Variational quantum classifiers (VQCs) are increasingly evaluated on real quantum processors, yet their deployment behaviour can depend on post-measurement readout and variation across independently trained models. We study a four-qubit VQC for three-class traffic-state classification on two IBM Heron r2 backends. In a single-seed comparison, a four-channel linear-projection readout achieves higher clean balanced accuracy than a three-channel direct-argmax readout (94.4% versus 78.9%), whereas their hardware recoveries on ibm_kingston are 53.0% and 94.4%, respectively. A matched synthetic-noise control shows that prediction stability does not imply higher absolute noisy accuracy. Across five independently trained direct-argmax models, the clearest backend difference occurs for the moderate class and favours ibm_marrakesh ; the congested-class difference is weaker and shot-dependent, whereas free-flow recall remains stable. Finite-shot disagreement follows the inverse of the class-mean margin ordering and decreases with increasing shots, but the hardware differences do not exhibit the same monotonic behaviour. Register-level assignment-matrix indicators show directional alignment with these differences without establishing causality. Temperature scaling reduces ECE, although the residual near 0.17 is empirical rather than structural. These results demonstrate the complementary roles of readout strategy, class-resolved metrics, decision margins, and seed-level replication in hardware evaluation of VQCs.
Variational quantum classifiers (VQCs) are increasingly evaluated on real quantum processors, yet their deployment behaviour can depend on post-measurement readout and variation across independently trained models. We study a four-qubit VQC for three-class traffic-state classification on two IBM Heron r2 backends. In a single-seed comparison, a four-channel linear-projection readout achieves higher clean balanced accuracy than a three-channel direct-argmax readout (94.4% versus 78.9%), whereas their hardware recoveries on ibm_kingston are 53.0% and 94.4%, respectively. A matched synthetic-noise control shows that prediction stability does not imply higher absolute noisy accuracy. Across five independently trained direct-argmax models, the clearest backend difference occurs for the moderate class and favours ibm_marrakesh ; the congested-class difference is weaker and shot-dependent, whereas free-flow recall remains stable. Finite-shot disagreement follows the inverse of the class-mean margin ordering and decreases with increasing shots, but the hardware differences do not exhibit the same monotonic behaviour. Register-level assignment-matrix indicators show directional alignment with these differences without establishing causality. Temperature scaling reduces ECE, although the residual near 0.17 is empirical rather than structural. These results demonstrate the complementary roles of readout strategy, class-resolved metrics, decision margins, and seed-level replication in hardware evaluation of VQCs.
H. Billah and M.M.H. Sarker contributed equally to this work. Interpreting a fixed quantum circuit's output as evidence of a hardware effect requires ruling out two things usually conflated with it: finite-sample fluctuation, and structure already present in the circuit's own noiseless output. Less commonly checked is a third: whether the reference circuit compared against is the one actually executed, once compiler-level qubit routing is accounted for. We execute a fixed 35-qubit, 243-gate diagnostic circuit on IBM Heron-family hardware (three runs, two backends) and show this third check is essential: the routed circuit submitted to hardware has different noiseless predictions than the logical circuit intended, for several qubits. Against the correct, routed-circuit reference, the aggregate deviation is significant in two of three runs, and four registers (0, 9, 14, 27) show a significant output-probability deviation in every run — though we distinguish a high-confidence subset, {0, 9, 27}, with tightly time-matched calibration, from qubit 14, included only on an exploratory basis given its calibration was retrieved roughly two weeks post-execution. Same-backend replicates correlate more strongly by the Pearson measure, though a rank-based (Spearman) test shows this gap does not survive at the level of register ordering, indicating substantial leverage from a small number of registers rather than a broadly robust pattern; these three executions also share a byte-identical payload and are not independent replications in the strictest sense. The mechanism is only partially resolved: independent single-qubit readout error and raw gate count are ruled out as complete explanations for most of these registers, but the simplest version of the leading hypothesis for the single most-affected register does not reproduce the observed cross-backend pattern even with live, backend-specific timing data. A second, structurally different circuit validated only in simulation calibrates the statistical procedure's behavior under one synthetic noise model; it does not itself establish generality to real hardware, and the specific register set found here should not be expected to generalize beyond this study. Note on scope: this is a methodology and case-study contribution centered on executed-circuit references for fixed-circuit diagnostics, not a characterization of IBM Heron hardware in general or a randomized-benchmarking experiment in the formal sense. Supporting data, the routed OpenQASM circuits, raw measurement counts, and analysis code for this study are available on request from the corresponding author. Affiliation: Sonicium Quantum Lab Ltd., Dhaka, Bangladesh.
