Invariance Audits for Quantum Kernels and Variational Rewinding: A Real-to-Hermitian Taxonomy of Projector, Flag, Anchor, and Density Geometry

Azadeh Alavi; Fatemeh Kouchmeshki; Hossein Akhoundi

REVIEW 2 major objections 6 minor 1 cited by

Quantum fidelity kernels and QVR return scores are exactly projector and anchor overlaps; they help only when the invariances they impose match the task.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 15:14 UTC pith:6OQNSYM2

load-bearing objection Solid hygiene paper: exact projector/anchor dictionary for fidelity kernels and QVR returns, useful flag obstruction, but mostly standard algebra plus toy witnesses. the 2 major comments →

arxiv 2607.07927 v1 pith:6OQNSYM2 submitted 2026-07-08 quant-ph

Invariance Audits for Quantum Kernels and Variational Rewinding: A Real-to-Hermitian Taxonomy of Projector, Flag, Anchor, and Density Geometry

Azadeh Alavi , Fatemeh Kouchmeshki , Hossein Akhoundi This is my paper

classification quant-ph MSC 53C3062H3068T0781P6815A18

keywords quantum machine learningquantum kernelsvariational rewindingHermitian projectorsGrassmann and flag manifoldsinvariance auditdensity geometryanchor overlap

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Machine-learning pipelines often replace raw vectors by unit directions, projectors, subspaces, ordered flags, quantum states, or density operators before any classifier is trained. This paper treats every such replacement as an invariance decision: a choice of which distinctions to keep and which to quotient out. It builds a self-contained real-to-Hermitian taxonomy that places Grassmann and flag kernels, quantum fidelity kernels, and QVR-style return scores on one geometric axis, and proves the exact identities that make a noiseless fidelity kernel equal the Hilbert–Schmidt product of rank-one Hermitian projectors and a QVR return probability equal an overlap with a learned anchor. Controlled vector, subspace, statevector, anomaly, finite-shot, and synthetic quotient-witness experiments show that these lifts succeed when their invariances match the label structure and fail correctly when the discarded information carries the label. A sympathetic reader cares because the taxonomy supplies a precise dictionary for auditing when quantum or geometric features are scientifically justified rather than treating them as opaque upgrades.

Core claim

The paper shows that a noiseless quantum fidelity kernel is exactly the Hilbert–Schmidt inner product of the associated rank-one Hermitian projectors, K(x,y)=|⟨ϕ(x)|ϕ(y)⟩|²=tr(Px Py), and that a QVR-style return probability is exactly an anchor-overlap score pθ(x)=tr(Px Aθ). Rank-r returns are complex Grassmann anchors; mixed or multimodal classes are density or positive-semidefinite anchors. On the real side, weighted flag kernels are positive semidefinite and block-gauge invariant, while whole-span Grassmann geometry cannot separate same-span block swaps that ordered flags can. Quantum and geometric lifts are therefore useful precisely when their imposed invariances match the task, and fai

What carries the argument

The exact trace identities K(x,y)=tr(Px Py) for fidelity kernels and pθ(x)=tr(Px Aθ) for QVR-style returns. These place both scores inside Hilbert–Schmidt geometry, separate pairwise kernel SVMs (many-anchor Hermitian margin models) from single-anchor return models, and let rank-r and density anchors be read as complex Grassmann and mixed-state extensions.

Load-bearing premise

That controlled statevector simulations on toy datasets and a synthetic time-series generator are enough to treat the taxonomy as general representation guidance, even though trainable entangling circuits, hardware noise, and real public anomaly benchmarks are left out of scope.

What would settle it

Build a same-span block-swap task where ordered flag features fail while whole-span Grassmann features succeed, or a native anomaly setting under the paper’s own encoder and anchor construction where the exact statevector return probability and the Hermitian trace score tr(Px A) systematically disagree beyond roundoff.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

