{"id":"4d551619-7d1e-4b0a-a4a9-5207ad7b4eff","arxiv_id":"2509.03233","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Fisher Linear Discriminant Analysis, trained on Pauli expectation values from simulated two-, three-, and four-qubit states, classifies entangled versus separable states with reported accuracies between 33% and 100% depending on the overlap regime.","lead":"This paper trains a classical machine-learning classifier, Fisher Linear Discriminant Analysis, on simulated measurements of two-, three-, and four-qubit states to sort them into entangled versus separable. It is a pedestrian application of a textbook classifier to small qubit systems, with idealized simulations and no comparison to existing entanglement detection methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'scalable' claim fails because FLDA's input and scatter matrices grow exponentially with qubit number, and the Werner benchmark is nearly tautological since label parameter p appears directly in the features.","rationale":"The reader rejected the paper primarily because the simulations are noiseless and accuracies do not transfer to practice. I agree with that, and the paper's own discussion says noise was not included and would lower accuracy. But I see a more load-bearing, structural problem: the word 'scalable' in the title and abstract is contradicted by the method's own resource requirements. The feature space is the full set of 4^N−1 Pauli expectation values; S_W is a dense matrix of that dimension; the generalized eigenproblem costs O(n^3). Nothing in Secs. 2–3 reduces this before the exponential cost is paid, and the paper presents no scaling law, only N=2,3,4. This is not a question of experimental noise but of the algorithm's intrinsic cost. In addition, the Werner-state experiments are weak evidence because the label-defining parameter p appears directly in the measured correlators, making the high accuracies partly tautological; the concurrence benchmark, where this leakage is absent, produces a below-random 0.33 test accuracy in the high-overlap regime. Together these points support the reader's REJECT verdict without changing it; the concern is about missing support for the central scalability claim rather than a new fatal flaw.","tokens_in":10792,"tokens_out":7576,"duration_ms":75824,"concrete_test":"Analytically count the resources required by the described FLDA pipeline for N=6 and N=8: number of Pauli features (4^N−1), memory to store S_W, and operations for the generalized eigenproblem. If the counts grow exponentially with no feature-selection or sampling mitigation, the 'scalable' claim is refuted by the paper's own Sec. 2–3 construction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is scalable entanglement detection, but its own pipeline (Sec. 3) constructs feature vectors from all 4^N−1 non-identity Pauli expectation values, then builds S_W and S_B of that size and solves a generalized eigenproblem at O(n^3) (Sec. 2). For N qubits this is exponential in N: estimating all Pauli correlators requires exponentially many measurement settings, S_W is (4^N−1)^2 entries, and the eigen-solve is O(4^{3N}). FLDA's reduction to k−1 dimensions occurs after this exponential cost, so the method is not scalable as claimed. Only N=2,3,4 are demonstrated, with no scaling analysis. The paper even concedes that 'the high dimensionality of multi-qubit systems also necessitates a large number of measurements.' Separately, the Werner benchmarks are weak evidence: in Eq. (6) the features include ⟨σ_i⊗σ_i⟩=p s_i, so the label threshold p>1/3 is encoded in a single feature; high accuracy there is nearly tautological. The one non-trivial two-qubit result, concurrence, yields test accuracy 0.33 in the high-overlap case (Table 2), below random. The discussion's admission that noise was omitted and would 'almost certainly lower these accuracy figures' further undercuts the practical-tool claim, but the exponential resource cost independently invalidates the scalability promise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes to use classical Fisher Linear Discriminant Analysis (FLDA) as a machine-learning tool for classifying quantum states as entangled or separable. Feature vectors are formed from expectation values of Pauli observables; the between-class and within-class scatter matrices S_B and S_W are computed on a labeled training set (labels from the Peres-Horodecki criterion or from concurrence), and the generalized eigenproblem