Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2610.29186

Readout Dependent Parameter Identifiability of Quantum Kernels

Hyong Keun Park, Jahoon Jeong, Joonkyo Shim, Hyunsoo Yoon

quant-phSubmitted Oct 21, 2026• Announced Oct 22, 2026• 0 citations
arXivHTMLPDF

Abstract

A circuit parameter can change measured features without changing their projected Gaussian kernel. We characterize this lost dependence and distinguish recovery by metric design from nonlinear feature expansion. With analytic feature coefficients, a fixed finite data basis affinely independent on the input domain, a fixed positive definite metric, and fixed bandwidth, the coefficient Gram matrix determines all-input kernel equivalence, global coordinate inactivity, and generic differential rank. An exact four-qubit family exhibits a sharp boundary: diagonal metrics give rank two, dense metrics reach but cannot exceed three, and no fixed affine expansion crosses this ceiling. One appended nonlinear scalar coordinate suffices for rank four and is minimal when the original features are retained. A single pairwise product achieves this with an explicit nondiagonal metric; a scalar sum of two products achieves it with a diagonal metric. Both use existing ideal local means. Exact certificates establish ranks in 54 selected circuit/readout settings. Matched finite-shot comparisons show that minimal representation size need not give the best sensitivity among the tested designs. Under a factorized measurement distribution, conditional product estimation cannot increase variance at a fixed design. Execution audits further separate kernel identifiability from compiled cost.

AI Summary

A trainable circuit parameter can change measured features while leaving their projected Gaussian kernel unchanged. This paper studies which parameter combinations remain identifiable after fixing the readout, metric and bandwidth, and whether changing the metric or expanding the features can recover lost dependence. For features $f_\theta(x)=C(\theta)h(x)+b(\theta)$, with analytic coefficients and a fixed finite basis whose image affinely spans its coefficient space, the coefficient Gram matrix $G=C^TMC$ characterizes equality of kernels on all input pairs. Its Jacobian determines global coordinate inactivity and generic local identifiable dimension. The proof uses injectivity of the map from coefficient Grams to pairwise distances. Exact coefficient tests give a finite decision procedure for a rational-frequency algebraic subclass, without an efficiency guarantee. An explicit four-qubit example separates metric and representation recovery: diagonal metrics give rank two, dense positive-definite metrics attain at most three, and no fixed affine expansion exceeds three. One appended nonlinear scalar feature suffices for rank four and is minimal when retaining the original features. These products use existing ideal local means, not additional measurement information. Exact symbolic identities, interval witnesses and rational minors establish ranks in 54 selected circuit/readout settings; rank need not increase when observables are added. Matched finite-shot simulations distinguish restored rank from statistical sensitivity: the smallest sufficient representation is not consistently best conditioned. Rank is local and generic, not global injectivity or predictive improvement. Parameter–data product angles lie outside the fixed-basis theorem, and strengthened compilation baselines leave no additional angle-task savings in the tested scope.

Some signs of AI-generated textIncrementalStrong assumptions

Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.

Comments
53 pages, 3 figures. Includes appendices. Code and reproducibility materials: https://github.com/parkfinity/readout_kernel_identifiability

Related papers

Metric Learning with Quantum Feature Spaces

Sharma, Vinayak, Sen, Lord, Padhy, Ashish +4 more

The $2^N$ dimensions of Hilbert space give quantum metric learning exponentially many mutually orthogonal states with which to separate classes, but the same exponential dimension also degrades its optimization landscape and the tractability of evaluating its metric. By letting a classical network predict all gate parameters of a fixed ansatz, we isolate the reachable state manifold and its fidelity metric. We prove that this separability is already present on the unentangled manifold $\mathbb{M}_{\text{unentangled}} = (\mathbb{CP}^1)^N$, which saturates the $2^N$ maximal orthogonal capacity needed for classification, while entanglement adds states only at intermediate separations, which are costly to resolve under finite sampling. The same size carries costs. Global fidelity causes exponential gradient collapse across both unentangled ($Θ(c^{-N})$) and entangled ($Θ(4^{-N})$) manifolds; as fidelities concentrate, the fidelity signal-to-noise ratio at a fixed shot budget also decays exponentially in $N$; and because $\mathbb{M}_{\text{unentangled}}$ already holds $2^N$ orthogonal states, expressive, unconstrained encoders under contrastive losses reach trivial minima by \emph{geometric shattering} even without entanglement. We introduce Quantum Projective Metric Learning (QPMeL), an exact $O(N)$ closed-form kernel on $\mathbb{M}_{\text{unentangled}}$ that circumvents circuit simulation. QPMeL resolves gradient collapse via a curriculum annealing from a $1$-local observable regime (restoring $Θ(1/N^2)$ polynomial variance) to the global product kernel. Empirically, QPMeL as a complete design exceeds eight and matches two of 11 published quantum few-shot designs.

