{"id":"02b4f315-db02-47ea-b535-39d917555890","arxiv_id":"2506.03779","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper proposes a roadmap for quantum operator-valued kernels and shows on simulated quantum channel estimation that they can outperform scalar-valued quantum kernels.","lead":"This position paper argues that quantum kernel machines should move from scalar-valued to operator-valued kernels to handle structured outputs and exploit entanglement. It defines entangled quantum operator-valued kernels, gives a quantum circuit for them, and reports an initial proof-of-concept on quantum channel estimation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Definition 3.1 does not establish that K is a positive semi-definite operator-valued kernel; since §5.2 admits the experimental σ^{i,j}=σ_iσ_j is non-Hermitian, the block Gram matrix may be indefinite and the proof-of-concept kernel regression may be ill-posed.","rationale":"Reading in good faith, this is a clearly hedged position paper: the authors state that QOVKs are not a 'magic bullet' and openly discuss open questions. The roadmap is coherent and the proof-of-concept is suggestive. The most load-bearing issue, however, is not the missing classical OVK baseline highlighted in the reader's weakest_assumption, but the fact that Definition 3.1 never proves the proposed object is a positive semi-definite operator-valued kernel in the sense of Definition A.1. If Eq. (1) is not psd, then calling it a kernel is unjustified, and the kernel ridge regression experiment in Section 5.1 is not a valid kernel method. The paper itself flags the relevant caveat in §5.2: the general construction does not require σ^{x,z}_X to be Hermitian, and the experiments use non-Hermitian products σ_iσ_j. This makes the validity gap concrete rather than hypothetical. The missing classical OVK baseline is a genuine interpretive concern about whether the reported gain is quantum-specific, but it only affects the strength of the empirical support; the psd gap affects whether the central mathematical object exists as claimed. Both are fixable: one can restrict σ^{x,z} to a form that guarantees psd (for example, prove the product form is psd via a Schur-product or Gram-matrix argument) and add a classical OVK baseline. Because the reader's conditional verdict already requires a psd-ness argument, my critique does not change the verdict; it re-centers the primary concern. I therefore recommend no change to the CONDITIONAL outcome, and agreement with the reader is partial because the emphasized weakness differs while the prescribed conditions overlap.","tokens_in":20941,"tokens_out":14188,"duration_ms":130110,"concrete_test":"Use a single input with σ^{x,x}=-I in Definition 3.1 and check whether K(x,x) is positive semidefinite; if it is not, the definition does not define a kernel. Then build the block Gram matrix for the Section 5.1 setting (U=CNOT∘SWAP, σ^{i,j}=σ_iσ_j, 100 random density matrices, 10 seeds) and compute its minimum eigenvalue; any negative eigenvalue would show the experimental QOVK is not psd, so the reported QSVK-vs-QOVK comparison is not a valid kernel-method comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Definition A.1 requires K(x,z)=K(z,x)* and block-wise positive semidefiniteness. Definition 3.1 never verifies these; it only imposes that U is non-separable. §5.2 explicitly says the general definition does not require σ^{x,z}_X to be Hermitian and that the channel-estimation experiments use σ^{i,j}_X=σ_iσ_j, products of density matrices that are generally non-Hermitian. For n=1, psd requires K(x,x)≥0. If σ^{x,x}=-I, then K(x,x)=-Tr_X[U(ρ_Y⊗I)U†]≤0, so Eq. (1) is not psd without additional assumptions on the feature matrices. Even for the special product form used in Table 1, no psd proof is supplied; if the assembled np×np Gram matrix has negative eigenvalues, the kernel ridge regression solution is not a valid kernel estimate. Thus the central construction of the paper—the entangled QOVK—is not yet a well-defined kernel, which makes the proof-of-concept support questionable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This position paper argues that research on quantum kernel machines should move from scalar-valued quantum kernels (QSVKs) to operator-valued quantum kernels (QOVKs), which can encode non-factorizable input-output interactions and exploit entanglement. The paper defines an entangled QOVK in Definition 3.1 via a partial trace over a dilated input-output system with a non-separable unitary U, sketches a quantum circuit implementation in Appendix D, and reports a proof-of-concept on quantum channel estimation in Section 5.1 showing that a QOVK recovers bit-flip and dephasing channels with much lower Frobenius error than a QSVK (Table 1). The paper also proposes a roadmap with four actions: quantum implementations of OVKs, entangled QOVKs, C*-algebraic kernels, and quantum structured prediction.","tokens_in":21180,"tokens_out":7802,"duration_ms":85786,"significance":"If