{1}(p,q_{\lambda};C_{1})=\lambda\sum_{x}2^{-n}d_{H}(x_{0},x)=\lambda n/2, \tag*{(51)} W 1 ( p , q λ ; C 1 ) = λ x ∑ 2 − n d H ( x 0 , x ) = λn /2 , (51)
using E x ∼ Unif [ d H ( x , x 0 ) ] = ∑ i = 1 n Pr [ x i ≠ ( x 0 ) i ] = n / 2 \mathbb{E}_{x\sim\operatorname{Unif}}[d_{H}(x,x_{0})]=\sum_{i=1}^{n}\Pr[x_{i}\ne(x_{0})_{i}]=n/2 E x ∼ Unif [ d H ( x , x 0 )] = ∑ i = 1 n Pr [ x i = ( x 0 ) i ] = n /2 .
Step 2: Total variation (claim (b)). TV ( p , q λ ) = 1 2 ∥ q λ − p ∥ 1 = 1 2 λ ∥ δ x 0 − p ∥ 1 = λ ( 1 − 2 − n ) \operatorname{TV}(p,q_{\lambda})=\frac{1}{2}\lVert q_{\lambda}-p\rVert_{1}=\frac{1}{2}\lambda\lVert\delta_{x_{0}}-p\rVert_{1}=\lambda(1-2^{-n}) TV ( p , q λ ) = 2 1 ∥ q λ − p ∥ 1 = 2 1 λ ∥ δ x 0 − p ∥ 1 = λ ( 1 − 2 − n ) .
Step 3: MMD (claim (c)). MMD is the RKHS norm of the mean-embedding difference, so MMD k ( p , q λ ) = ∥ μ p − μ q λ ∥ H k = λ ∥ μ δ x 0 − μ p ∥ H k = λ MMD k ( p , δ x 0 ) \operatorname{MMD}_{k}(p,q_{\lambda})=\lVert\mu_{p}-\mu_{q_{\lambda}}\rVert_{\mathcal{H}_{k}}=\lambda\lVert\mu_{\delta_{x_{0}}}-\mu_{p}\rVert_{\mathcal{H}_{k}}=\lambda\operatorname{MMD}_{k}(p,\delta_{x_{0}}) MMD k ( p , q λ ) = ∥ μ p − μ q λ ∥ H k = λ ∥ μ δ x 0 − μ p ∥ H k = λ MMD k ( p , δ x 0 ) . Set a : = E x ∼ p [ k ( x , x 0 ) ] a:=\mathbb{E}_{x\sim p}[k(x,x_{0})] a := E x ∼ p [ k ( x , x 0 )] . By the binomial expansion of the Hamming-distance generating function,
a = 2 − n ∑ d = 0 n ( n d ) e − γ d = ( 1 + e − γ 2 ) n . (52) a=2^{-n}\sum_{d=0}^{n}\binom{n}{d}e^{-\gamma d}=\left(\frac{1+e^{-\gamma}}{2}\right)^{n}. \tag*{(52)} a = 2 − n d = 0 ∑ n ( d n ) e − γ d = ( 2 1 + e − γ ) n . (52)
Translation invariance of d H d_{H} d H gives E x ′ ∼ p [ k ( x , x ′ ) ] = a \mathbb{E}_{x'\sim p}[k(x,x')]=a E x ′ ∼ p [ k ( x , x ′ )] = a for every fixed x x x , hence E x , x ′ ∼ p [ k ( x , x ′ ) ] = a \mathbb{E}_{x,x'\sim p}[k(x,x')]=a E x , x ′ ∼ p [ k ( x , x ′ )] = a ; with k ( x 0 , x 0 ) = 1 k(x_{0},x_{0})=1 k ( x 0 , x 0 ) = 1 the point-mass expansion yields MMD k 2 ( p , δ x 0 ) = a + 1 − 2 a = 1 − a \operatorname{MMD}_{k}^{2}(p,\delta_{x_{0}})=a+1-2a=1-a MMD k 2 ( p , δ x 0 ) = a + 1 − 2 a = 1 − a . Squaring the RKHS relation shows claim (c).
