This project pre-registers and reports a blinded, pre-registered validation of HilbertBench, a non-intrusive diagnostic framework for quantum machine learning (QML). HilbertBench records quantum-circuit executions into append-only, cryptographically sealed traces and computes mechanism-level diagnostics — barren plateaus, shot noise, expressibility, optimization convergence, circuit structure, and hardware noise — offline, without perturbing the experiment under test. The instrument is open-source and frozen at version 1.0.0 (Zenodo DOI: 10.5281/zenodo.21009245). Purpose. We test a single question: can an analyst with no knowledge of how a QML run was produced correctly identify why it failed — barren plateau, shot starvation, noise domination, or none — using only the recorded trace and the tool's built-in analyzers? To our knowledge this is the first blinded, pre-registered validation of a QML diagnostic, modeled on the methodology of clinical trials. Design. Researcher A generates a labeled corpus of runs with planted failure modes (each label known by construction), publishes a SHA-256 cryptographic commitment of the answer key, and — only after this pre-registration is timestamped — hands the blinded runs to an independent diagnostician, Researcher B, who has not seen the corpus design. The corpus has two strata: a simulated set with clean, isolated ground truth (the primary analysis) and a real-hardware subset on a single IBM device (real-device credibility), diagnosed with primary plus optional secondary labels. Analyzer thresholds are frozen at library defaults on a disjoint development corpus before any test data is generated. Expected outcomes. We report diagnosis accuracy against the 25% four-class chance baseline with 95% Wilson confidence intervals, per-class precision/recall, and a full confusion matrix — whatever the result. We additionally pre-register a hardware prediction: on real devices, barren plateaus and noise domination are physically entangled (the circuit depth that induces an intrinsic plateau also collapses fidelity), which the confusion matrix is expected to reveal. Supporting analyses recover the McClean (2018) barren-plateau scaling exponent and validate calibration-based fidelity prediction against real hardware. All data, code, seeds, and the answer key are archived for full reproducibility.
This project pre-registers and reports a blinded, pre-registered validation of HilbertBench, a non-intrusive diagnostic framework for quantum machine learning (QML). HilbertBench records quantum-circuit executions into append-only, cryptographically sealed traces and computes mechanism-level diagnostics — barren plateaus, shot noise, expressibility, optimization convergence, circuit structure, and hardware noise — offline, without perturbing the experiment under test. The instrument is open-source and frozen at version 1.0.0 (Zenodo DOI: 10.5281/zenodo.21009245). Purpose. We test a single question: can an analyst with no knowledge of how a QML run was produced correctly identify why it failed — barren plateau, shot starvation, noise domination, or none — using only the recorded trace and the tool's built-in analyzers? To our knowledge this is the first blinded, pre-registered validation of a QML diagnostic, modeled on the methodology of clinical trials. Design. Researcher A generates a labeled corpus of runs with planted failure modes (each label known by construction), publishes a SHA-256 cryptographic commitment of the answer key, and — only after this pre-registration is timestamped — hands the blinded runs to an independent diagnostician, Researcher B, who has not seen the corpus design. The corpus has two strata: a simulated set with clean, isolated ground truth (the primary analysis) and a real-hardware subset on a single IBM device (real-device credibility), diagnosed with primary plus optional secondary labels. Analyzer thresholds are frozen at library defaults on a disjoint development corpus before any test data is generated. Expected outcomes. We report diagnosis accuracy against the 25% four-class chance baseline with 95% Wilson confidence intervals, per-class precision/recall, and a full confusion matrix — whatever the result. We additionally pre-register a hardware prediction: on real devices, barren plateaus and noise domination are physically entangled (the circuit depth that induces an intrinsic plateau also collapses fidelity), which the confusion matrix is expected to reveal. Supporting analyses recover the McClean (2018) barren-plateau scaling exponent and validate calibration-based fidelity prediction against real hardware. All data, code, seeds, and the answer key are archived for full reproducibility.
