{"id":"85d3cfb5-2f11-4bf6-9d8a-b97f73f3768a","arxiv_id":"1908.11762","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Supervised classifiers distinguish coherent from stochastic single-qubit noise on GST data, with near-perfect accuracy after feature engineering and margin-based robustness to sampling noise.","lead":"This paper shows that off-the-shelf machine learning classifiers can reliably tell apart two kinds of single-qubit noise, coherent and stochastic, using gate set tomography data. It maps out when simple linear classifiers work and when feature engineering is required, and tests robustness to finite sampling.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stochastic noise is sampled with a nonzero-mean generator (Eq. A8/A11) while coherent noise has zero mean, so the classifier may be learning first-order error magnitude rather than coherence.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the stochastic ensemble has a nonzero first-order mean generator while the coherent ensemble does not. This is not a remote edge case; it is visible in the paper's own equations (A4 vs A8/A11) and conflicts with the paper's stated definition of stochastic noise in Section III.1. The concern is load-bearing because if the classifier exploits this mean shift, the paper has not demonstrated that ML can learn a coherence discriminator; it may only have learned a noise-strength or first-order-decay threshold. The concern is mitigated but not resolved by Table III, which shows fixed-η separability for some subsets: separability of the exact noiseless feature vectors does not establish that the trained classifiers use coherence-related structure, nor does it quantify accuracy at each η. No code or data is provided, which compounds the difficulty of checking the simulations, but the primary logical flaw remains the uncontrolled first-order mean. Because the paper is explicitly a proof of principle and the required control is straightforward, the appropriate verdict is conditional pending the matched-η / matched-error-strength experiment, matching the reader's conditional verdict.","tokens_in":47234,"tokens_out":12192,"duration_ms":140549,"concrete_test":"Re-run the complete training/evaluation protocol of Section IV separately for each of the 19 η values, and also for coherent/stochastic pairs matched to equal average gate infidelity. Report cross-validated accuracy for each η (e.g., η = 10^-4, 10^-3, 10^-2, 0.1). If per-η accuracy at realistic small η is near chance or substantially below the pooled ~0.99 values, the reported separability is driven by first-order error magnitude rather than by coherence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central demonstration is undercut by an asymmetry in the two generative ensembles. Coherent errors are drawn from a zero-mean Gaussian Hamiltonian (Eq. A4), so the ensemble mean of the error generator is zero. Stochastic errors are instead drawn from a folded normal with positive rates (Eq. A8), and Eq. A11 shows ⟨h_jk⟩ ∝ η > 0: a deterministic, first-order dissipative component. This contradicts the paper's own definition in Section III.1 that stochastic fluctuations have zero expected value, and it means the two classes differ in their first-order average generator, not only in their coherence. Because η spans 1e-4 to 0.5, with equal sample counts across large and small strengths, a linear classifier can attain high pooled accuracy by thresholding on this first-order error-magnitude signal, e.g., labeling any dataset with a systematic O(η) depolarizing shift as stochastic. Fixing η does not remove the confound: the mean shift itself is proportional to η, so the stochastic class remains systematically displaced even at fixed strength. No per-η or matched-error-strength experiment is reported, so the claim that the learned map distinguishes coherent from stochastic noise is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a rubric for developing 'machine-learned' QCVV protocols, in which a supervised classifier replaces the statistical model and estimator of a conventional QCVV pipeline, and applies that rubric to the binary task of deciding whether single-qubit gate noise is coherent or stochastic. Training data are synthetic GST data sets produced by pyGSTi simulations of two noise ensembles: Gaussian unitary (coherent) perturbations and random-basis Lindblad (stochastic) decay. For each GST experiment design index L, the authors test linear separability with a linear program whose dual certifies inseparability, evaluate five classifiers under default and tuned hyperparameters, examine the low-dimensional structure of the L=1 data via PCA and MDS, introduce two quadratic feature maps that restore linear separability for L=1, and show that an SVM's margin predicts robustness to finite-sample noise. The main empirical claims are that GST data sets are linearly separable for L>1, that the L=1 data have a 'radio dish' geometry, and that feature engineering lets linear classifiers match or approach nonlinear ones.","tokens_in":47491,"tokens_out":9113,"duration_ms":89792,"significance":"As a proof of principle for using ML to learn QCVV data-processing maps, the paper is executed with unusual care: linear separability is certified by an LP with a dual witness, hyperparameters are tuned by explicit grid search, generalization is checked by repeated cross-validation and held-out 'hero tests', and the simulations use the standard pyGSTi toolchain. The rubric in Section II is a genuinely useful contribution that could structure future work in this area. However, the central demonstration is undermined by a mismatch between the two generative noise ensembles: the stochastic ensemble has a nonzero first-order generator mean while the