arXiv:2312.01655•Dec 4, 2023•1 citation

Statevector-to-Hardware Reconstruction of a Four-Qubit ZZ Quantum Kernel: A Single-Backend Case Study of Three Execution Jobs

Rostyslav Sipakov

Hardware noise and finite sampling perturb the fidelity estimates forming a quantum-kernel Gram matrix. We measured how far three hardware-reconstructed Gram matrices depart from an exact statevector reference for one frozen four-qubit ZZ feature map on N=24 indoor air-quality windows, executed on ibm_fez at 1024 shots per circuit in three single, non-interleaved jobs: baseline, dynamical decoupling alone, and gate twirling alone. All were complete, finite, and positive-semidefinite. Off-diagonal root-mean-squared error (RMSE) against the reference was 0.0878, 0.0864, and 0.0427; full-matrix centered kernel alignment (CKA) ranged 0.933-0.989 and the post hoc diagonal-excluded (U-centered) CKA 0.816-0.986. The gate-twirled job deviated least on every reported geometry axis; its baseline contrasts are deletion-stable for the Spearman, mean-absolute-error, RMSE, and full-matrix CKA diagnostics, while the Pearson and diagonal-excluded contrasts fall just below that convention. Dynamical decoupling was not separated from the baseline. The observed error exceeded both finite-shot reference scales, so, under those sampling-only models, sampling does not explain it. Centered kernel-target alignment did not track reconstruction fidelity and stayed at or below each label-permutation reference: implementation fidelity and task relevance are distinct diagnostic axes. All configuration-level statements describe three realized jobs on one backend; no mitigation-efficacy, classifier-superiority, forecasting, or quantum-advantage claim is made.

quant-pharXiv:2607.20377•Quantum Reports•Jul 23, 2026•1 citation

Adaptive Low-Rank Quantum Kernel Acquisition under Finite-Shot Measurements: Perturbation Guarantees for QSVMs

Duong Do, Duong Tung Nguyen

Quantum kernel methods require estimating entries of a dense Gram matrix through repeated circuit measurements, making kernel construction expensive and statistically noisy. For a dataset of size $n$, dense construction requires $O(n^2)$ kernel queries, while existing approximation methods typically assume inexpensive access to kernel entries or use non-adaptive sampling. We study adaptive low-rank quantum kernel acquisition when the kernel is accessible only through a finite-shot quantum overlap oracle. Our framework, ICAQKL, uses residual-driven incomplete Cholesky construction to acquire informative kernel columns sequentially and form a positive-semidefinite rank-$r$ surrogate, reducing the training-kernel acquisition cost to $O(nrN_s)$ circuit executions for $r\ll n$, where $N_s$ is the number of shots per queried entry. The central difficulty is that finite-shot errors perturb both the queried kernel values and the adaptive pivot sequence that determines the low-rank factor. We develop a non-asymptotic perturbation framework for this stochastic adaptive process, establishing concentration over the realized queries, stability of noisy pivot selection and factor construction, and propagation of the coupled sampling and approximation errors to the QSVM objective, optimizer, prediction function, and classification margin. The resulting rank--shot conditions characterize when ICAQKL approaches the full QSVM induced by the same quantum feature map. Experiments on benchmark and biomedical datasets show favorable accuracy--query-cost tradeoffs, competitive performance with the full QSVM, and competitive or improved performance relative to the evaluated Nyström baselines under comparable settings. ICAQKL is intended to enable measurement-efficient and theoretically controlled use of a selected quantum kernel, rather than to establish quantum advantage over classical kernels.

