{"id":"ef0a8223-e337-4fe9-ada5-6927b58fb82f","arxiv_id":"2501.15359","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A neural network trained with one clean qubit on an NMR quantum processor learns a quantum data embedding that lifts MNIST 0/1 classification accuracy from 54% to 98%.","lead":"This paper trains a neural network to shape how images are loaded into a small quantum computer, using a type of quantum processor that is much easier to build than full quantum computers. The method improved handwritten digit classification from 54% to 98% accuracy in experiments on a nuclear magnetic resonance quantum processor.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 98%-vs-54% classification comparison changes both the quantum embedding and the classical preprocessing, so the DQC1-trained embedding's contribution is not isolated.","rationale":"The paper reports a credible proof-of-principle experiment, but the central empirical comparison is underdetermined. The reader's conditional verdict already identifies the classical-preprocessing confound as a second load-bearing premise; I agree that this is the most serious issue. I do not make the reader's primary concern about the loss being a poor surrogate for trace distance the main attack, because for the one-layer ZZ feature map used in the experiment, the DQC1-measured Hilbert-Schmidt inner product equals the overlap between the embedded states, so the implemented objective is better aligned with state distinguishability than the general formulation suggests. The missing baseline remains decisive: without a classically trained NN of the same architecture, the 54%-to-98% gain could be entirely a classical feature-learning effect. The data-split ambiguity noted by the reader is a secondary concern; the proposed ablation also forces explicit reporting of held-out accuracy, which would settle whether the 98% figure is a generalization result. Since the concern is concrete and addressable with one additional experiment, keeping the manuscript conditional is appropriate rather than rejecting or accepting outright.","tokens_in":14532,"tokens_out":15689,"duration_ms":147979,"concrete_test":"Retrain the same neural-network architecture on the same 500 MNIST 0/1 images, with the same PCA preprocessing, using a purely classical loss: cross-entropy on the network outputs (or a classical surrogate that maximizes Fisher separation between classes in output space), without any DQC1 or quantum feedback. Freeze the resulting g(x,w) and feed it into the identical one-layer ZZ feature map and the identical trained PQC classifier. Report the same metrics as the paper: PQC accuracy on a held-out test split, mean trace distance on training and test pairs, and the PQC loss curves. If this classical-NN baseline reaches or exceeds 98% accuracy (or produces comparable trace distance), the reported improvement is attributable to the classical neural-network preprocessing rather than to the DQC1-trained quantum embedding, and the headline claim must be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The main empirical claim is that NQE-DQC1 improves classification from 54% (traditional ZZ embedding) to 98% on 500 MNIST 0/1 images. In the NQE pipeline, PCA-reduced features are first passed through a trainable classical neural network g(x,w), and the network outputs are then fed into the ZZ feature map. In the 'traditional' baseline, the neural network is absent, so raw PCA features go directly into the same feature map. The comparison therefore varies two things at once: learned classical nonlinear preprocessing and the DQC1-based quantum trace-distance training. Because MNIST 0/1 is an easy binary problem and the NN operates on only 5 PCA components, a classically trained NN of the same architecture could plausibly produce well-separated clusters before any quantum encoding, making the downstream ZZ feature map and 4-parameter PQC little more than a nonlinear readout. The trace-distance increase shown in Fig. 2(c) is computed after this NN transformation, so it measures the full NQE map, not the contribution of the DQC1 training objective in particular. Without an ablation that uses the same NN architecture but trains it with a purely classical loss, the 98%-vs-54% headline cannot be attributed to the NQE-DQC1 protocol's central mechanism. I note that for the implemented one-layer ZZ feature map, V(gx) = P(gx)H^⊗n, so the HS inner product measured by DQC1 coincides with the overlap of the embedded states; thus the reader's surrogate-mismatch concern is weaker for this experiment. The missing classical-NN baseline is the primary threat to the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Neural Quantum Embedding via DQC1 (NQE-DQC1), a hybrid classical-quantum method for learning quantum feature maps for binary classification. A classical neural network g(x,w) maps input data to parameters of a ZZ feature map, and the network is trained by minimizing Eq. (1), a squared deviation between the normalized Hilbert-Schmidt inner product of the embedding unitaries and a label-matching target. The HS inner product is estimated experimentally with a DQC1 circuit on a four-qubit NMR processor. The authors report that training loss