A practitioner can audit a quantum or geometric pipeline by checking whether the label depends on radius, global phase, basis choice, or block order that the chosen lift erases.
When a class is multimodal or occupies an extended region of Hilbert space, rank-one pure anchors can be replaced by rank-r Grassmann or density anchors without leaving the same geometric family.
Same-span block-swap problems require ordered flag features; whole-span Grassmann geometry is guaranteed to fail on them.
Finite-shot binomial sampling can be layered on the same trace score without changing the underlying identity; only the estimate, not the geometry, changes.
A quantum-kernel SVM is a many-anchor Hermitian margin model and need not match single-anchor QVR performance even when both use the same encoder.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same invariance-audit checklist could be applied to other hybrid pipelines (variational classifiers with different observables, re-uploading designs) to decide when entanglement is necessary versus when a classical projector kernel already captures the geometry.
A matched public real time-series anomaly study—identical encoder, return measurement, shot model, noise model, and training budget for circuit and trace implementations—would turn the paper’s synthetic anomaly audit into a transferable comparison protocol.
If the taxonomy is adopted, QML reporting could routinely include an explicit invariance statement (what is quotiented out) alongside accuracy tables.
The block-swap witness suggests a broader family of quotient-witness tests for any manifold embedding used in classical or quantum machine learning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Solid hygiene paper: exact projector/anchor dictionary for fidelity kernels and QVR returns, useful flag obstruction, but mostly standard algebra plus toy witnesses.

read the letter

The punchline is that this is a careful taxonomy paper, not a quantum-advantage paper. It rewrites noiseless fidelity kernels and QVR-style returns as Hermitian projector/anchor overlaps, adds a clean real-side flag-versus-Grassmann obstruction, and runs controlled witnesses that match those invariances. That framing is useful; the load-bearing equations are elementary and correct.

What is actually new is the packaging: a self-contained real-to-Hermitian dictionary that separates unit directions, real projectors, Grassmann, ordered flags, pure-state projectors, rank-r Grassmann anchors, density centers, and Hermitian decision operators for kernel SVMs. On the real side, the weighted flag kernel PSD/block-gauge proof and the same-span block-swap witness are clean and do real work—whole-span Grassmann must fail when block order carries the label, and the tables show exactly that. On the quantum side, Props. 7–9 are immediate from Born’s rule and cyclicity (K = tr(Px Py), pθ = tr(Px Aθ)), with Ky Fan for the best rank-r anchor. The paper is honest about scope: no hardware claims, no trainable-entangler claims, experiments as invariance audits.

Soft spots are real but already flagged by the authors. The algebraic core is largely standard matrix analysis and QI (Nielsen–Chuang, Schuld’s kernel view, Grassmann literature). Novelty is unification and diagnostic language more than new theorems. Evidence is scikit-learn toys, a synthetic anomaly series, and statevector/finite-shot layers; code is proprietary. That is fine for a hygiene paper and insufficient for operational guidance—exactly as §11 says. Tables are modest and consistent with the non-claims; product-angle kernels being strong is expected once you have the closed form.

Who it is for: people who design or audit quantum/geometric lifts and want a shared language for when projectivization, subspace, flag, or density quotients are appropriate. A serious referee should see it. I would engage with the dictionary and the flag witness; I would not treat the toy tables as general validation. Send it to peer review.

Referee Report

2 major / 6 minor

Summary. The manuscript develops a real-to-Hermitian taxonomy for auditing representation choices in quantum machine learning. It formalizes Grassmann and weighted flag projector kernels (positive semidefiniteness and block-gauge invariance), gives a same-span block-swap obstruction showing when ordered flags succeed while whole-span Grassmann fails, and proves that a noiseless fidelity kernel equals the Hilbert–Schmidt product of rank-one Hermitian projectors, K(x,y)=|⟨ϕ(x)|ϕ(y)⟩|²=tr(PxPy), while a QVR-style return probability equals an anchor overlap pθ(x)=tr(PxAθ). Rank-r returns are identified with complex Grassmann anchors and mixed models with density anchors. Controlled toy, synthetic witness, statevector, finite-shot, and anomaly experiments are presented as invariance diagnostics rather than advantage benchmarks; the paper explicitly disclaims hardware speedup and quantum advantage.