S_B w = λ S_W w is solved to obtain a projection direction. The authors report train/test accuracies for two-qubit Werner and concurrence states, three-qubit GHZ-based Werner, PPT, 'alternative PPT', and biseparable states, and one four-qubit Werner case, and conclude that the method is simple, efficient, scalable, interpretable, and promising for real-world experiments.","tokens_in":11069,"tokens_out":13480,"duration_ms":109305,"significance":"The FLDA formalism itself is standard, and applying it to Pauli-expectation-value features is a natural idea that could in principle yield an interpretable baseline for entanglement classification. However, the manuscript does not substantiate its central claims: the scalability claim is contradicted by the exponential feature-space construction, the three-qubit and 'alternative PPT' benchmarks contain physics errors, the Werner benchmark is near-tautological because the label parameter appears directly in the features, and the one non-trivial two-qubit result falls below chance in the high-overlap case. The paper also provides no code, data, hyperparameters, or error bars, and its own discussion concedes that noise was omitted. As a result, the reported accuracies do not establish a practical or scalable entanglement-detection method.","major_comments":[{"comment":"The state called an 'alternative PPT entangled state' in Eq. (11) is not entangled. Expanding ρ = (1/8)(I⊗I⊗I + I⊗σ_z⊗σ_z + σ_z⊗I⊗σ_z + σ_z⊗σ_z⊗I) in the computational basis gives a diagonal matrix with entries 1/2 on |000⟩ and |111⟩ and zero elsewhere, i.e., ρ = (|000⟩⟨000| + |111⟩⟨111|)/2, which is a separable mixture of product states. Therefore Table 5 classifies separable states against separable states, and the Discussion's interpretation of the lower accuracy for the 'PPT entangled state' is based on a mislabeled class.","section":"Section 3.2, Eq. (11), Table 5"},{"comment":"The text states both 'full separability for p≤1/3' and 'separable across all bipartitions for p≤1/5' for the three-qubit Werner state; these two statements are mutually inconsistent. Moreover, in tripartite systems the PPT condition is necessary but not sufficient for full separability, so a separability threshold cannot be read off from the PPT eigenvalue. Since the classification methodology is carried over from Sec. 3.1, the labels used for Table 3 appear to treat PPT-entangled states with 1/5 < p ≤ 1/3 as separable, which would contaminate the training data and invalidate the reported accuracy as a measure of entanglement detection.","section":"Section 3.2, Eq. (9), Table 3"},{"comment":"The central scalability claim is not supported by the proposed pipeline. The feature space is the set of all 4^N−1 non-identity Pauli expectation values, so forming the feature vectors, estimating S_W and S_B, and solving the generalized eigenproblem at O(n^3) with n=4^N−1 all require resources exponential in the qubit number N. Only N=2,3,4 are demonstrated, with no scaling analysis, and the Discussion concedes that 'the high dimensionality of multi-qubit systems also necessitates a large number of measurements.' The paper therefore does not deliver scalable entanglement detection as claimed.","section":"Section 3, 'Feature vectors' and 'Computational Efficiency'; Section 2"},{"comment":"For the Werner-state benchmark, the label is essentially contained in the features. In Eq. (6), the features include ⟨σ_i⊗σ_i⟩ = p s_i, so the separable/entangled boundary p=1/3 is a simple threshold on a single feature. High accuracy on this family is therefore near-tautological and does not demonstrate that FLDA has learned a nontrivial entanglement signature; the same parametric dependence carries over to the GHZ-based three- and four-qubit Werner states.","section":"Section 3.1, Eq. (6), Tables 1, 3, 7"},{"comment":"The concurrence-based two-qubit experiment, which is the non-trivial test, reports a test accuracy of 0.33 in the high-overlap case, below the random-guessing baseline of 0.5, with a Fisher criterion of only 0.012. This directly contradicts the Discussion's statement that 'even in high-overlap scenarios, FLDA often maintains high classification accuracy.' The single most informative numerical result in the paper is therefore evidence against, not for, the method's practical effectiveness.","section":"Section 3.1, Table 2; Section 4"},{"comment":"The evaluation is not reproducible as reported. No code or data are provided; the regularization parameter ϵ in