arXiv:2611.03777•Nov 2, 2026•0 citations

Empirical Evaluation of Quantum Kernel Methods for Classification: A Comparative Study with Classical Kernels

Ronak Bhalgami

Quantum kernel methods are the most concrete near-term proposal for quantum machine learning on classical data, yet published benchmarks disagree on whether they outperform tuned classical kernels, and the disagreement cannot be resolved from the literature because the comparisons are not protocol-controlled. We argue that the obstacle is one of measurement, not theory: the usual setup fixes the quantum feature map's data scaling while grid-searching the classical kernel's hyperparameters, recomputes nothing per fold, and reports cost in simulator time. We therefore treat the comparison itself as the object of study and impose three controls. The bandwidth is tuned on the same grid budget as the classical $γ$, the kernel is recomputed inside every fold from fold-fitted preprocessing, and cost is counted in hardware circuit evaluations. On top of these we run a factorial design of 324 cells in which dimensionality, training size, class separability and data structure are explicit covariates, together with four standard datasets, a shots-versus-gate-noise ablation, and a cost sweep. The dominant effect we measure is not quantum: tuning the feature map's bandwidth is worth 0.120 accuracy on synthetic data and 0.339 on real data, exceeding every quantum-classical difference we observe. Tuned quantum kernels are then statistically indistinguishable from a tuned RBF on all four real datasets. The advantage on synthetic data is confined almost entirely to data generated by the feature map itself, and it shrinks with dimensionality (-0.0045 per qubit) while growing with sample size. Finite sampling costs an order of magnitude more accuracy than gate and readout noise, and a fidelity kernel needs $5\times10^{7}$ circuits at n = 10,000. Protocol variance is thus sufficient to explain the field's contradictory conclusions.

quant-pharXiv:2610.42029•Nov 2, 2026•0 citations

Bandwidth, Representation Loss, and Temporal Drift in Quantum Kernel Fraud Detection: A Pre-Specified Falsification Study

Mohamed Faisal Sindhi, Hammaz Buksh

Abstract Quantum kernel results are often reported without holding the input representation, training sample, tuning budget and temporal evaluation protocol fixed against a matched classical kernel. This pre-specified, protocol-locked falsification study uses the Elliptic++ transaction tables (203,769 transactions, 49 time steps) under a locked temporal split (training steps 1–24, validation 25–29, test 30–49 in 4 locked windows). The primary metric is pooled Recall@5%. A frozen full-dimensional gradient-boosting model reached 0.381, and every model, including this one, collapsed in the final window. Graph structure and selective reranking were tested first and falsified. Compressing 182 features to 4, 6 and 8 principal components retained 34%, 43% and 51% of the variance and 64%, 50% and 33% of the full-dimensional model’s pooled Recall@5%. In the primary, default-bandwidth analysis, an entangling fidelity quantum kernel was below the strongest matched classical kernel at 4, 6 and 8 qubits (pooled Recall@5% differences −0.078, −0.203 and −0.146). In a secondary, post hoc bandwidth correction, explicit scale tuning raised the entangling kernel’s Recall@5% by 0.070, 0.121 and 0.109; the corrected kernel remained below the strongest matched classical kernel at 6 and 8 qubits and was statistically unresolved at 4 qubits. The selected scale lay at the lower boundary of the tested grid, and the apparent effect of entanglement depended on bandwidth. No repeatable predictive benefit of the quantum kernel over matched classical kernels was observed in the tested setting; the findings do not extend beyond this dataset, protocol, representation, feature-map family, sample size and bandwidth range.