Significance. If accepted as stated, the paper supplies a clean, self-contained dictionary that unifies real projective/Grassmann/flag geometry with Hermitian projector, anchor, and density scores used in quantum kernels and QVR-style returns. The central trace identities (Props. 7–9) are elementary consequences of Born’s rule and cyclicity, but packaging them with real-side guardrails (Props. 2–4), the same-span block-swap witness, and an explicit separation of pairwise kernel SVMs from single-anchor returns is useful for the field. Strengths include fully written proofs, conservative limitations (§11), explicit non-claims about hardware advantage, and experiments framed as falsifiable invariance witnesses (Tables 4, 7, 9) rather than operational benchmarks. The contribution is taxonomic and diagnostic rather than a new algorithmic or complexity result; that is appropriate for the stated scope.

major comments (2)

Sections 6–7 and Code availability: the numerical tables (especially the same-span block-swap witness in Table 4, the native anomaly audit in Table 7, and the synthetic quotient witnesses in Table 9) are load-bearing for the abstract’s claim that controlled experiments “support the same conclusion.” Executable code is proprietary and only aggregate outputs are offered upon request. For those tables to function as independent invariance witnesses, the manuscript needs enough non-proprietary detail—exact synthetic time-series generator equations, seed/split protocol, and pseudocode for the rank-selection and threshold rules—to allow third-party reproduction of the reported ROC–AUC/F1/mismatch counts and the block-swap accuracies. Without that, the empirical half remains only partially auditable even though the algebraic identities stand alone.
§5.5 and Table 11 vs. Tables 6 and 10: the conceptual distinction between a pairwise quantum-kernel SVM (affine Hermitian decision operator B in the span of training projectors) and a single-anchor QVR return is central and correctly stated algebraically. Table 10 already warns that its “context” columns are not the same constructed score as the trace-consistency probe. The abstract and §7.5 still risk being read as a performance ranking of “QVR vs kernels vs density.” A short, explicit statement that no fair head-to-head optimization budget is claimed—and that Table 6 is a classwise protocol summary, not a refutation of QVR for its original one-class setting—should be placed next to Table 6 so the diagnostic framing cannot be misread as a model bake-off.

minor comments (6)

Proposition 6 and Table 5: the product-angle kernel’s closed form is correctly used to caution against reading strong toy accuracy as quantum advantage; consider stating the classical evaluation cost O(N²q) already in the table caption for readers who skip §9.
Table 2 caption asserts that vector RBF is the best overall article-level model on all four datasets; the main text should briefly note whether that ranking used the same nested validation protocol as the non-vector rows, to avoid any appearance of post-hoc selection (§6.3).
§4.3: the fixed (untrained) entangling reuploading baseline is weaker in Table 5; the text already warns against over-interpreting this, but a single sentence cross-referencing Limitation 4 in the table discussion would help.
Notation: both Uθ and Vθ appear for the learned unitary in §5; standardize to one symbol.
References [13–14] introduce QVR; a one-sentence clarification that the present anomaly audit is not a reproduction of those benchmarks (already in §1.1 and §11) could also appear in the Table 7 caption for skimmers.
MSC codes and keywords are appropriate; “invariance audit” is used as a methodological phrase—define it once in §1 when first introduced.

Circularity Check

1 steps flagged

No load-bearing circular derivation; core quantum identities are elementary definitional rewrites packaged as a taxonomy, not fitted predictions or self-citation chains.

specific steps

renaming known result [Abstract; §1.2 Contributions 3–4; Prop. 7 (Eq. 24); Props. 8–9 (Eqs. 30–36)]
"We prove that a noiseless quantum fidelity kernel is exactly a Hermitian projector kernel... We prove that a QVR-style return probability is exactly an anchor-overlap score... |⟨ϕ(x)|ϕ(y)⟩|² = tr(Px Py)... |⟨0|Vθ|ϕ(x)⟩|² = tr(Px Aθ)."

These equalities are the standard Hilbert–Schmidt form of pure-state fidelity and Born’s rule after defining Aθ := V* Π V. Presenting them as numbered proofs and as the quantum half of a new taxonomy renames textbook projector geometry rather than deriving a non-definitional prediction. The reduction is mild and explicit; it does not force empirical claims or hide a fit.