Step 4: Strict inclusion. For any λ \lambda λ in the interval (17) , claim (a) gives W 1 ( p , q λ ; C 1 ) = λ n / 2 ≤ τ W 1 W_{1}(p,q_{\lambda};C_{1})=\lambda n/2\le\tau_{W_{1}} W 1 ( p , q λ ; C 1 ) = λn /2 ≤ τ W 1 with W 1 > τ W 1 − η W_{1}>\tau_{W_{1}}-\eta W 1 > τ W 1 − η , so q λ ∉ R W 1 q_{\lambda}\notin\mathcal{R}_{W_{1}} q λ ∈ / R W 1 and the one-sided margin condition of D 4 \mathcal{D}_{4} D 4 holds, triggering the MMD escalation. Because MMD k 2 ( p , q λ ) \operatorname{MMD}_{k}^{2}(p,q_{\lambda}) MMD k 2 ( p , q λ ) is increasing in λ \lambda λ by claim (c), its minimum over the interval is at the left endpoint; if condition (18) holds there, then MMD k 2 ( p , q λ ) > τ M M D \operatorname{MMD}_{k}^{2}(p,q_{\lambda})>\tau_{\mathrm{MMD}} MMD k 2 ( p , q λ ) > τ MMD throughout, the MMD vote is FAIL, and the conservative AND rule returns FAIL. Thus q λ ∈ R b i ∖ R W 1 q_{\lambda}\in\mathcal{R}_{\mathrm{bi}}\setminus\mathcal{R}_{W_{1}} q λ ∈ R bi ∖ R W 1 , and since R W 1 ⊆ R b i \mathcal{R}_{W_{1}}\subseteq\mathcal{R}_{\mathrm{bi}} R W 1 ⊆ R bi by construction, R b i ⊋ R W 1 \mathcal{R}_{\mathrm{bi}}\supsetneq\mathcal{R}_{W_{1}} R bi ⊋ R W 1 . The operational claim is a direct substitution: at the knife-edge endpoint λ = 2 τ W 1 / n \lambda=2\tau_{W_{1}}/n λ = 2 τ W 1 / n (the limit η → 0 \eta\to0 η → 0 ) the right-hand side of (18) is 4.1 × 10 − 4 , 2.3 × 10 − 4 , 1.5 × 10 − 4 , 1.1 × 10 − 4 4.1\times10^{-4},2.3\times10^{-4},1.5\times10^{-4},1.1\times10^{-4} 4.1 × 1 0 − 4 , 2.3 × 1 0 − 4 , 1.5 × 1 0 − 4 , 1.1 × 1 0 − 4 at n = 3 , 4 , 5 , 6 n=3,4,5,6 n = 3 , 4 , 5 , 6 , each exceeding τ M M D = 10 − 4 \tau_{\mathrm{MMD}}=10^{-4} τ MMD = 1 0 − 4 ; the corresponding largest admissible windows are η max ( n ) = 0.025 , 0.017 , 0.009 , 0.001 \eta_{\max}(n)=0.025,0.017,0.009,0.001 η m a x ( n ) = 0.025 , 0.017 , 0.009 , 0.001 from Eq. (20) .
Transport value of the Family 2 correlation shift. Let p = Unif ( { 0 , 1 } n ) p=\operatorname{Unif}(\{0,1\}^{n}) p = Unif ({ 0 , 1 } n ) and q ( x ) = 2 − n [ 1 + β ( − 1 ) x 0 ⊕ x 1 ] q(x)=2^{-n}[1+\beta(-1)^{x_{0}\oplus x_{1}}] q ( x ) = 2 − n [ 1 + β ( − 1 ) x 0 ⊕ x 1 ] with β ∈ ( 0 , 1 ] \beta\in(0,1] β ∈ ( 0 , 1 ] , and let C 1 C_{1} C 1 be the Hamming cost. We show W 1 ( p , q ; C 1 ) = β / 2 W_{1}(p,q;C_{1})=\beta/2 W 1 ( p , q ; C 1 ) = β /2 .
Upper bound. The signed measure q − p q-p q − p carries + 2 − n β +2^{-n}\beta + 2 − n β on each of the 2 n − 1 2^{n-1} 2 n − 1 strings with x 0 = x 1 x_{0}=x_{1} x 0 = x 1 and − 2 − n β -2^{-n}\beta − 2 − n β on each of the 2 n − 1 2^{n-1} 2 n − 1 strings with x 0 ≠ x 1 x_{0}\ne x_{1} x 0 = x 1 , so the total excess is β / 2 \beta/2 β /2 . Pair each excess string x x x with the deficit string obtained by flipping x 0 x_{0} x 0 , a bijection between the two sets with d H = 1 d_{H}=1 d H = 1 on every pair. Transporting along it moves the whole excess at unit cost, so W 1 ( p , q ; C 1 ) ≤ β / 2 W_{1}(p,q;C_{1})\le\beta/2 W 1 ( p , q ; C 1 ) ≤ β /2 .
Lower bound. Take f ( x ) = 1 2 ( − 1 ) x 0 ⊕ x 1 f(x)=\frac{1}{2}(-1)^{x_{0}\oplus x_{1}} f ( x ) = 2 1 ( − 1 ) x 0 ⊕ x 1 . Then ∣ f ( x ) − f ( y ) ∣ ∈ { 0 , 1 } |f(x)-f(y)|\in\{0,1\} ∣ f ( x ) − f ( y ) ∣ ∈ { 0 , 1 } , and it equals 1 only when x x x and y y y differ in x 0 x_{0} x 0 or in x 1 x_{1} x 1 , in which case d H ( x , y ) ≥ 1 d_{H}(x,y)\ge1 d H ( x , y ) ≥ 1 ; hence f f f is C 1 C_{1} C 1 -Lipschitz. Its pairing with the discrepancy is
⟨ f , q − p ⟩ = ∑ x 1 2 ( − 1 ) x 0 ⊕ x 1 2 − n β ( − 1 ) x 0 ⊕ x 1 = β 2 , (53) \langle f,q-p\rangle=\sum_{x}\frac{1}{2}(-1)^{x_{0}\oplus x_{1}}2^{-n}\beta(-1)^{x_{0}\oplus x_{1}}=\frac{\beta}{2}, \tag*{(53)} ⟨ f , q − p ⟩ = x ∑ 2 1 ( − 1 ) x 0 ⊕ x 1 2 − n β ( − 1 ) x 0 ⊕ x 1 = 2 β , (53)
since the squared sign is one at every x x x . By Kantorovich duality [2 ] , W 1 ( p , q ; C 1 ) ≥ β / 2 W_1(p, q; C_1) \ge\beta/2 W 1 ( p , q ; C 1 ) ≥ β /2 .
The two bounds meet, so the value is exact. We have additionally confirmed it by exact linear programming at n = 2 n = 2 n = 2 –7 7 7 and β ∈ { 0.05 , 0.3 } \beta\in\{0.05, 0.3\} β ∈ { 0.05 , 0.3 } , where the primal optimum, β / 2 \beta/2 β /2 , and the dual pairing agree to machine precision. □