Purpose and background. Variational quantum machine-learning runs fail silently: barren plateaus, shot starvation, and hardware noise all present as a loss that stops improving. The HilbertBench instrument records circuit executions non-intrusively into sealed traces and attributes such failures offline. Our first registered study (OSF 10.17605/OSF.IO/2PWHU) validated the human-plus-instrument workflow at ceiling (72/72 blinded diagnoses) on a corpus in which each run expressed one dominant failure mechanism by construction. A ceiling result on a corpus designed to be diagnosable, however, cannot locate the instrument's limits. This study measures those limits and asks a sharper question: as a planted failure mechanism is weakened from strongly expressed to physically absent, where does detection fail, and does a human diagnostician reading the full trace evidence outperform a frozen mechanical decision rule precisely in that transition region? Design. An independent diagnostician (Researcher B from the first study, who has never seen any run in this corpus, its composition, or its answer key) diagnoses 63 blinded, simulator-generated runs from their traces alone, assigning each a primary label, an optional secondary label, a confidence, and free-text notes. The runs are produced by the byte-identical generators of the first study and span graded mechanism strengths that bracket each detection boundary, plus two-mechanism mixtures and mechanism-free controls. The class balance, per-run composition, knob levels, seeds, and generation order are withheld from the diagnostician until unsealing; the full run-level assignment is committed by SHA-256 and revealed at unsealing. Because this registration itself specifies the composition and grid, it is embargoed until the diagnosis is complete and unsealed, so that its contents cannot reach the diagnostician before she finishes her single blinded pass; the registration timestamp and the SHA-256 commitments establish that the design was fixed in advance regardless of the embargo. The comparator is a frozen mechanical rule (the naive reading of the analyzer verdicts, with a fixed mechanism precedence) whose outputs on the blinded corpus are hash-committed before the diagnostician receives the corpus, so the comparator cannot react to her results. Design rationale (simulator-only). Locating a detection boundary requires planting exactly one mechanism at an exactly known strength per run. On real NISQ hardware, single mechanisms cannot be planted cleanly -- the circuit depth that makes a landscape barren also collapses fidelity, a co-occurrence we registered and confirmed in the first study -- so hardware would introduce an uncontrolled second mechanism into every run and invalidate the answer key. This follows the standard two-arm structure of diagnostic validation: real-device (clinical) validity was established in the blinded hardware stratum of the first study; the present study is the analytical arm (limit of detection), performed on controlled references by design. The noise arm still uses calibrated device noise models with ISA transpilation, so device physics remains represented while ground truth stays reproducible. Primary outcome. A paired, one-sided exact McNemar test comparing the human's and the rule's primary labels against the planted ground truth, over the discordant runs (H1: the human is correct on more discordant runs than the rule). The paired design is used because in the transition region the knob-defined label and the trace evidence genuinely diverge; both graders face the identical ambiguity on the identical runs, and the pairing cancels it. Secondary outcomes and predictions. (H2) The human recovers the specific rule-failure classes identified previously -- borderline-SNR runs the rule grades "marginal" rather than "dominated," and shallow converged trajectories whose low variance mimics a plateau -- at a higher rate than the rule. (H3) The human's reported confidence is lower inside the transition regions than on strongly expressed anchor runs. We also report human-versus-rule detection-rate curves by planted strength, two-mechanism flag-set recall, and false positives on mechanism-free runs. Analysis and integrity. Diagnosis is a single pass. The scoring program is committed in code before any run exists; no scoring choice is made after data are collected. The diagnostician's filled sheet is hash-committed before the answer key, the corpus composition, and the rule outputs are unsealed. All commitments are auditable: repository commit b35ee2b (protocol and scorer), assignment hash 540dfefccdf1d33312a7ab3711c2486040ace46c86c2ef312fb1b1e03f35b08a, answer-key hash e7dbf167da9a7dc33219e7721a697c28b2abefa6a76a85c5bf3725417baa0139, and rule-outputs hash 6da0c442e2ca143916921f162a6a3e01f90e0fca46964bbc22dd9e5d70a12c49. Status at registration. The corpus has been generated and blinded and the comparator outputs committed; the diagnostician has not yet received the corpus, and no outcome data exist.