coherent ensemble does not, so the learned decision surfaces may be separating on a first-order error-magnitude signal rather than on coherence. This is a fixable but load-bearing issue; with a matched-strength or per-η analysis, the paper could support its claims, but as written the claim that ML can learn to distinguish coherent from stochastic noise is not established.","major_comments":[{"comment":"The two simulated noise classes differ not only in coherence but in their first-order average generator. Coherent errors are drawn from a zero-mean Gaussian Hamiltonian (Eq. (A4), with ⟨He⟩=0 in Eq. (A9)), while the stochastic rates are drawn from a folded normal (Eq. (A8)) whose mean is nonzero: Eq. (A11) gives ⟨h_jk⟩ ∝ η > 0. Since η spans 10^-4 to 0.5 with equal sample counts per class, a linear classifier can attain high pooled accuracy by thresholding on the systematic O(η) dissipative shift rather than on coherence, and the confound survives at fixed η because the mean shift itself is proportional to η. This contradicts the paper's own definition in Section III.1 that stochastic errors are zero-mean fluctuations; the nonzero mean is computed in Eq. (A11) but never analyzed in the results. The central claim that the learned classifiers distinguish coherent from stochastic noise therefore requires a matched-error-strength control, per-η accuracy reporting, or an explicit analysis showing that the decision surface does not rely on the first-order generator.","section":"Appendix A, Eqs. (A4)-(A11); Section III.1"},{"comment":"The paper's restricted-η separability checks do not resolve the confound. Table III reports that fixed-η subsets of C1 are linearly separable, but this is exactly what a separator acting on the O(η) mean displacement of the stochastic ensemble would produce; separability at fixed η does not demonstrate separation on the basis of coherence. Similarly, the undersampling check in Table IV compares train and test accuracy pooled over all η values, so the ~0.97 test accuracy is also consistent with a mean-threshold explanation. Because no per-η accuracies, confusion matrices, or feature-ablation results are reported anywhere in Section IV, the proxy hypothesis is not ruled out by the present data.","section":"Table III; Section IV A 1"},{"comment":"The finite-sample robustness demonstration is internal to the confounded generative model: the SVM is trained and evaluated on the same folded-normal stochastic ensemble, so the computed margin and the accuracy-vs-noise curves certify robustness of a mean-shift separator rather than of a coherence classifier. The claim that the φSQ and φPP feature maps do not increase finite-sample sensitivity is likewise evaluated only on this ensemble. This section should be re-run, or at minimum reinterpreted, once the ensemble-matching issue is addressed.","section":"Section IV E, Figure 9"}],"minor_comments":[{"comment":"In the RBF SVM paragraph, 'in the original feautre space' should read 'feature space'.","section":"Section III 4 b"},{"comment":"Reference [59] is broken: the sentence 'We use “◦.' is incomplete and should be finished or removed.","section":"References, [59]"},{"comment":"The caption does not state what the points in the swarmplot represent (presumably per-fold accuracies across the K=20 splits); this should be clarified, and error bars or a box plot would improve readability.","section":"Figure 6"},{"comment":"The 'radio dish' interpretation is presented as strongly supported by the Choi-Jamiolkowski analogy, but the argument is explicitly heuristic (the text concedes 'the exact structure may not be comparable'); the qualitative claim should be labeled as such or backed by a quantitative test.","section":"Section IV B"},{"comment":"The statement that 'we do not use the extra 20900 feature vectors' appears in both Section IV A 1 and Appendix C; the hero-test procedure should be described once and referenced thereafter.","section":"Section IV A 1 and Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The central confound identified in the stress-test note is real: the stochastic ensemble has a nonzero first-order generator mean (Eq. A11) while the coherent ensemble does not, so the high accuracies, the linear-separability results, and even the 'radio dish' geometry are all consistent with a mean-shift surrogate. The fix is within the manuscript's scope: re-generating data with a centered or matched-error-strength stochastic ensemble, or adding per-η accuracy curves and an ablation of first-order features, would resolve the concern. The rubric contribution and the methodological care (LP certificates, cross-validation, hero tests) make the paper worth a revision rather than rejection. If the authors decline to run the control, the claims should be weakened to 'classifiers can separate these two particular generative ensembles' and the conclusions substantially reframed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a genuinely careful proof-of-principle: the rubric for building machine-learned QCVV maps is clear, the separability analysis uses an LP-based witness rather than guesswork, and the finite-sample margin test is the right way to think about robustness. Second, the central demonstration is compromised by an asymmetry in the simulated noise ensembles. Coherent errors are drawn from a zero-mean Gaussian Hamiltonian (Eq. A4), while stochastic errors are drawn from a folded normal (Eq. A8), so the stochastic generator has a nonzero mean proportional to η (Eq. A11). That contradicts the paper's own definition in Section III.1 that stochastic fluctuations have zero expected value, and it means the two classes differ in first-order error magnitude, not only in coherence. A linear classifier can achieve high pooled accuracy simply by thresholding on that shift. Restricting η does not remove the problem, since the shift grows with η, and no matched-error-strength or per-η classification experiment is reported. So the claim that the learned map distinguishes coherent from stochastic noise is not yet established.