Abstract Variational Quantum Classifiers (VQCs) represent a promising approach forquantum-enhanced machine learning on Noisy Intermediate-Scale Quantum(NISQ) processors. This paper presents a comprehensive comparative studyinvestigating the impact of asymmetric readout errors on VQC performanceacross system scales. Through systematic numerical simulations of both 2-qubitand 4-qubit architectures trained on synthetic binary classification tasks, weanalyze how qubit role dependence and scalability interact with noise resilience.Our 2-qubit results demonstrate the classical optimizer’s remarkable capacityto compensate for measurement qubit noise, achieving 28.3% lower empiricalloss than theoretical predictions, while auxiliary qubit noise proves significantlymore detrimental. Extending to 4-qubit systems reveals persistent compensatorymechanisms, though with different quantitative patterns: 10% measurementqubit noise yields optimal accuracy (0.500 vs. 0.380 ideal), representing a 31.6%improvement. We observe consistent theoretical-empirical alignment across scales,with theoretical MSE values for 4-qubit systems ranging from 0.9423-0.9462for 5-20% error rates. The study demonstrates that while fundamental noiseresilience patterns persist during scaling, quantitative trade-offs and optimaloperating points shift, providing crucial insights for practical VQC deploymenton asymmetric NISQ hardware.
Variational quantum classifiers (VQCs) are often reported to perform poorly on biomedical tabular data, but such conclusions depend on the encoding circuit, the training budget, and the comparison protocol. Methods. We compared 4-qubit VQCs, a quantum-kernel SVM (QSVM), a majority-class control, and tuned classical models (on four PCA features and on all features) on Wisconsin breast cancer (n = 569), Coimbra biomarkers (n = 116), Golub leukemia gene expression (n = 72, 7,129 genes), and SEER breast-cancer registry data (n = 4,024; 15% deaths). All models were evaluated on identical blocks of 5-fold × 5-seed stratified cross-validation with preprocessing fitted on training folds, nested tuning, and Nadeau–Bengio-corrected inference with Holm correction. Trained classifiers were also evaluated on three IBM Quantum devices. Results. On the three small datasets, a VQC with the conventional two-repetition ZZ feature map was close to chance (accuracy 0.688, 0.594, 0.509), whereas a product (Z) feature map reached 0.900, 0.797, and 0.658; at matched repetition counts the Z-minus-ZZ gap was +0.12 to +0.21 and, after Holm correction, significant on Wisconsin only. More evaluations (up to 1,500) and six initializations did not close the ZZ deficit on Wisconsin. With the encoding chosen by inner cross-validation, the VQC exceeded the majority control on Wisconsin and Golub and stayed below classical models on Wisconsin and in several Golub contrasts. No statistically significant difference was detected between the encoding-selected VQC and the classical baselines on Coimbra, or between the Z-map QSVM and PCA-4 classical models; Bayesian equivalence analysis supported practical equivalence only for the QSVM on Wisconsin. On SEER the encoding mattered little: the VQC exceeded the majority control in balanced accuracy (0.548 versus 0.500) but ranked patients worse than logistic regression and random forest (ROC-AUC 0.678 versus 0.745 and 0.733). Full-feature classical models were strongest overall. On hardware (class-balanced test subsets, not natural-prevalence performance), accuracy relative to noise-free simulation changed by −0.071 to +0.048 per run (pooled ibm_fez, Wisconsin: −0.024, 95% CI −0.060 to +0.008), using 5.5 QPU-minutes. Conclusions. Under a fixed training budget, encoding choice was associated with large differences in VQC accuracy on small datasets, so single-encoding negative results should not be generalized; because the Z and ZZ maps also differ in entanglement structure, we do not attribute the gap to encoding alone. No quantum advantage was observed.