•Oct 13, 2026•0 citations

Readout-Rank Laws for Isotropic Quantum Tangents

Marwan Ait Haddou

Deep parameterized quantum circuits may remain sensitive to a parameter change while the observables retained by a learning model barely respond. We study this separation for a fixed computational-basis measurement. For a pure-state tangent, we compare the quantum Fisher information $F_Q$, the Fisher information $F_{\rm full}$ in the complete bitstring distribution, and the largest variance-normalized response $\mathcal I_{\mathcal A}$ available to a diagonal readout space $\mathcal A$. If the joint state--tangent frame is Haar random, we prove that the two successive information fractions are independent Beta variables whose means are $1/2$ and $r/(2^n-1)$, where $r$ is the centered dimension of the readout. Consequently, even the joint span of all computational-basis Pauli strings through any fixed weight $k$ retain only $O(n^k2^{-n})$ of the full-record information. Exact-statevector experiments across six circuit families show increasing finite-size agreement with this hierarchy in five nonconserving ensembles as the circuit depth grows. A number-conserving family departs strongly from the isotropic prediction even after correcting the support and readout rank, showing that rank alone is insufficient without tangent isotropy.

quant-pharXiv:2608.07628•Aug 10, 2026•2 citations
A circuit parameter can change measured features without changing their projected Gaussian kernel. We characterize this lost dependence and distinguish recovery by metric design from nonlinear feature expansion. With analytic feature coefficients, a fixed finite data basis affinely independent on the input domain, a fixed positive definite metric, and fixed bandwidth, the coefficient Gram matrix determines all-input kernel equivalence, global coordinate inactivity, and generic differential rank. An exact four-qubit family exhibits a sharp boundary: diagonal metrics give rank two, dense metrics reach but cannot exceed three, and no fixed affine expansion crosses this ceiling. One appended nonlinear scalar coordinate suffices for rank four and is minimal when the original features are retained. A single pairwise product achieves this with an explicit nondiagonal metric; a scalar sum of two products achieves it with a diagonal metric. Both use existing ideal local means. Exact certificates establish ranks in 54 selected circuit/readout settings. Matched finite-shot comparisons show that minimal representation size need not give the best sensitivity among the tested designs. Under a factorized measurement distribution, conditional product estimation cannot increase variance at a fixed design. Execution audits further separate kernel identifiability from compiled cost.
A trainable circuit parameter can change measured features while leaving their projected Gaussian kernel unchanged. This paper studies which parameter combinations remain identifiable after fixing the readout, metric and bandwidth, and whether changing the metric or expanding the features can recover lost dependence.
For features fθ(x)=C(θ)h(x)+b(θ)f_\theta(x)=C(\theta)h(x)+b(\theta)fθ​(x)=C(θ)h(x)+b(θ), with analytic coefficients and a fixed finite basis whose image affinely spans its coefficient space, the coefficient Gram matrix G=CTMCG=C^TMCG=CTMC characterizes equality of kernels on all input pairs. Its Jacobian determines global coordinate inactivity and generic local identifiable dimension. The proof uses injectivity of the map from coefficient Grams to pairwise distances. Exact coefficient tests give a finite decision procedure for a rational-frequency algebraic subclass, without an efficiency guarantee.
An explicit four-qubit example separates metric and representation recovery: diagonal metrics give rank two, dense positive-definite metrics attain at most three, and no fixed affine expansion exceeds three. One appended nonlinear scalar feature suffices for rank four and is minimal when retaining the original features. These products use existing ideal local means, not additional measurement information. Exact symbolic identities, interval witnesses and rational minors establish ranks in 54 selected circuit/readout settings; rank need not increase when observables are added.
Matched finite-shot simulations distinguish restored rank from statistical sensitivity: the smallest sufficient representation is not consistently best conditioned. Rank is local and generic, not global injectivity or predictive improvement. Parameter–data product angles lie outside the fixed-basis theorem, and strengthened compilation baselines leave no additional angle-task savings in the tested scope.