full rationale

The paper’s central algebraic claims (Prop. 7: noiseless fidelity equals tr(Px Py); Props. 8–9: QVR-style return equals tr(Px Aθ), with rank-r and density extensions) are standard Hilbert–Schmidt / Born-rule identities. Once Px and Aθ are defined as the projectors the paper writes down, the equalities hold by cyclicity of the trace and the definition of the success projector; the short proofs simply expand those definitions. That is near-definitional bookkeeping, not a circular prediction of an independent quantity from a fitted parameter. The real-side guardrails (weighted flag PSD and block-gauge invariance, same-span obstruction, spectral-activation identity) are likewise elementary and self-contained. There is no fitted-input-called-prediction loop, no uniqueness theorem imported from the authors’ prior work, and no load-bearing self-citation chain: QVR is cited to Baker et al. and PennyLane (external). The controlled block-swap, phase/radius, and anomaly witnesses are diagnostic checks of the stated invariances, not statistically forced forecasts. The only mild circularity-adjacent feature is packaging textbook projector identities as numbered “proofs” and a “real-to-Hermitian taxonomy” contribution; that is renaming/organization of known geometry rather than a forced derivation. Score 1 reflects that mild packaging, not a broken derivation chain. Honest finding: no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central algebraic claims rest almost entirely on standard linear algebra and quantum measurement axioms; experimental conclusions additionally assume that toy/synthetic controlled tasks are informative about representation choice. No new physical constants are fitted. Free parameters are experimental knobs (flag weights, shot counts, simple noise rates, validation-selected ranks) that affect tables but not the trace identities. Invented entities are methodological (audit taxonomy, witnesses), not new particles or forces.

free parameters (5)

flag block weights α_b ≥ 0
Chosen nonnegative weights in the weighted flag kernel KF; they define the kernel family but are not derived from data in the proofs.
anchor rank r (validation-selected)
Rank of success projectors / Grassmann anchors selected on validation ROC–AUC with PR–AUC and smaller-rank tie-breakers in the anomaly and classwise audits.
finite-shot counts S ∈ {128,512,1024,4096}
Hand-chosen observation budgets for binomial sampling layers; affect mismatch counts, not the exact trace identity.
simple noise model (depolarizing 0.03, readout flip 0.02)
Ad hoc transparent observation noise rates used only in one anomaly row; not fitted to hardware.
product-angle scale α and encoder/preprocess choices
Circuit and preprocessing variants selected under nested validation; free modeling choices for numerical tables.

axioms (5)

domain assumption Born rule: measurement success probability equals ⟨ψ|Π|ψ⟩ for a projector Π (and unitary conjugates).
Invoked throughout §4–5 to equate return probabilities with tr(Px Aθ).
standard math Hilbert–Schmidt inner product and cyclicity of trace on Hermitian operators.
Used in Props. 2, 7–9 to obtain kernel PSD and fidelity/return identities.
standard math Ky Fan maximum principle for tr(ρA) over rank-r projectors.
Cited as Prop. 10 to justify optimal unrestricted rank-r anchors.
domain assumption Ideal noiseless statevector model is the right object for the algebraic audit; shots/noise are observation layers only.
Stated in scope, §6, and Limitations; underpins all exact-identity claims.
ad hoc to paper Controlled toy/synthetic tasks with known label-bearing invariances are valid witnesses for representation audits.
Experimental protocol §6–7 treats scikit-learn toys and synthetic block-swap/phase/radius/anomaly generators as diagnostic rather than operational benchmarks.

invented entities (2)

Real-to-Hermitian invariance-audit taxonomy (projector/flag/anchor/density dictionary) no independent evidence
purpose: Classify representation objects by quotient and score type for QML and geometric ML.
Organizational construct of the paper; not a physical entity. Independent evidence is limited to the paper’s own controlled witnesses.
Same-span block-swap witness task no independent evidence
purpose: Show when whole-span Grassmann features must fail while ordered flags succeed.
Synthetic diagnostic constructed for the paper’s Prop. 3; useful but paper-internal.

pith-pipeline@v1.1.0-grok45 · 19556 in / 3805 out tokens · 50914 ms · 2026-07-10T15:14:10.399644+00:00 · methodology

0 comments

read the original abstract

Machine-learning models often replace vectors by normalized directions, projectors, covariances, subspaces, ordered flags, quantum states, or density operators before any classifier is fitted. This replacement is an invariance decision: it determines which distinctions are kept and which are quotiented out. We develop a self-contained real-to-Hermitian taxonomy for auditing such representations in quantum machine learning. On the real side, we formalize Grassmann and flag projector kernels, prove positive semidefiniteness and block-gauge invariance of a weighted flag kernel, and give a same-span block-swap witness showing when whole-span Grassmann geometry must fail while ordered flags succeed. On the quantum side, we prove that a noiseless fidelity kernel is exactly the Hilbert--Schmidt inner product between the associated rank-one Hermitian projectors, and that a QVR-style return probability is exactly an overlap score between the input projector and a learned anchor operator. Rank-constrained returns correspond to complex Grassmann anchors, while mixed or multimodal class models are naturally represented by density or positive-semidefinite anchors. Controlled vector, subspace, statevector, anomaly, finite-shot, and quotient-witness experiments support the same conclusion: quantum and geometric lifts are useful when their invariances match the task, and fail correctly when discarded information is label-bearing. The paper makes no hardware-speedup or quantum-advantage claim.