the central construction can be made rigorous, this is a timely and potentially influential proposal: it reframes the quantum kernel debate around structured prediction and operator-valued outputs, where classical kernel methods are less dominant. The paper has several strengths: Definition 3.1 is explicit, the circuit derivation in Appendix D is concrete, the proof-of-concept is a genuine simulation without fitted constants, and the authors acknowledge that entanglement alone does not guarantee advantage. However, the current manuscript does not establish that the proposed object is a valid kernel in the operator-valued sense, and the empirical support does not isolate a quantum or entanglement-specific effect. These issues are load-bearing because the proof-of-concept is the main quantitative evidence for the position.","major_comments":[{"comment":"The paper does not establish that the operator-valued function in Eq. (1) is a positive semi-definite operator-valued kernel. Appendix A (Definition A.1 and Theorem A.3) requires K(x,z)=K(z,x)* and block-wise positive semi-definiteness for a reproducing kernel to exist, but Definition 3.1 imposes only non-separability of U and makes no assumptions on the feature matrices sigma^{x,z}_X. Section 5.2 explicitly states that sigma^{x,z}_X is not required to be Hermitian and that the experiments use sigma^{i,j}_X = sigma_i sigma_j, which are generally non-Hermitian. For n=1, if sigma^{x,x}_X = -I, then K(x,x) = -Tr_X[U(rho_Y tensor I)U^dag] is negative semi-definite, so the proposed class is not psd without additional assumptions. Even for the specific product form used in Table 1, no psd proof is supplied; if the assembled np x np Gram matrix has negative eigenvalues, the kernel ridge regression solution reported in Section 5.1 is not a valid kernel estimate. The authors should either restrict Definition 3.1 to feature matrices satisfying a block-psd condition or prove psd for the constructions used in the experiments and the circuit.","section":"Definition 3.1, Eq. (1), Appendix A, Section 5.2"},{"comment":"The experiment compares the QOVK only against a scalar-valued quantum kernel (QSVK), so the observed improvement does not identify the source of the gain. The QOVK differs from the QSVK in two ways: it uses operator-valued outputs and it uses a non-separable unitary U. A classical operator-valued kernel with a suitable output kernel could plausibly achieve comparable or better recovery on this matrix-valued regression task, which would weaken the paper's central claim that the advantage comes from quantum operator-valued structure or entanglement. To support the position, the authors should add a classical OVK baseline (for example, k(sigma_i, sigma_j) L with a tuned positive definite output kernel L) and, ideally, a separable quantum OVK with the same input feature matrix, so that the effect of operator-valued outputs can be separated from the effect of the non-separable U.","section":"Section 5.1, Table 1, Appendix B"},{"comment":"There is an inconsistency between the kernel used in the experiments and the kernel for which a quantum circuit is provided. Appendix D derives a specific feature matrix sigma^{x,z}_X = (rho_x + rho_z + <psi_x|psi_z> |psi_x><psi_z| + <psi_z|psi_x> |psi_z><psi_x|) / (2(1+| <psi_z|psi_x> |^2)) in Eq. (8), which is Hermitian and positive semi-definite. Section 5.1, however, states that the experiments use sigma^{i,j}_X = sigma_i sigma_j, products of density matrices that are generally non-Hermitian. These are different constructions: the circuit prepares the former, while the experiments evaluate the latter. The paper should clarify which feature matrix is used in each place, and either provide a circuit for the experimental construction or validate the circuit construction experimentally.","section":"Section 5.1 vs Appendix D, Eqs. (8) and (10)"},{"comment":"The proof-of-concept fixes the interaction unitary U to the composition of a CNOT and a SWAP gate and does not report any sensitivity analysis. Since the reported advantage may depend on this ad hoc choice, the authors should vary U (for example, random two-qubit unitaries or different entangling gates) and report the resulting error distribution. Without this, the experiment demonstrates only that one particular entangled construction works on one task, not that the QOVK framework generally provides the claimed structural advantage.","section":"Section 5.1, Appendix B"}],"minor_comments":[{"comment":"The notation paragraph describes the inner product and tensor product in bra-ket notation, but the final sentence is slightly confusing because it lists the tensor product immediately after the definition of the inner product; consider separating these two definitions more clearly.","section":"Section 2.2, Notation paragraph"},{"comment":"Section 5.1 refers to 'Eq. (10)' for the QOVK definition, but Eq. (10) appears only in Appendix D, not in the main text. The authors should renumber or add a forward reference to the