The 2N2^N2N dimensions of Hilbert space give quantum metric learning exponentially many mutually orthogonal states with which to separate classes, but the same exponential dimension also degrades its optimization landscape and the tractability of evaluating its metric. By letting a classical network predict all gate parameters of a fixed ansatz, we isolate the reachable state manifold and its fidelity metric. We prove that this separability is already present on the unentangled manifold Munentangled=(CP1)N\mathbb{M}_{\text{unentangled}} = (\mathbb{CP}^1)^NMunentangled​=(CP1)N, which saturates the 2N2^N2N maximal orthogonal capacity needed for classification, while entanglement adds states only at intermediate separations, which are costly to resolve under finite sampling. The same size carries costs. Global fidelity causes exponential gradient collapse across both unentangled (Θ(c−N)Θ(c^{-N})Θ(c−N)) and entangled (Θ(4−N)Θ(4^{-N})Θ(4−N)) manifolds; as fidelities concentrate, the fidelity signal-to-noise ratio at a fixed shot budget also decays exponentially in NNN; and because Munentangled\mathbb{M}_{\text{unentangled}}Munentangled​ already holds 2N2^N2N orthogonal states, expressive, unconstrained encoders under contrastive losses reach trivial minima by \emph{geometric shattering} even without entanglement. We introduce Quantum Projective Metric Learning (QPMeL), an exact O(N)O(N)O(N) closed-form kernel on Munentangled\mathbb{M}_{\text{unentangled}}Munentangled​ that circumvents circuit simulation. QPMeL resolves gradient collapse via a curriculum annealing from a 111-local observable regime (restoring Θ(1/N2)Θ(1/N^2)Θ(1/N2) polynomial variance) to the global product kernel. Empirically, QPMeL as a complete design exceeds eight and matches two of 11 published quantum few-shot designs.
Hardware noise and finite sampling perturb the fidelity estimates forming a quantum-kernel Gram matrix. We measured how far three hardware-reconstructed Gram matrices depart from an exact statevector reference for one frozen four-qubit ZZ feature map on N=24 indoor air-quality windows, executed on ibm_fez at 1024 shots per circuit in three single, non-interleaved jobs: baseline, dynamical decoupling alone, and gate twirling alone. All were complete, finite, and positive-semidefinite. Off-diagonal root-mean-squared error (RMSE) against the reference was 0.0878, 0.0864, and 0.0427; full-matrix centered kernel alignment (CKA) ranged 0.933-0.989 and the post hoc diagonal-excluded (U-centered) CKA 0.816-0.986. The gate-twirled job deviated least on every reported geometry axis; its baseline contrasts are deletion-stable for the Spearman, mean-absolute-error, RMSE, and full-matrix CKA diagnostics, while the Pearson and diagonal-excluded contrasts fall just below that convention. Dynamical decoupling was not separated from the baseline. The observed error exceeded both finite-shot reference scales, so, under those sampling-only models, sampling does not explain it. Centered kernel-target alignment did not track reconstruction fidelity and stayed at or below each label-permutation reference: implementation fidelity and task relevance are distinct diagnostic axes. All configuration-level statements describe three realized jobs on one backend; no mitigation-efficacy, classifier-superiority, forecasting, or quantum-advantage claim is made.
Quantum kernel methods require estimating entries of a dense Gram matrix through repeated circuit measurements, making kernel construction expensive and statistically noisy. For a dataset of size nnn, dense construction requires O(n2)O(n^2)O(n2) kernel queries, while existing approximation methods typically assume inexpensive access to kernel entries or use non-adaptive sampling. We study adaptive low-rank quantum kernel acquisition when the kernel is accessible only through a finite-shot quantum overlap oracle. Our framework, ICAQKL, uses residual-driven incomplete Cholesky construction to acquire informative kernel columns sequentially and form a positive-semidefinite rank-rrr surrogate, reducing the training-kernel acquisition cost to O(nrNs)O(nrN_s)O(nrNs​) circuit executions for r≪nr\ll nr≪n, where NsN_sNs​ is the number of shots per queried entry. The central difficulty is that finite-shot errors perturb both the queried kernel values and the adaptive pivot sequence that determines the low-rank factor. We develop a non-asymptotic perturbation framework for this stochastic adaptive process, establishing concentration over the realized queries, stability of noisy pivot selection and factor construction, and propagation of the coupled sampling and approximation errors to the QSVM objective, optimizer, prediction function, and classification margin. The resulting rank--shot conditions characterize when ICAQKL approaches the full QSVM induced by the same quantum feature map. Experiments on benchmark and biomedical datasets show favorable accuracy--query-cost tradeoffs, competitive performance with the full QSVM, and competitive or improved performance relative to the evaluated Nyström baselines under comparable settings. ICAQKL is intended to enable measurement-efficient and theoretically controlled use of a selected quantum kernel, rather than to establish quantum advantage over classical kernels.