\n\nWhat the paper does well: the separability results for L>1 versus L=1 are new and interesting, and the feature-engineering maps φSQ and φPP are natural and well-motivated. The cross-validation and hero tests are honest, and the LP certificates make the geometry claims reproducible. The paper also lays out the costs of the ML approach clearly. No code or data ships, which is a secondary reproducibility gap.\n\nSoft spots in proportion: the mean-generator confound is load-bearing; the missing code is minor. The paper's own text supplies the contradiction, so this is not a manufactured concern.\n\nBottom line: this is a paper for people working on automated QCVV, and it deserves a serious referee. The referee should ask the authors to either fix the stochastic model to have genuinely zero-mean generators or to run experiments with matched first-order error strengths, per η. Without that, the proof-of-principle is about error magnitude, not coherence.","headline":"A careful ML-for-QCVV proof of principle whose central claim is undercut by a mean-generator asymmetry between the two simulated noise classes.","tokens_in":47980,"tokens_out":3902,"would_cite":false,"duration_ms":35845,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper reports that machine-learning classifiers can distinguish coherent from stochastic single-qubit noise using gate set tomography data, and that the two classes become linearly separable after feature engineering.","keywords":["machine-learned QCVV","gate set tomography","coherent noise","stochastic noise","support vector machine","feature engineering","single-qubit noise classification"],"falsifier":"Regenerate both training ensembles with matched first- and second-order error statistics -- for example, zero-mean Hamiltonian perturbations for coherent noise and zero-mean Lindblad jump coefficients for stochastic noise, with equal variances -- then retrain the same classifiers. If accuracy falls to chance, the original separation was driven by the average-error-size confound rather than by coherence.","tokens_in":47030,"feed_emoji":"⚛️","tokens_out":9755,"duration_ms":82469,"temperature":0.7,"pith_summary":"This paper asks whether machine learning can take over the data-processing step of quantum device characterization: instead of fitting a statistical model to experimental data, an algorithm learns the map from data to the property of interest directly. As a test case, it trains supervised classifiers to decide whether the noise on a single qubit is coherent (a fixed miscalibration) or stochastic (a randomly fluctuating control field), using the circuits of gate set tomography. The paper reports that the binary task is learnable: with enough circuit depth the two noise classes separate cleanly, and even the shortest circuits become linearly separable once simple feature engineering is applied, such as squaring outcome frequencies or adding their pairwise products. It also finds that a support-vector-machine classifier stays accurate under finite-sample statistical noise once the fluctuations fall below the geometric margin of its separating hyperplane, suggesting that machine-learned quantum characterization could automate parts of device diagnostics.","feed_headline":"Machine learning can label qubit noise type from GST data","feed_subtitle":"On simulated GST data, simple classifiers separate coherent from stochastic; engineered features beat sampling noise.","key_machinery":"The central object is the decision surface learned by a supervised classifier, viewed geometrically as a surface that separates gate set tomography feature vectors labeled “coherent” from those labeled “stochastic”. The argument is carried by two pieces: a linear-programming test for linear separability that can certify, via a theorem of alternatives, when no separating hyperplane exists; and the feature maps $\\varphi_{\\mathrm{SQ}}$ and $\\varphi_{\\mathrm{PP}}$, which append $f_j^2$ or $f_j f_k$ terms to the raw outcome-frequency vector and thereby unroll the curved “radio dish” geometry of the shallowest GST data so that a hyperplane can separate the classes. The geometric margin of the learned hyperplane is what connects training accuracy to robustness under finite-sample noise.","core_discovery":"The paper's central claim is that a supervised classifier can learn to distinguish coherent from stochastic single-qubit noise directly from gate set tomography (GST) data, without constructing a statistical noise model. For GST experiment designs with circuit depth index $L>1$, the paper finds that the collections of simulated GST frequencies are linearly separable: a hyperplane can separate feature vectors generated by coherent noise from those generated by stochastic noise with near-perfect accuracy once hyperparameters are tuned. For the shallowest design ($L=1$), the raw data are not linearly separable, but the paper proves that they become linearly separable after adding polynomial features -- the squares of individual outcome frequencies or all pairwise products -- and that linear classifiers then match intrinsically nonlinear ones. The paper also demonstrates that a maximum-margin linear SVM trained on noiseless engineered features handles finite-sampling noise reliably once the statistical fluctuations are below the hyperplane's geometric margin. The intended lesson is that off-the-shelf machine learning can automatically produce a working quantum-characterization data-processing