Abstract Claims of quantum advantage in machine learning often rest on classical comparators that lack the same features, on simulation alone, and on unreported hardware cost. QBioBench is a benchmark of quantum machine learning on peripheral biosignals that holds quantum and classical models to the same features and reports calibration and hardware cost. It comprises three clinical tasks, 12-lead ECG diagnosis (PTB-XL), single-lead apnea detection (Apnea-ECG), and wearable stress detection, each reduced to six to eight shared dimensions. A factorial grid of projected and fidelity quantum kernels and variational classifiers is screened in noisy simulation, and the promoted configurations are evaluated in an exploratory study on an IBM Nighthawk r2 processor on the full test pools of 510 apnea minutes and 512 PTB-XL records. Every prediction is recovered from the device records, intervals resample recordings or patients, and classical baselines are trained on the same units and tested on the same executed inputs. No quantum model reaches its matched support vector machine in ROC-AUC on either primary task, and none is separably better in balanced accuracy. The angle-encoded fidelity and projected kernels reach ROC-AUC 0.697 and 0.693 on apnea, paired differences of − 0.127 and − 0.119 to the support vector machine, and both are classical by construction, since the entangling layer of the encoding cancels from the fidelity kernel and leaves closed-form local expectations in the projected kernel. A variational classifier comes closest, reaching 0.797 on apnea (difference − 0.012 , 95% interval [ − 0.052 , + 0.027 ] ), and falls to chance without its two-qubit gates. Test sets of the size that circuit budgets favor are unreliable, as random 80-minute subsets of the apnea pool give the fidelity kernel a central 95% range of ROC-AUC from 0.589 to 0.797. Readout mitigation changed ROC-AUC by at most 0.015.
Quantum processors are now used as feature generators for classical models on tabular data. A circuit encodes each record, the device returns expectation values, and a classical model learns from them. A recent report claimed large gains from such features, gains that appeared only on noisy hardware. These claims are hard to check, because a quantum device is not a static function. Its state changes from job to job and from one calibration to the next, and that drift can leak information about the outcome into the features. We propose a five-step protocol for testing claims of this kind. It runs a circuit that carries no data in the same jobs as the real one. It checks whether the hardware leaves a recognizable mark on each job. It compares random and natural assignments of records to jobs, replays the null results under other submission orders, and closes the leak with a job-purged backtest. We apply it, pre-registered, to the HSBC and IBM study of fill probabilities in bond trading, which reported a large hardware-only gain. We use a public analog task and run the study's circuit on a 129-qubit chain of an IBM Heron processor. Under controlled assignment the hardware features give no gain over classical features, and the device reproduces the ideal features about as well as shot noise allows. Yet every job carries a distinct fingerprint. A circuit with no data at all predicts fills well above chance when records are batched by outcome and the same job sits on both sides of the train/test split. Removing that overlap removes the effect. The protocol costs under an hour of processor time and is designed for any hardware feature or kernel pipeline on tabular data.
Quantum Machine Learning (QML) integrates the principles of quantum mechanics superposition, entanglement, and interference into mathematical and logical operations that classical computers cannot emulate. Despite the limitations of today’s noisy intermediate-scale quantum (NISQ-ERA) devices, these platforms already sustain small- to medium-scale machine-learning workloads, catalyzing rapid expansion in QML research. Guided by PRISMA methodology, we systematically screened 5,463 publications published between January 2015 and June 2025 and identified 160 high-quality studies that implemented or evaluated QML algorithms on real or simulated quantum processors. The selected literature converges into four thematic categories: (i) quantum adaptations of classical algorithms, (ii) variational and hybrid quantum–classical models, (iii) quantum linear-algebra accelerators, and (iv) quantum-inspired generative and reinforcement techniques. Application domains span image and signal classification, cybersecurity, finance, chemistry, and healthcare, with benchmarks such as MNIST, CIC-IDS2019 and credit-risk corpora featuring prominently in the reviewed corpus. Separately from the reviewed literature, Sect. 6 reports our own five-seed benchmark of five quantum and two classical classifiers on five public tabular datasets (Breast Cancer, Diabetes, Iris, Sonar and Bank Marketing). Comparative analyses reveal that hybrid QML systems typically match or marginally exceed the performance of state-of-the-art classical baselines while employing fewer parameters. Nevertheless, a clear, reproducible quantum advantage remains elusive, hindered by data-encoding overheads, circuit-depth constraints, environmental noise, and barren-plateau training phenomena. By integrating algorithmic trends, empirical outcomes, and hardware considerations, this review delineates QML’s present capabilities and identifies critical technological milestones error mitigation, scalable qubit architectures, and efficient data-loading schemes needed to unlock its full potential.