Eq. (5), the normalization/standardization choice, the train/test split, and the procedure for generating the 10,000 samples from N trials = 1000 are not specified; and no error bars are given. In addition, the simulations are noiseless, and the Discussion admits that experimental noise 'would almost certainly lower these accuracy figures.' Since the stated application is real-world experiments, these omissions are load-bearing for the paper's central claim.","section":"Section 3.1, Algorithm 1; Section 4"}],"minor_comments":[{"comment":"The matrices are displayed as 3×5 arrays, but the text says W is a 1×15 projection vector; please clarify how the flattening is done.","section":"Section 3.1, Eqs. (7)-(8)"},{"comment":"The quantity 'FLD Threshold' is never defined, and the classification rule in Algorithm 1 (closest projected mean) should be reconciled with the threshold values reported in the tables.","section":"Tables 1-7"},{"comment":"The relationship between 'N trials = 1000', 'N train = 10000', and the number of features is unclear; please specify how independent samples are generated and whether the features are noisy or exact expectation values.","section":"Section 3.1, Algorithm 1"},{"comment":"The claims of 'systematically evaluate' and 'high accuracy' are not matched by the four-qubit section, which contains a single high-overlap case.","section":"Abstract and Introduction"},{"comment":"References [27] and [35] are duplicate entries of the same paper; please remove one.","section":"References"},{"comment":"Axis labels and units are missing, and the text reference 'see Fig. 3, 2' in Section 3.1 is malformed.","section":"Figures 2-8"}],"recommendation":"reject","confidential_remarks":"To the editor: The manuscript has multiple load-bearing technical errors (the 'alternative PPT' state is separable, the three-qubit PPT/separability threshold is internally inconsistent) and its central scalability claim conflicts with its own resource estimates. The evaluation also lacks code, data, noise modeling, and baselines. These issues go beyond presentation and cannot be fixed by local revision; I recommend rejection, though a thoroughly revised and resubmitted version that resizes the claims to a small-scale demonstration might be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a textbook classifier (Fisher LDA from 1936) applied to two-, three-, and four-qubit states, with no code, no baselines, no noise, and a title claim that the pipeline itself contradicts. I would not send this to referees.\n\nWhat the paper does well: it is clearly written and honest about its limits. The authors explicitly say noise was not included and would almost certainly lower the accuracies. They report the bad result honestly too: test accuracy 0.33 for the high-overlap concurrence case, below random. The FLDA equations are standard and correct. The observation that the discriminant vector W can point to the most informative Pauli observables is a genuine, if modest, practical advantage.\n\nWhere it falls down. The central claim is scalability, but the method's feature vector is all 4^N-1 non-identity Pauli expectation values, so the scatter matrices have exponential size and the generalized eigenproblem costs O(4^{3N}). FLDA's dimensionality reduction happens after that exponential cost, not before it. Demonstrations go only to N=4, with no scaling analysis, and the paper even concedes that high dimensionality requires many measurements. So 'scalable' is unsupported in the only direction that would matter.\n\nThe benchmark choice is also weak evidence. For the Werner states, the label threshold p>1/3 is essentially encoded in a single feature: Eq. (6) gives c_ii = p s_i, so the classifier is close to reading off the label. The one non-trivial two-qubit test, the concurrence case, yields a sub-random 0.33 in the high-overlap regime, which the paper reports but does not meaningfully discuss. No error bars, no measurement-setting counts, no comparison against the SVM and neural-network entanglement detectors the paper itself cites. The citation pattern is fine but mostly standard background plus the authors' own related work.