decreases, the trace distance between class-conditional embedded states increases, and a subsequent parameterized quantum circuit achieves 98% classification accuracy on 500 MNIST 0/1 images, versus 54% with a traditional ZZ feature map. They also transfer the trained embedding to IBM superconducting processors and include supplementary experiments on Fashion-MNIST and satellite images.","tokens_in":14815,"tokens_out":6044,"duration_ms":58092,"significance":"If the central claims hold, the work would be a useful demonstration that a subuniversal DQC1 model can train a classical network-to-quantum embedding on an ensemble quantum processor and that the trained embedding transfers to other hardware. The experimental execution appears careful: the DQC1 trace-estimation circuit is standard, the measured training loss tracks simulation, and the IBM transfer experiment adds credibility. The paper also ships numerical and experimental loss curves for a second dataset in the supplement. However, the headline accuracy comparison conflates the quantum embedding with classical preprocessing, and the relationship between the optimized loss and the stated trace-distance objective is not derived for the general case. These issues must be resolved before the main claims can be accepted.","major_comments":[{"comment":"The text states that Eq. (1) implements the goal of maximizing the trace distance between the class-conditional states rho+ and rho-, but the loss actually minimizes the squared deviation of normalized Hilbert-Schmidt inner products of the embedding unitaries from the label-matching target. The equivalence between these two objectives is asserted, not derived. For the specific single-layer ZZ feature map used here (M=1), V(gx)=P(gx)H^⊗n, so the normalized HS inner product coincides with the fidelity between the embedded states, which partly closes the gap in the implemented experiment. Nevertheless, the general claim in the protocol overview and the discussion of CPTP-map contractivity should be accompanied by a derivation or an explicit condition under which Eq. (1) is equivalent to maximizing trace distance.","section":"§Protocol overview, Eq. (1)"},{"comment":"The central 98%-versus-54% comparison changes two variables at once: the NQE condition prepends the trained classical neural network g(x,w) to the ZZ feature map, while the 'traditional' baseline feeds raw PCA features directly into the same feature map. Because MNIST 0/1 with five PCA components is already a very easy binary problem, the improved separation and accuracy could be caused entirely by the classical neural-network preprocessing, with no contribution from the DQC1-based training objective. To support the headline claim, the authors should include an ablation in which the same neural network architecture is trained with a purely classical loss (or used with random weights) and then fed into the same quantum embedding and PQC; the measured accuracy in that condition would isolate the effect of the NQE-DQC1 training.","section":"§Classification results, Fig. 4"},{"comment":"The reported 98.0% accuracy is computed over 'all 500 images' with no separate training/test breakdown, even though the dataset is partitioned into training and test sets and Fig. 4(a) shows only eight images per set. Generalization is a central claim of the paper, so the authors should report test-set accuracy separately, along with the number of PQC training iterations, the batch selection procedure, and the variance over repeated PQC training runs. Without these, the reader cannot distinguish memorization of the training set from genuine separation.","section":"§Classification results, Fig. 4(b)"},{"comment":"The neural network g(x,w) is never defined: no layer count, width, activation function, initialization, optimizer, or learning-rate schedule is given, and the paper does not state how ∇L_NQE is obtained from the DQC1-estimated loss. Since g is the trainable component of the protocol and the confounded-baseline question depends on its capacity, this omission prevents reproduction and makes it impossible to assess whether the network alone solves the classification task.","section":"§Experimental scheme, NQE training"}],"minor_comments":[{"comment":"There are several typos and inconsistent references: 'nuclear magnatic resonance' on page 2, 'sturcture' on page 4, and 'Eq. (B1)' in the main text should refer to Eq. (1).","section":"Throughout"},{"comment":"The PCA preprocessing is under-specified: the input dimensionality before PCA, whether PCA is fit on the training set only, and the normalization applied to the five PCA components should be stated.","section":"§Experimental scheme"},{"comment":"The trace-distance error bars are standard deviations over 20 randomly chosen pairs; please clarify whether the 20-pair selection was fixed across iterations and whether multiple independent NQE training runs were performed, since the reported curves otherwise reflect a single training trajectory.","section":"§Fig. 2(c)"},{"comment":"For the Fashion-MNIST and satellite datasets, only PQC loss curves are reported; end-task classification accuracies on held-out test data should be reported to support the claim that NQE improves these additional benchmarks.","section":"§Supplemental Material, Appendix E"}],"recommendation":"major_revision","confidential_remarks":"The experimental work appears competently executed and the DQC1 implementation is credible, but the attribution of the 98%-vs-54% result to the quantum protocol is the main editorial risk. Please ask the authors to state explicitly the incremental contribution over Refs. [14,15] and to add the missing ablation and architecture details."