Quantum kernel methods are the most concrete near-term proposal for quantum machine learning on classical data, yet published benchmarks disagree on whether they outperform tuned classical kernels, and the disagreement cannot be resolved from the literature because the comparisons are not protocol-controlled. We argue that the obstacle is one of measurement, not theory: the usual setup fixes the quantum feature map's data scaling while grid-searching the classical kernel's hyperparameters, recomputes nothing per fold, and reports cost in simulator time. We therefore treat the comparison itself as the object of study and impose three controls. The bandwidth is tuned on the same grid budget as the classical γγγ, the kernel is recomputed inside every fold from fold-fitted preprocessing, and cost is counted in hardware circuit evaluations. On top of these we run a factorial design of 324 cells in which dimensionality, training size, class separability and data structure are explicit covariates, together with four standard datasets, a shots-versus-gate-noise ablation, and a cost sweep. The dominant effect we measure is not quantum: tuning the feature map's bandwidth is worth 0.120 accuracy on synthetic data and 0.339 on real data, exceeding every quantum-classical difference we observe. Tuned quantum kernels are then statistically indistinguishable from a tuned RBF on all four real datasets. The advantage on synthetic data is confined almost entirely to data generated by the feature map itself, and it shrinks with dimensionality (-0.0045 per qubit) while growing with sample size. Finite sampling costs an order of magnitude more accuracy than gate and readout noise, and a fidelity kernel needs 5×1075\times10^{7}5×107 circuits at n = 10,000. Protocol variance is thus sufficient to explain the field's contradictory conclusions.
Abstract Quantum kernel results are often reported without holding the input representation, training sample, tuning budget and temporal evaluation protocol fixed against a matched classical kernel. This pre-specified, protocol-locked falsification study uses the Elliptic++ transaction tables (203,769 transactions, 49 time steps) under a locked temporal split (training steps 1–24, validation 25–29, test 30–49 in 4 locked windows). The primary metric is pooled Recall@5%. A frozen full-dimensional gradient-boosting model reached 0.381, and every model, including this one, collapsed in the final window. Graph structure and selective reranking were tested first and falsified. Compressing 182 features to 4, 6 and 8 principal components retained 34%, 43% and 51% of the variance and 64%, 50% and 33% of the full-dimensional model’s pooled Recall@5%. In the primary, default-bandwidth analysis, an entangling fidelity quantum kernel was below the strongest matched classical kernel at 4, 6 and 8 qubits (pooled Recall@5% differences −0.078, −0.203 and −0.146). In a secondary, post hoc bandwidth correction, explicit scale tuning raised the entangling kernel’s Recall@5% by 0.070, 0.121 and 0.109; the corrected kernel remained below the strongest matched classical kernel at 6 and 8 qubits and was statistically unresolved at 4 qubits. The selected scale lay at the lower boundary of the tested grid, and the apparent effect of entanglement depended on bandwidth. No repeatable predictive benefit of the quantum kernel over matched classical kernels was observed in the tested setting; the findings do not extend beyond this dataset, protocol, representation, feature-map family, sample size and bandwidth range.
Deep parameterized quantum circuits may remain sensitive to a parameter change while the observables retained by a learning model barely respond. We study this separation for a fixed computational-basis measurement. For a pure-state tangent, we compare the quantum Fisher information FQF_QFQ​, the Fisher information FfullF_{\rm full}Ffull​ in the complete bitstring distribution, and the largest variance-normalized response IA\mathcal I_{\mathcal A}IA​ available to a diagonal readout space A\mathcal AA. If the joint state--tangent frame is Haar random, we prove that the two successive information fractions are independent Beta variables whose means are 1/21/21/2 and r/(2n−1)r/(2^n-1)r/(2n−1), where rrr is the centered dimension of the readout. Consequently, even the joint span of all computational-basis Pauli strings through any fixed weight kkk retain only O(nk2−n)O(n^k2^{-n})O(nk2−n) of the full-record information. Exact-statevector experiments across six circuit families show increasing finite-size agreement with this hierarchy in five nonconserving ensembles as the circuit depth grows. A number-conserving family departs strongly from the isotropic prediction even after correcting the support and readout rank, showing that rank alone is insufficient without tangent isotropy.