\n\nWho is this for? Possibly a student looking for a first example of FLDA on qubit states. It does not advance entanglement detection beyond the cited ML literature, and the practical-tool and scalability promises are not supported by the evidence in the paper. I'd desk reject rather than spend referee time on it.","headline":"A clean but routine FLDA demo on small simulated states; the 'scalable' claim is contradicted by the method's own exponential feature-space cost.","tokens_in":11630,"tokens_out":1327,"would_cite":false,"duration_ms":14897,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a classical linear classifier, trained on Pauli expectation values, separates entangled from separable two-, three-, and four-qubit states with high accuracy in noise-free simulations.","keywords":["quantum entanglement","entanglement detection","Fisher linear discriminant analysis","Werner states","positive partial transpose","machine learning","quantum state classification","Pauli observables"],"falsifier":"Re-run the two-, three-, and four-qubit experiments with the same state families and FLDA pipeline, but estimate each Pauli expectation value from a finite number of projective measurements (for example, 1000 shots) rather than from the exact trace. If the high-overlap test accuracies collapse toward chance — in particular if the two-qubit Werner 0.89 and three-qubit GHZ 0.87 figures cannot be reproduced — the noiseless-simulation claim is isolated to an ideal limit and the practical detection claim fails.","tokens_in":10568,"feed_emoji":"⚛️","tokens_out":12486,"duration_ms":103600,"temperature":0.7,"pith_summary":"The paper claims that Fisher linear discriminant analysis (FLDA), a classical statistical classifier, can detect quantum entanglement by learning a single linear projection of measured Pauli expectation values. In noise-free simulations on two-, three-, and four-qubit systems it reports high test accuracy for Werner states, positive-partial-transpose states, and a biseparable family, with accuracy degrading as the entangled and separable classes overlap more heavily. The largest drop appears in the two-qubit concurrence-based family, where high-overlap test accuracy is reported as 0.33. The motivation is practical: FLDA reduces the exponentially large space of Pauli observables to a one-dimensional score and returns a discriminant vector that identifies which measurements carry the most entanglement information. The paper states explicitly that noise was not included in its idealized simulations, so the accuracies describe an ideal measurement limit rather than an immediate experimental method.","feed_headline":"One classical projection separates entangled from separable qubits","feed_subtitle":"Trained on Pauli expectation values, FLDA classifies entangled versus separable states in noise-free simulations.","key_machinery":"The central object is the Fisher linear discriminant vector $\\mathbf{w}$, obtained by maximizing the Fisher criterion $J(\\mathbf{w}) = \\frac{\\mathbf{w}^{T} S_B \\mathbf{w}}{\\mathbf{w}^{T} S_W \\mathbf{w}}$ and solving the generalized eigenvalue problem $S_B \\mathbf{w} = \\lambda S_W \\mathbf{w}$, where $S_B$ is the between-class scatter and $S_W$ the within-class scatter of the feature vectors. The feature vectors are built from measurement outcomes, specifically expectation values $\\langle \\sigma_k \\rangle = \\mathrm{tr}(\\rho \\sigma_k)$ of single- and multi-qubit Pauli observables, so the method is a direct pipeline from quantum measurements to a classical one-dimensional score. The labeled examples are Werner states with mixing parameter $p$, PPT entangled states, and biseparable states, with separability assigned by the Peres-Horodecki criterion or by the construction of the state. When $S_W$ is singular it is regularized as $S_W + \\epsilon I$, and classification is done by comparing the projected test point to the projected class means.","core_discovery":"The central claim is that a Fisher linear discriminant computed from exact Pauli expectation values separates entangled from separable states with high accuracy for the state families tested. For two-qubit Werner states the classifier reaches 0.89, 0.92, and 1.00 test accuracy in the high-, medium-, and low-overlap regimes; for the two-qubit concurrence-parametrized family the high-overlap accuracy is 0.33 while the low-overlap accuracy is 1.00. Three-qubit GHZ-type Werner states score 0.87, 0.96, and 1.00 across the same overlap levels, three-qubit PPT families score between 0.65 and 1.00, and the tested three-qubit biseparable and four-qubit Werner cases reach 1.00 even at high overlap. The paper interprets the decline with overlap as the expected behavior of a linear method: when the projected class means are close, the Fisher criterion is small and