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Gram-Certified Resource Continuation for Structured Quantum Representation Audits
quant-ph 2026-07 conditional novelty 5.5

A coarse spectral flag that is δ_c-suboptimal transfers to a fine isometrically lifted problem with suboptimality at most δ_c+2ε, certified by a 2m amplitude Gram matrix, and continuation is justified only when mismat...

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · cited by 1 Pith paper · 2 internal anchors

[1]

Bhatia.Positive Definite Matrices

R. Bhatia.Positive Definite Matrices. Princeton University Press, 2007

work page 2007
[2]

R. A. Horn and C. R. Johnson.Matrix Analysis, 2nd edition. Cambridge University Press, 2013

work page 2013
[3]

M. A. Nielsen and I. L. Chuang.Quantum Computation and Quantum Information, 10th anniversary edition. Cambridge University Press, 2010

work page 2010
[4]

Edelman, T

A. Edelman, T. A. Arias, and S. T. Smith. The geometry of algorithms with orthogonality constraints.SIAM Journal on Matrix Analysis and Applications, 20(2):303–353, 1998

work page 1998
[5]

Absil, R

P.-A. Absil, R. Mahony, and R. Sepulchre.Optimization Algorithms on Matrix Manifolds. Princeton University Press, 2008

work page 2008
[6]

Hamm and D

J. Hamm and D. D. Lee. Grassmann discriminant analysis: A unifying view on subspace-based learning. InProceedings of the 25th International Conference on Machine Learning, pages 376–383, 2008

work page 2008
[7]

Harandi, M

M. Harandi, M. Salzmann, and R. Hartley. From manifold to manifold: Geometry-aware dimensionality reduction for SPD matrices. InEuropean Conference on Computer Vision, pages 17–32, 2014

work page 2014
[8]

Schuld, I

M. Schuld, I. Sinayskiy, and F. Petruccione. An introduction to quantum machine learning. Contemporary Physics, 56(2):172–185, 2015

work page 2015
[9]

Havlicek, A

V. Havlicek, A. D. Corcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, and J. M. Gambetta. Supervised learning with quantum-enhanced feature spaces.Nature, 567:209–212, 2019

work page 2019
[10]

M. Schuld. Supervised quantum machine learning models are kernel methods.arXiv preprint arXiv:2101.11020, 2021. 19

work page internal anchor Pith review Pith/arXiv arXiv 2021
[11]

Schuld, R

M. Schuld, R. Sweke, and J. J. Meyer. Effect of data encoding on the expressive power of variational quantum-machine-learning models.Physical Review A, 103:032430, 2021

work page 2021
[12]

Perez-Salinas, A

A. Perez-Salinas, A. Cervera-Lierta, E. Gil-Fuster, and J. I. Latorre. Data re-uploading for a universal quantum classifier.Quantum, 4:226, 2020

work page 2020
[13]

J. S. Baker, H. Horowitz, S. K. Radha, S. Fernandes, C. Jones, N. Noorani, V. Skavysh, P. Lamontagne, and B. C. Sanders. Quantum variational rewinding for time series anomaly detection.arXiv preprint arXiv:2210.16438, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[14]

Quantum detection of time series anomalies

PennyLane Demos. Quantum detection of time series anomalies. https://pennylane.ai/dem os/tutorial_univariate_qvr, accessed July 2026

work page 2026
[15]

FidelityQuantumKernel documentation

Qiskit Machine Learning. FidelityQuantumKernel documentation. https://qiskit-communi ty.github.io/qiskit-machine-learning/stubs/qiskit_machine_learning.kernels.Fi delityQuantumKernel.html, accessed July 2026

work page 2026
[16]

Data encoding

IBM Quantum Learning. Data encoding. https://quantum.cloud.ibm.com/learning/cou rses/quantum-machine-learning/data-encoding, accessed July 2026

work page 2026
[17]