Abstract—Quantum Machine Learning (QML) operating on Noisy Intermediate-Scale Quantum (NISQ) hardware represents an emerging computational paradigm to address high-dimensional optimization problems that challenge classical Von Neumann supercomputing architectures. By mapping complex feature vectors into exponentially large 2^N Hilbert state spaces via Parameterized Quantum Circuits (PQCs) and Variational Quantum Algorithms (VQAs), QML targets computational advantages in combinatorial optimization, high-dimensional classification, and quantum-enhanced molecular modeling. However, the theoretical promise of NISQ-era variational computing is severely constrained by three fundamental bottlenecks: (1) Barren Plateaus, characterized by the exponential concentration of measure and vanishing variance of cost function gradients; (2) Hardware Decoherece Noise and readout error rates that corrupt quantum state fidelity; and (3) Classical-to-Quantum State Preparation bottlenecks that can erase theoretical computational speedups. This paper delivers an exhaustive, mathematically formal technical investigation and empirical benchmark of QML architectures for high-dimensional optimization. We formalize the Haar-distributed geometric mechanics of barren plateaus under 2-design unitary assumptions, evaluate barren plateau mitigation through shallow local cost operators and layer-by-layer parameter initialization, and contrast Variational Quantum Eigensolvers (VQE) with Quantum Kernel Support Vector Machines (Q-SVM). Furthermore, we construct an empirical benchmarking testbed executing across both state-vector simulators and superconducting quantum backends (IBM Quantum Heron / Eagle architectures), evaluating 127-qubit Hamiltonian optimization and high-dimensional financial portfolio allocation. Our findings demonstrate that while global cost functions suffer exponential gradient decay Var[grad] ~ O(2^-N) across circuit depth d >= 4, local Hamiltonian decomposition maintains trainable variance, achieving a 98.2% ground-state fidelity under zero-noise extrapolation (ZNE) error mitigation. Finally, we formulate open challenges in quantum error mitigation scaling, non-perturbative barren plateau escapes, and quantum data loading efficiency. Index Terms—Quantum Machine Learning (QML), Parameterized Quantum Circuits (PQCs), Variational Quantum Eigensolver (VQE), Barren Plateaus, Quantum Kernel Methods, NISQ Computing, Error Mitigation, Hilbert Space.
This preregistration describes a confirmatory study of an upstream conditioning architecture in a quantum error-correction workflow. The project’s purpose is to test whether an intervention applied prior to the primary decoder produces a real and reproducible decoder-relevant effect, remains compatible with more than one downstream decoder family, generates a measurable exposed-observable statistical footprint, and improves or preserves performance across a preregistered regime rather than only at a single operating point. The study is intentionally limited to architecture-level evaluation, with the public claim set confined to exposed observables, frozen transformations, preregistered statistical tests, explicit negative controls, and bounded interpretation rules. The confirmatory design is organized around five preregistered tests: category existence, decoder compatibility, structural footprint, observable statistical signature, and regime robustness. Across these tests, the central comparison is between a baseline control pipeline and a conditioned pipeline in which the upstream intervention is applied while all non-conditioning settings are held fixed. The study uses frozen observables and frozen execution rules, including matched seeds, shot counts, code distances, physical error rates, geometry, and baseline decoder settings, so that any supported effect can be interpreted as evidence about the architecture under test rather than as a byproduct of post hoc tuning. To preserve a clean-room boundary, the registration restricts primary analysis to exposed, decoder-facing quantities only. These include logical failure, an exposed defect-count summary, and an exposed decoder-difficulty proxy, along with a locked scalar error-load construction derived from control-only standardization. The study does not rely on decoder-internal state variables, hidden confidence traces, or implementation-specific internal diagnostics in its primary confirmatory analyses. This structure is intended to let readers evaluate whether the architecture leaves a reproducible footprint in decoder-facing behavior without disclosing reconstruction-level mechanism details. Expected outcomes are defined in confirmatory rather than speculative terms. If the preregistered hypotheses are supported, the study is expected to provide evidence that an upstream conditioning intervention exists as a real architectural category in QEC, acts before the primary decoder, is compatible with downstream decoding, changes decoder-relevant exposed structure, and produces a reproducible observable statistical footprint in logical-failure behavior. If support is absent, mixed, or null, those outcomes will still be recorded under the frozen rules without rescue through post hoc metric changes, reweighting, or reinterpretation. In either case, the expected result of the registration is a bounded, auditable public record of whether this architecture class shows reproducible category-level behavior under a preregistered confirmatory design. The preregistration is not intended to establish a unique mechanism identity, identify a privileged implementation, disclose implementation-specific reconstruction details, or support claims about exact targeting logic, suppression laws, subset-sizing logic, telemetry-governed control, or broader cross-domain theory narratives. Public-facing interpretation is therefore intentionally limited to category validation and exposed-observable behavior. This is meant to support rigorous public evaluation while preserving strict clean-room limits around protected implementation detail.