appendix.","section":"Section 5.1, Eq. (10) reference"},{"comment":"In the definition of the dephasing channel, diag(sigma) should be explicitly identified as the matrix obtained by setting all off-diagonal entries of sigma to zero; the current wording is acceptable but could be more precise.","section":"Appendix B.1, dephasing channel definition"},{"comment":"The sentence 'The input feature matrix sigma^{x,z}_X produced by the quantum circuit is a density matrix ... In contrast, the general definition ... does not require sigma^{x,z}_X to be Hermitian' highlights a tension that should be resolved in the main definition, not only in the circuit discussion.","section":"Section 5.2"},{"comment":"The paper promises that the code will be publicly available at a GitLab URL; adding a link to the actual repository and a short description of the software dependencies would improve reproducibility for the proof-of-concept.","section":"Global"}],"recommendation":"major_revision","confidential_remarks":"The position is appropriate for an ML venue, and the authors are honest about the preliminary nature of the evidence. The main technical blocker is the unproven positive semi-definiteness of Definition 3.1; this is fixable by adding conditions on sigma or by proving psd for the specific constructions used. The experimental comparison also needs a classical OVK baseline to support the quantum-specific interpretation. I recommend major revision rather than rejection, because the core proposal is defensible and the required changes are local in scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a genuine position paper with a new construction—entangled quantum operator-valued kernels—and a concrete swap-test circuit. But the construction isn't shown to be a valid kernel, and the experiments don't isolate the quantum part. The psd gap is the one to worry about.\n\nWhat's new: definition (1) is a real extension of classical entangled kernels to the quantum setting, and the circuit in Fig. 6/Appendix D is worked out in detail. The channel-estimation simulation is a real proof-of-concept, not a fit-to-data exercise, and the paper is honest about its preliminary nature.\n\nThe soft spots, in order of severity. First, Definition 3.1 calls K an operator-valued kernel without verifying the psd condition. The paper explicitly says in Sec 5.2 that σ^{x,z}_X need not be Hermitian and that the experiments use products of density matrices, which are generally non-Hermitian. The stress-test concern is correct: for arbitrary feature matrices, K can be indefinite (take σ^{x,x} = -I). Kernel ridge regression with an indefinite Gram matrix is not a well-posed kernel method. This is a load-bearing omission, not a cosmetic one.\n\nSecond, the comparison is QOVK vs QSVK only. A classical operator-valued kernel baseline would show whether the gain comes from operator-valuedness or from quantum entanglement. The paper interprets the gain as support for the quantum construction, but that interpretation isn't tested. Third, the interaction unitary U is fixed by hand (CNOT+SWAP); the result may be sensitive to that choice.\n\nThese are fixable. The roadmap and circuit are worth keeping. The authors need to either prove psd for a useful subclass of feature matrices or explicitly state the open problem and soften the claims. And they should add a classical OVK baseline before asserting a quantum advantage.\n\nThis paper is for researchers in quantum kernel methods and structured prediction. It's a position paper, so the value is in the direction and the initial evidence, not in a theorem. I'd send it to peer review—the idea is timely and the circuit is a concrete starting point—but I'd make acceptance conditional on addressing the psd question and the missing baseline.","headline":"The paper's QOVK construction is new and the circuit is concrete, but the missing psd proof and the lack of a classical OVK baseline leave the quantum advantage claim unsupported.","tokens_in":21725,"tokens_out":3437,"would_cite":false,"duration_ms":32724,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Quantum kernel machines should move beyond scalar-valued kernels to operator-valued kernels, whose entangled form can exploit non-factorizable input-output structure.","keywords":["quantum kernel methods","operator-valued kernels","entanglement","structured prediction","quantum channel estimation","kernel ridge regression","C*-algebraic kernels","quantum machine learning"],"falsifier":"Run the same bit-flip and dephasing channel estimation task with a classical operator-valued kernel (for example a separable OVK with an output kernel learned by cross-validation) under identical training data, noise, and evaluation protocol; if it achieves errors near the QOVK's 0.101 and 0.096, the case for a quantum-specific entangled advantage collapses. A second check is to replace the hand-fixed CNOT+SWAP unitary $U$ with a randomly chosen non-separable unitary: if recovery errors jump back to QSVK levels, the result depends