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful new thing here is an experimental demonstration that DQC1 can train a neural network to produce a quantum embedding on an NMR processor, and that the resulting embedding transfers to IBM superconducting hardware. The NQE concept and the Hilbert-Schmidt loss come from the group's own earlier work (Ref [14]); the DQC1-specific training loop, the NMR implementation, and the cross-platform transfer are the new pieces. The experiment looks self-consistent: the loss converges, trace distance increases for both training and test pairs, and the IBM results match the simulations.\n\nThe main soft spot is the headline comparison. The 98%-versus-54% classification result changes two things at once: the NQE pipeline inserts a trainable classical neural network before the ZZ feature map, while the baseline uses raw PCA features. So the comparison does not isolate the contribution of the DQC1-training objective. A classical neural network with the same architecture, trained with a standard loss on the same five PCA components, could plausibly produce comparable separation on MNIST 0/1, making the quantum embedding a near-trivial readout. Without an ablation that holds the NN architecture fixed and trains it with a classical loss, or at least reports the class separation before quantum encoding, the 98% number cannot be attributed to the NQE-DQC1 mechanism. This is the primary threat to the claim, and it is addressable.\n\nThe reader's worry that the loss is misaligned with trace distance is weaker than it first appears. For the implemented one-layer ZZ feature map, V(gx) = D(gx)H^⊗n, so the real part of the normalized Hilbert-Schmidt inner product equals the real part of the state fidelity. Optimizing toward 0/1 targets does push trace distance in the right direction. Minor caveat: the loss uses only the real part, leaving the imaginary part of the fidelity unconstrained, but the experimental trace-distance curves suggest this is not a practical issue here.\n\nThe data split for the 98% accuracy is ambiguous; it is not clear whether the 500 images are the training set, the test set, or the union. The authors should state this explicitly. That is a minor reporting issue.\n\nThis is a solid experimental methods paper for the QML-on-NMR subfield and deserves a serious referee. The referee should ask for the missing classical-NN baseline and a clear statement of the train/test split before the 98%-versus-54% claim is taken at face value.","headline":"Solid DQC1-on-NMR proof of principle, but the headline classification gain is confounded by the classical neural network and a missing baseline.","tokens_in":15450,"tokens_out":5134,"would_cite":false,"duration_ms":43752,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68"],"pacs":["03.67.Lx","76.60.-k"],"model":"deepseek-v4-flash","headline":"One clean qubit can train a quantum data embedding, lifting MNIST '0/1' classification on NMR hardware from 54% to 98%.","keywords":["neural quantum embedding","DQC1","one clean qubit","quantum machine learning","quantum data embedding","NMR quantum processor","trace distance","Hilbert–Schmidt inner product"],"falsifier":"A decisive ablation: take the trained network outputs $g(x,w)$ for the 500 images and classify them with a classical linear or two-layer model, bypassing the quantum feature map and the parameterized circuit entirely. If that classical pipeline already reaches roughly 98% accuracy, the improvement is attributable to classical preprocessing rather than to the quantum embedding; if it stays near 54%, the separation is genuinely created by the embedding. A second consistency check is to compute the exact trace distance $\\|\\rho_+ - \\rho_-\\|_1$ between the class-conditional states after each NQE iteration and compare it with the NQE loss; if the loss decreases while the trace distance saturates or drops, the surrogate objective is not tracking the separation it claims to maximize.","tokens_in":14307,"feed_emoji":"🧠","tokens_out":27978,"duration_ms":226755,"temperature":0.7,"pith_summary":"The paper's central claim is that the choice of quantum data embedding—normally a fixed, hand-picked feature map—can instead be learned, and learned with very little quantum hardware. The authors train a classical neural network whose outputs set the angles of a $ZZ$-feature map, using the DQC1 protocol to estimate the overlap between two embedding circuits from a single readout on one clean qubit. On a four-qubit NMR processor encoding handwritten '0' and '1' images, the trained embedding increases the trace distance between the two classes and lifts classification accuracy from 54% (near random guessing) to 98% for the same shallow classifier. If this holds, ensemble-style