map for this simplified binary property.","pith_inferences":["The simulated ensembles are not matched in their first-order statistics: stochastic noise has a nonzero average error generator proportional to the noise strength, while coherent noise has zero-mean Hamiltonian perturbations. The reported classifiers may therefore be solving a magnitude-detection problem in disguise, and a test with matched error moments would separate ‘coherence’ from ‘size of er","The “radio dish” structure seen in the low-dimensional embeddings suggests the coherence label is recoverable from distance-like quantities in a Choi-state picture, where pure states envelop mixed states; if so, a single hand-designed invariant such as the purity of the estimated process would probably give the same separation, making the engineered polynomial features a data-driven stand-in for t","A natural extension, implicit in the paper's own discussion, is a regression version that estimates a continuous coherence measure (such as unitarity) from the same GST data; the high accuracies reported for the binary problem suggest that regression target should also be learnable.","If this recipe generalizes beyond one qubit, quantum device diagnostics could be built as libraries of trained classifiers attached to fixed experiment designs, with experts contributing labels rather than statistical models; the dominant obstacle would be generating representative training data as processors grow."],"forward_implications":["A user could classify the noise on a qubit as coherent or stochastic by running standard GST circuits and applying a trained linear classifier, with no need to fit a full gate set or choose a noise model.","Increasing the GST circuit-depth index improves separability: for $L>1$, raw outcome frequencies are linearly separable, so a hyperplane classifier reaches high accuracy after hyperparameter tuning.","Feature engineering such as $\\varphi_{\\mathrm{SQ}}$ or $\\varphi_{\\mathrm{PP}}$ makes even the shallowest GST design linearly separable, showing that the information distinguishing the two noise types is present in short circuits but requires nonlinear feature maps to expose.","A maximum-margin SVM is a practical choice for this task because its accuracy degrades gracefully as finite-sample noise is added, staying essentially perfect once $1/\\sqrt{N_{\\mathrm{samples}}}$ falls below the margin.","The same rubric and training-data approach could be reused for other single-qubit properties, since only the property labels and performance measure would need to change."],"supporting_citations":[{"why":"supplies the gate set tomography experiment design whose circuits and outcome frequencies form the data for classification","marker":"[6]"},{"why":"introduces unitarity benchmarking, the established QCVV technique for estimating noise coherence that this classification task simplifies","marker":"[23]"},{"why":"demonstrates unitarity benchmarking on real hardware, supporting the practical relevance of distinguishing coherent from stochastic noise","marker":"[24]"},{"why":"provides the implementations of the LDA, QDA, perceptron, linear SVM, and RBF SVM classifiers used in the experiments","marker":"[60]"},{"why":"the software used to simulate noisy gate sets and generate the labeled GST training and testing feature vectors","marker":"[66]"},{"why":"supplies the theorem of alternatives used to certify whether each data collection is linearly separable","marker":"[98]"}],"fun_headline_variants":["ML separates coherent and stochastic qubit noise from GST","Machine learning classifies single-qubit noise from GST data","SVM distinguishes coherent vs stochastic noise in GST","Qubit noise type revealed by ML on GST data","Simple ML sorts qubit noise into coherent or stochastic"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulated experiments isolate coherence as the thing that differs between the two classes; in fact the stochastic-noise ensemble has a nonzero average error that grows with the noise-strength parameter while the coherent-noise ensemble has zero-mean error, so the classifier may be detecting error magnitude rather than coherence itself.","fun_headline_variants_meta":{"raw":{"variants":["ML separates coherent and stochastic qubit noise from GST","Machine learning classifies single-qubit noise from GST data","SVM distinguishes coherent vs stochastic noise in GST","Qubit noise type revealed by ML on GST data","Simple ML sorts qubit noise into coherent or stochastic"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00024,"raw_usage":{"total_tokens":1522,"prompt_tokens":951,"completion_tokens":571,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":494}},"tokens_in":567,"tokens_out":571,"duration_ms":5838,"temperature":1.0,"reasoning_tokens":494,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:06:56.295398+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Regenerate both training ensembles with matched first- and second-order error statistics -- for example, zero-mean Hamiltonian perturbations for coherent noise and zero-mean Lindblad jump coefficients for stochastic noise, with equal variances -- then retrain the same classifiers. If accuracy falls to chance, the original separation was driven by the average-error-size confound rather than by coherence.","supporting_citations":[{"cited_title":"Comparing Experiments to the Fault-Tolerance Threshold","cited_arxiv_id":"1510.05653","evidence_quote":"the software used to simulate noisy gate sets and generate the labeled GST training and testing feature vectors"},{"cited_title":"Agrawal, R","cited_arxiv_id":null,"evidence_quote":"supplies the theorem of alternatives used to certify whether each data collection is linearly separable"}],"review_version":1}