on the specific $U$ rather than the entangled class.","tokens_in":20744,"feed_emoji":"⚛️","tokens_out":8867,"duration_ms":87057,"temperature":0.7,"pith_summary":"Quantum kernel research has stalled, in this paper's telling, because it has been testing scalar-valued kernels on standard classification and regression problems where classical kernel methods already excel. The paper argues that the field should move to operator-valued kernels, which return matrices rather than numbers and can represent structured outputs together with their internal dependencies. Its central proposal is the entangled quantum operator-valued kernel (QOVK), built from a non-separable unitary that couples input and output registers, allowing output correlations to depend on inputs in a joint, non-factorizable way. A proof-of-concept on quantum channel estimation shows the QOVK recovering bit-flip and dephasing channels far more accurately than a scalar-valued quantum kernel (Frobenius errors 0.101 and 0.096 versus 0.569 and 0.525). If this view is right, the meaningful testbed for quantum advantage in kernel machines is structured prediction, not scalar classification.","feed_headline":"Operator-valued quantum kernels beat scalar kernels on channel recovery","feed_subtitle":"Proof-of-concept on quantum channel estimation cuts error from 0.569 to 0.101, pointing toward structured prediction.","key_machinery":"The central object is the entangled quantum operator-valued kernel of Definition 3.1, $K(x,z)=\\operatorname{Tr}_X[U_{YX}(\\rho_Y\\otimes\\sigma^X_{x,z})U_{YX}^\\dagger]$, in which $U_{YX}$ is a unitary that cannot be written as $A_Y\\otimes B_X$, $\\rho_Y$ is a density matrix on the output space, and $\\sigma^X_{x,z}$ is a feature matrix extracted from inputs $x$ and $z$. The non-separable unitary carries the argument: it couples input and output subsystems so that correlations among outputs can depend on the input representation, a property separable kernels cannot express. Around this object the paper builds two supporting mechanisms: a swap-test quantum circuit that produces the feature matrix and evaluates the kernel on a quantum device, and a C*-algebraic generalization via reproducing kernel Hilbert C*-modules, in which kernel outputs can themselves be quantum operators such as gates and density matrices.","core_discovery":"The paper claims that scalar-valued quantum kernels lack the degrees of freedom needed to exploit intrinsically quantum resources, and that realizing the potential of quantum kernel machines requires operator-valued kernels. Its definition of an entangled QOVK is $K(x,z)=\\operatorname{Tr}_X[U_{YX}(\\rho_Y\\otimes\\sigma^X_{x,z})U_{YX}^\\dagger]$, where $U_{YX}$ is a non-separable unitary, $\\rho_Y$ is an output density matrix, and $\\sigma^X_{x,z}$ is an input feature matrix; the non-separability is what lets the kernel represent input-dependent output correlations rather than factorizing them. The paper shows that quantum scalar-valued kernels, including fidelity kernels, are recovered as special separable cases of this construction. It also gives a swap-test quantum circuit that prepares the feature matrix and evaluates the kernel, and reports a proof-of-concept where kernel ridge regression with an entangled QOVK estimates single-qubit bit-flip and dephasing channels with much lower Frobenius error than a QSVK (0.101 and 0.096 versus 0.569 and 0.525). The intended contribution is a road map: develop entangled QOVKs, bring in C*-algebraic reproducing kernel Hilbert C*-modules, and concentrate applications on structured prediction.","pith_inferences":["Beyond the paper: a classical operator-valued kernel with a good output kernel may reproduce the channel-estimation gains, which would show the advantage comes from operator-valued outputs rather than quantum entanglement; the paper does not test this control.","Beyond the paper: because the unitary $U$ is fixed by hand as CNOT+SWAP, the result may depend on matching $U$ to the channel structure; learning $U$ from data or averaging over random non-separable unitaries would clarify whether the entangled class, not a single gate, is responsible.","Beyond the paper: the operator-valued setting could connect quantum kernel methods to quantum neural tangent kernels and variational feature learning, since learned unitaries on a coupled input-output system would interpolate between kernel machines and trained quantum models; the authors only gesture at learned feature maps.","Beyond the paper: if the QOVK advantage persists under a classical OVK control, the framework suggests a concrete scaling test, increase output dimension $p$ and channel complexity and ask whether the gap between entangled QOVK and classical OVK grows, which would indicate where quantum resources start to matter."],"forward_implications":["Quantum kernel machines gain a natural application regime, structured prediction, where outputs carry dependencies that scalar kernels cannot encode, so quantum advantages can be assessed