quantum machines usually set aside as too weak for machine learning become practical trainers for embeddings that later run on other hardware, and the field's bottleneck shifts from the classifier circuit to the data-loading stage.","feed_headline":"One clean qubit retrains quantum embedding: MNIST 0/1 jumps 54% to 98%","feed_subtitle":"One clean qubit in an NMR sample trains the embedding; the same weights then run on other chips.","key_machinery":"The load-bearing object is the DQC1 overlap estimator: starting from one pure probe qubit and $n$ maximally mixed register qubits, $\\rho_0 = |0\\rangle\\langle 0| \\otimes I/2^n$, and applying $U = H_1 V_c^\\dagger(g_{x_2}) V_c(g_{x_1}) H_1$ with controlled feature maps, the probe's expectation value $\\langle \\sigma_z \\rangle = \\mathrm{Re}[\\mathrm{Tr}(V(g_{x_1})V^\\dagger(g_{x_2}))]/2^n$ returns the normalized Hilbert–Schmidt inner product between the two embedding unitaries in a single readout. Substituted into the NQE loss, that number makes the whole pipeline—neural network, feature map, and loss—differentiable with respect to the network weights $w$, so ordinary gradient descent can learn the embedding. Two supporting facts carry the argument: trace distance is contractive under CPTP maps, so the embedding is the only stage where class separation can be created, and the NMR platform initializes the required one-clean-qubit state with gate fidelities above 99.5%, making the experimental overlap estimates trustworthy.","core_discovery":"On the paper's own terms, the discovery is that the separation of quantum-embedded data can be optimized end-to-end with only one clean qubit. The NQE-DQC1 protocol composes a classical neural network $g(x,w)$ with the $ZZ$-feature map $V(\\phi) = (\\exp[i\\sum_k \\phi_k Z_k + \\phi_{n+k}Z_kZ_{k+1}] H^{\\otimes n})^M$ at $M=1$, so the network's outputs set the feature-map angles. The training signal is the normalized Hilbert–Schmidt inner product $\\langle V(g_{x_i}), V(g_{x_j})\\rangle_{HS} = \\mathrm{Tr}(V(g_{x_i})V^\\dagger(g_{x_j}))/2^n$, which DQC1 retrieves as the $z$-expectation value of a single probe qubit prepared in $|0\\rangle\\langle 0| \\otimes I/2^n$. Minimizing the squared deviation of that overlap from 1 for same-class pairs and 0 for different-class pairs is the protocol's objective, and the paper takes this to maximize the trace distance between the class-averaged states $\\rho_+$ and $\\rho_-$; because any later quantum operation is a contractive map, that distance bounds every downstream classifier. Experimentally the loss converges around the tenth iteration, the measured trace distance rises on both training and test images, and a small two-layer parameterized circuit reaches 98% accuracy with the trained embedding versus 54% with the raw feature map, with the same trained network also guiding classification on a superconducting processor.","pith_inferences":["A natural control the paper does not run: feed the trained network outputs $g(x,w)$ directly to a classical classifier, skipping the quantum feature map and PQC. If those features already separate the 500 images at near 98%, part of the reported gain is classical preprocessing; if they stay near 54%, the separation truly happens in the embedding.","The NQE loss is an indirect objective—it matches the average overlap of whole embedding circuits to the label pattern rather than the fidelity of the specific embedded states used at inference. A direct test would train against $\\|\\rho_+ - \\rho_-\\|_1$ itself on a small register and compare the trace distance actually achieved.","The paper demonstrates transfer in one direction, from NMR-trained embeddings to a superconducting processor; the complementary direction—training on a platform with higher gate fidelities and transferring back—is untested, though nothing in the protocol forbids it.","Because the single-readout overlap estimate is independent of the number of register qubits, the training signal's cost scales with the complexity of implementing the controlled feature maps rather than with register size—a structural hint that this training scheme, if correct, has room to grow beyond three encoding qubits."],"forward_implications":["If the protocol is right, ensemble quantum systems such as NMR become a practical training resource for quantum machine learning: the DQC1 training loop needs only one clean qubit and a single readout per overlap estimate, which is exactly the regime these platforms can support.","A fixed, hand-picked feature map can be the binding constraint on a quantum classifier: with the same shallow parameterized circuit, accuracy on the MNIST 0/1 task jumps from 54% to 98% purely by replacing the raw feature map with the trained one.","The trained embedding transfers across hardware: the neural network weights learned on the NMR processor were reused to run classification on a cloud superconducting processor, with results tracking numerical simulation.","The protocol generalizes beyond the demonstration: supplementary runs on Fashion-MNIST and satellite imagery, under simulated NMR noise and on a 127-qubit superconducting processor, show the same qualitative advantage of NQE embedding over the raw $ZZ$-feature