Toy datasets

scikit-learn developers. Toy datasets. https://scikit-learn.org/stable/datasets/toy_d ataset.html, accessed July 2026

work page 2026
[18]

Common pitfalls and recommended practices

scikit-learn developers. Common pitfalls and recommended practices. https://scikit-learn .org/stable/common_pitfalls.html, accessed July 2026. 20

work page 2026

[1] [1]

Bhatia.Positive Definite Matrices

R. Bhatia.Positive Definite Matrices. Princeton University Press, 2007

work page 2007

[2] [2]

R. A. Horn and C. R. Johnson.Matrix Analysis, 2nd edition. Cambridge University Press, 2013

work page 2013

[3] [3]

M. A. Nielsen and I. L. Chuang.Quantum Computation and Quantum Information, 10th anniversary edition. Cambridge University Press, 2010

work page 2010

[4] [4]

Edelman, T

A. Edelman, T. A. Arias, and S. T. Smith. The geometry of algorithms with orthogonality constraints.SIAM Journal on Matrix Analysis and Applications, 20(2):303–353, 1998

work page 1998

[5] [5]

Absil, R

P.-A. Absil, R. Mahony, and R. Sepulchre.Optimization Algorithms on Matrix Manifolds. Princeton University Press, 2008

work page 2008

[6] [6]

Hamm and D

J. Hamm and D. D. Lee. Grassmann discriminant analysis: A unifying view on subspace-based learning. InProceedings of the 25th International Conference on Machine Learning, pages 376–383, 2008

work page 2008

[7] [7]

Harandi, M

M. Harandi, M. Salzmann, and R. Hartley. From manifold to manifold: Geometry-aware dimensionality reduction for SPD matrices. InEuropean Conference on Computer Vision, pages 17–32, 2014

work page 2014

[8] [8]

Schuld, I

M. Schuld, I. Sinayskiy, and F. Petruccione. An introduction to quantum machine learning. Contemporary Physics, 56(2):172–185, 2015

work page 2015

[9] [9]

Havlicek, A

V. Havlicek, A. D. Corcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, and J. M. Gambetta. Supervised learning with quantum-enhanced feature spaces.Nature, 567:209–212, 2019

work page 2019

[10] [10]

M. Schuld. Supervised quantum machine learning models are kernel methods.arXiv preprint arXiv:2101.11020, 2021. 19

work page internal anchor Pith review Pith/arXiv arXiv 2021

[11] [11]

Schuld, R

M. Schuld, R. Sweke, and J. J. Meyer. Effect of data encoding on the expressive power of variational quantum-machine-learning models.Physical Review A, 103:032430, 2021

work page 2021

[12] [12]

Perez-Salinas, A

A. Perez-Salinas, A. Cervera-Lierta, E. Gil-Fuster, and J. I. Latorre. Data re-uploading for a universal quantum classifier.Quantum, 4:226, 2020

work page 2020

[13] [13]

J. S. Baker, H. Horowitz, S. K. Radha, S. Fernandes, C. Jones, N. Noorani, V. Skavysh, P. Lamontagne, and B. C. Sanders. Quantum variational rewinding for time series anomaly detection.arXiv preprint arXiv:2210.16438, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022

[14] [14]

Quantum detection of time series anomalies

PennyLane Demos. Quantum detection of time series anomalies. https://pennylane.ai/dem os/tutorial_univariate_qvr, accessed July 2026

work page 2026

[15] [15]

FidelityQuantumKernel documentation

Qiskit Machine Learning. FidelityQuantumKernel documentation. https://qiskit-communi ty.github.io/qiskit-machine-learning/stubs/qiskit_machine_learning.kernels.Fi delityQuantumKernel.html, accessed July 2026

work page 2026

[16] [16]

Data encoding

IBM Quantum Learning. Data encoding. https://quantum.cloud.ibm.com/learning/cou rses/quantum-machine-learning/data-encoding, accessed July 2026

work page 2026

[17] [17]

Toy datasets

scikit-learn developers. Toy datasets. https://scikit-learn.org/stable/datasets/toy_d ataset.html, accessed July 2026

work page 2026

[18] [18]

Common pitfalls and recommended practices

scikit-learn developers. Common pitfalls and recommended practices. https://scikit-learn .org/stable/common_pitfalls.html, accessed July 2026. 20

work page 2026