We report a set of pre-registered experiments on the three 156-qubit IBM Heron r2 processors reachable on a single open-plan account (ibm_kingston, ibm_marrakesh, ibm_fez), carried out on 14–15 October 2026. The object of study is not computational speedup but channel reliability: whether fixed symbolic messages can be delivered intact through a present-day noisy processor when the processor's per-wire readout response is measured fresh in the same sitting, and whether two such processors can exchange messages with per-position verification. Messages are 17-symbol sequences drawn from a 13-symbol test alphabet whose preparations collapse to nine distinguishable readout classes. A receiver combining per-wire selection from a fresh characterization, grammar-constrained decoding, and a corpus-frequency tie-break delivered 36/36 messages on ibm_kingston in two independent campaigns, 36/36 on ibm_marrakesh, and 35/36 on ibm_fez, against stale-plan and uncharacterized controls ranging from 2/36 to 30/36 depending on backend and control type. The grammar-constrained receiver without the corpus prior scored 30/36, 30/36, and 29/36 on the three backends; its losses are a single systematic mechanism — a deterministic tie on one battery message — identified from saved counts and closed by the prior. In a separate experiment, ibm_kingston and ibm_marrakesh exchanged self-introduction streams authored by each backend from its own fresh characterization, through a six-stage symmetric relay, with 204/204 position-flights received exactly and zero substitutions. In a counts-only blind test, an independent decoder with no shared codebook recovered the discreteness and grouping structure of two authored streams (pooled pairwise grouping agreement 0.971 against a pre-registered 0.90 strong-pass bar) while overestimating the alphabet size (reported 8, true value 6). A census of repeated characterizations shows that the ranking of a device's clearest wires turns over within hours (same-day top-17 wire overlap of 1–2 of 17 on ibm_marrakesh), while per-wire response quality moves by only ~0.01, so characterizations are within-sitting instruments. No speedup, no machine networking, and no machine understanding is claimed. The full receiver implementation is withheld as a trade secret; verification is offered through the published job IDs, retained raw counts, and tiered third-party reproduction.
We investigate whether the three normalized components of a thermodynamic stability metric, normalized fidelity I, coherence ratio ρ = T2/T1, and readout error S, can serve as training-free input features for machine learning models that predict the Φ-threshold qubit-quality class across IBM Quantum hardware. Using a strict methodology that excludes Φ from the feature set and evaluates generalization via backend-split evaluation, we find that on 445 qubits from three IBM Quantum backends, (I, ρ, S) predict the Φ-threshold class with 98.4% average balanced accuracy on held-out backends without retraining. Feature ablation reveals that ρ alone achieves 94.6% balanced accuracy and carries 70 to 78% of feature importance, while I alone and S alone are near random at 52.2% and 50.9% respectively. All four tested model architectures (Random Forest, Gradient Boosting, Neural Network, SVM) exceed 80% balanced accuracy under the strict protocol. For the narrower task of predicting T2-based hardware quality, a data-driven threshold of 0.60 outperforms the physics-derived threshold of 0.25 by 16.2 percentage points. These results establish (I, ρ, S) as a model-agnostic, training-free feature representation for predicting the Φ-threshold qubit-quality class across the tested IBM Quantum backends, with coherence ratio ρ carrying most of the predictive signal. The methods described in this paper are the subject of U.S. Provisional Patent Application No. 63/956,800 (filed January 9, 2026). No license to implement or commercialize the described methods is granted by this publication. All rights reserved.
Comparisons of quantum data-encoding strategies are rarely accompanied by a registered estimand, an explicit treatment of where simulated noise is applied, or a classical reference point. This study reports a registered simulation study of the interaction between data encoding (angle versus ZZ feature map) and circuit noise (noiseless versus single-Pauli errors with probability p = 0.02 after every CNOT, plus readout flips) for an 8-qubit variational quantum classifier trained with SPSA. Three binary datasets, ten confirmatory seeds (240 runs) and circuit size as a secondary factor were used. Hyperparameters were tuned symmetrically on separate seeds and frozen before the confirmatory runs. The registered primary estimand, the Encoding x Noise interaction on validation balanced accuracy, was -0.154 (95% seed-cluster bootstrap CI [-0.181, -0.124]), which supports the registered directional hypothesis. ZZ-encoded models trained under noise met the pre-declared convergence criterion in 1 of 60 runs, against 48 of 60 for angle encoding, so training failure and signal degradation are not separated by the registered design. An exploratory follow-up, in which ZZ models were trained without noise and evaluated under noise, showed accuracy falling to 0.55-0.63 at p = 0.02, which points to the number of noisy two-qubit gates (56 additional CNOTs) as the main driver. The registered transfer hypothesis was not supported: robust selection did not differ from naive selection (+0.003, CI [-0.012, +0.016]). A class-balanced logistic regression exceeded every quantum configuration on validation accuracy. No quantum advantage is claimed.