map."],"supporting_citations":[{"why":"Defines DQC1, the one-clean-qubit model this protocol uses to estimate the normalized trace of a unitary from a single probe measurement.","marker":"[10]"},{"why":"Introduces the $ZZ$-feature map for quantum-enhanced feature spaces; this is the embedding family NQE parameterizes and the baseline its accuracy is compared against.","marker":"[6]"},{"why":"The earlier neural quantum embedding framework this work extends; it motivates the trace-distance objective and supports the claim that NQE absorbs dimensionality reduction.","marker":"[14]"},{"why":"Standard reference for the Hilbert–Schmidt inner product and trace distance used in the loss and the evaluation, and for the statement that noise contracts trace distance.","marker":"[20]"},{"why":"Supplies the state-discrimination bound connecting trace distance to the minimum achievable misclassification probability.","marker":"[27]"},{"why":"Provides the contractivity inequality showing no subsequent CPTP map can increase the trace distance set by the embedding.","marker":"[37]"},{"why":"The MNIST handwritten-digit dataset on which the 54%-versus-98% classification experiment is run.","marker":"[39]"},{"why":"The optimal-control pulse-engineering method that gives the above-99.5% gate fidelities supporting the NMR demonstration.","marker":"[44]"},{"why":"The spatial-averaging procedure that prepares the pseudo-pure state from which the one-clean-qubit initial state is built on the NMR processor.","marker":"[45]"},{"why":"The parameter-shift rule used to train the parameterized classifier circuit after NQE embedding.","marker":"[46–49]"}],"fun_headline_variants":["One clean qubit trains quantum embedding: accuracy jumps to 98%","NQE-DQC1: neural embedding training with one clean qubit","98% accuracy from one-qubit-driven quantum embedding","DQC1 enables neural embedding: MNIST accuracy 98% vs 54%","Neural network sets embedding angles with one clean qubit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that matching the average overlap of whole embedding circuits to the label pattern—the quantity DQC1 measures with one clean qubit—faithfully maximizes how far apart the two classes of embedded quantum states actually end up, and that the large accuracy gain comes from the quantum embedding rather than from the classical neural network's preprocessing of the images.","fun_headline_variants_meta":{"raw":{"variants":["One clean qubit trains quantum embedding: accuracy jumps to 98%","NQE-DQC1: neural embedding training with one clean qubit","98% accuracy from one-qubit-driven quantum embedding","DQC1 enables neural embedding: MNIST accuracy 98% vs 54%","Neural network sets embedding angles with one clean qubit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000591,"raw_usage":{"total_tokens":2860,"prompt_tokens":1125,"completion_tokens":1735,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":741,"completion_tokens_details":{"reasoning_tokens":1643}},"tokens_in":741,"tokens_out":1735,"duration_ms":11883,"temperature":1.0,"reasoning_tokens":1643,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:22:01.035452+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive ablation: take the trained network outputs $g(x,w)$ for the 500 images and classify them with a classical linear or two-layer model, bypassing the quantum feature map and the parameterized circuit entirely. If that classical pipeline already reaches roughly 98% accuracy, the improvement is attributable to classical preprocessing rather than to the quantum embedding; if it stays near 54%, the separation is genuinely created by the embedding. A second consistency check is to compute the exact trace distance $\\|\\rho_+ - \\rho_-\\|_1$ between the class-conditional states after each NQE iteration and compare it with the NQE loss; if the loss decreases while the trace distance saturates or drops, the surrogate objective is not tracking the separation it claims to maximize.","supporting_citations":[{"cited_title":"Havl´ıcek, A","cited_arxiv_id":null,"evidence_quote":"Introduces the $ZZ$-feature map for quantum-enhanced feature spaces; this is the embedding family NQE parameterizes and the baseline its accuracy is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Standard reference for the Hilbert–Schmidt inner product and trace distance used in the loss and the evaluation, and for the statement that noise contracts trace distance."},{"cited_title":"Bae and L.-C","cited_arxiv_id":null,"evidence_quote":"Supplies the state-discrimination bound connecting trace distance to the minimum achievable misclassification probability."},{"cited_title":"Siudzi ´nska, S","cited_arxiv_id":null,"evidence_quote":"Provides the contractivity inequality showing no subsequent CPTP map can increase the trace distance set by the embedding."},{"cited_title":"Khaneja, T","cited_arxiv_id":null,"evidence_quote":"The optimal-control pulse-engineering method that gives the above-99.5% gate fidelities supporting the NMR demonstration."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The spatial-averaging procedure that prepares the pseudo-pure state from which the one-clean-qubit initial state is built on the NMR processor."}],"review_version":1}