where classical methods are known to struggle.","Entanglement becomes a concrete design resource in kernels: non-separable unitaries generate kernel classes strictly larger than fidelity and separable kernels, so experiments can test whether entanglement improves learning.","The operator-valued kernel matrix grows from $n\\times n$ to $np\\times np$, giving quantum linear-system solvers and amplitude-amplification routines a larger matrix dimension on which their polylogarithmic scaling could, in principle, translate into a practical edge.","The C*-algebraic RKHM route lets kernel outputs be quantum objects such as density operators and gates, which would make tasks like quantum channel learning native to the kernel framework.","Quantum channel estimation is proposed as a first benchmark where the QOVK clearly separates from the QSVK, giving the community a concrete task for further study."],"supporting_citations":[{"why":"Established quantum feature spaces and quantum kernels, the object class the paper argues is too restrictive.","marker":"Havlíček et al., 2019"},{"why":"Defined quantum machine learning in feature Hilbert spaces and the fidelity kernel used as the scalar-valued baseline.","marker":"Schuld & Killoran, 2019"},{"why":"Supplies negative generalization results for scalar-valued quantum kernels that the paper attributes to the scalar setting.","marker":"Huang et al., 2021"},{"why":"Inductive-bias analysis of quantum kernels, another negative result motivating the shift to richer kernels.","marker":"Kübler et al., 2021"},{"why":"Introduced operator-valued kernels for learning vector-valued functions, the classical framework being generalized.","marker":"Micchelli & Pontil, 2005"},{"why":"Introduced entangled kernels beyond separability, the direct basis for Definition 3.1 of entangled QOVKs.","marker":"Huusari & Kadri, 2021"},{"why":"Introduced learning in reproducing kernel Hilbert C*-modules, underpinning the C*-algebraic roadmap.","marker":"Hashimoto et al., 2023b"},{"why":"Swap test used as the building block for the proposed quantum circuit implementation of the QOVK.","marker":"Buhrman et al., 2001"},{"why":"Formulated quantum channel identification, the proof-of-concept task in Section 5.1.","marker":"Fujiwara, 2001"}],"fun_headline_variants":["Quantum kernels need operator-valued forms to exploit entanglement","Scalar quantum kernels capped; operator-valued unlock entanglement","Entangled operator kernels cut channel recovery error nearly 6x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the dramatic improvement in the channel-estimation experiment comes from the entangled quantum operator-valued kernel itself, yet the experiment compares QOVK only against a scalar-valued quantum kernel, so the gain could in principle come from the operator-valued output structure alone and be matched by a classical operator-valued kernel.","fun_headline_variants_meta":{"raw":{"variants":["Quantum kernels need operator-valued forms to exploit entanglement","Scalar quantum kernels capped; operator-valued unlock entanglement","Entangled operator kernels cut channel recovery error nearly 6x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000853,"raw_usage":{"total_tokens":3760,"prompt_tokens":1048,"completion_tokens":2712,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":664,"completion_tokens_details":{"reasoning_tokens":2661}},"tokens_in":664,"tokens_out":2712,"duration_ms":18878,"temperature":1.0,"reasoning_tokens":2661,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:55:33.776700+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same bit-flip and dephasing channel estimation task with a classical operator-valued kernel (for example a separable OVK with an output kernel learned by cross-validation) under identical training data, noise, and evaluation protocol; if it achieves errors near the QOVK's 0.101 and 0.096, the case for a quantum-specific entangled advantage collapses. A second check is to replace the hand-fixed CNOT+SWAP unitary $U$ with a randomly chosen non-separable unitary: if recovery errors jump back to QSVK levels, the result depends on the specific $U$ rather than the entangled class.","supporting_citations":[{"cited_title":"Quantum machine learning in feature hilbert spaces","cited_arxiv_id":null,"evidence_quote":"Defined quantum machine learning in feature Hilbert spaces and the fidelity kernel used as the scalar-valued baseline."},{"cited_title":"Entangled kernels - beyond separability","cited_arxiv_id":null,"evidence_quote":"Introduced entangled kernels beyond separability, the direct basis for Definition 3.1 of entangled QOVKs."},{"cited_title":"Quantum fingerprinting","cited_arxiv_id":null,"evidence_quote":"Swap test used as the building block for the proposed quantum circuit implementation of the QOVK."},{"cited_title":"Quantum channel identification problem","cited_arxiv_id":null,"evidence_quote":"Formulated quantum channel identification, the proof-of-concept task in Section 5.1."}],"review_version":1}