classification becomes unreliable. It also treats the discriminant vector as an interpretable output that ranks the Pauli observables by their usefulness for entanglement detection.","pith_inferences":["A natural test the paper does not run is to inject finite shot noise by sampling Pauli outcomes according to the Born rule instead of using exact expectation values; the paper's own discussion predicts that noise inflates the within-class scatter and lowers accuracy, and the high-overlap concurrence case already sits at 0.33 in the noiseless limit.","The FLDA boundary is functionally an entanglement witness: the discriminant vector defines a linear functional on the space of Pauli expectation values whose sign is the classification decision, so the trained classifier could be recast as a witness whose validity for the tested families could in principle be certified rather than only measured empirically.","Testing on states outside the parametrized families — random mixed states, bound entangled states in other dimensions, or states whose separability boundary is nonlinear in Pauli space — would show whether the linear boundary generalizes or is specific to Werner-type structure.","The four-qubit perfect score at high overlap should be read as an upper bound on what the method can do, since the two- and three-qubit results show accuracy falling with overlap; realistic extensions with noise would likely require more informative features or nonlinear classifiers."],"forward_implications":["A single scalar projection of Pauli expectation values suffices to separate entangled from separable states for the Werner, PPT, and biseparable families tested, so FLDA offers a low-cost classification route in noise-free settings.","The discriminant vector ranks observables by their influence on the entanglement decision, which can guide experimental measurement selection toward a small set of informative Pauli terms.","Class separability, as measured by the Fisher criterion, tracks classification accuracy: well-separated classes reach perfect test accuracy, while high-overlap classes degrade, most sharply for the two-qubit concurrence case at 0.33.","Because FLDA reduces the $4^N-1$-dimensional Pauli feature space to one dimension for a two-class problem, the per-classification cost depends on the number of measured features, not on the exponential Hilbert-space dimension."],"supporting_citations":[{"why":"Supplies the classical FLDA formulation and the Fisher criterion that the whole classification pipeline is built on.","marker":"[43,44]"},{"why":"Provides the generalized eigenvalue problem used to compute the optimal projection vector.","marker":"[45]"},{"why":"Defines the three-qubit mixed-state families and PPT entangled states that are used as test cases.","marker":"[46]"},{"why":"Supplies the alternative three-qubit PPT state representation tested in the classification experiments.","marker":"[47]"},{"why":"Provides the two-qubit circuit and concurrence-based entanglement parameterization used to generate labeled training data.","marker":"[27]"}],"fun_headline_variants":["Fisher discriminant on Pauli values marks entangled states","100% accuracy on 4-qubit Werner states via FLDA","FLDA classifies entangled vs separable in noise-free sims","Pauli expectation values unlock Fisher entanglement test"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracies depend on measurement data with no noise or statistical fluctuation — the simulations use exact average outcomes — and the paper itself notes that real experimental noise would almost certainly lower them.","fun_headline_variants_meta":{"raw":{"variants":["Fisher discriminant on Pauli values marks entangled states","100% accuracy on 4-qubit Werner states via FLDA","FLDA classifies entangled vs separable in noise-free sims","Pauli expectation values unlock Fisher entanglement test"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000538,"raw_usage":{"total_tokens":2559,"prompt_tokens":898,"completion_tokens":1661,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":1596}},"tokens_in":514,"tokens_out":1661,"duration_ms":11004,"temperature":1.0,"reasoning_tokens":1596,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:31:13.466578+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the two-, three-, and four-qubit experiments with the same state families and FLDA pipeline, but estimate each Pauli expectation value from a finite number of projective measurements (for example, 1000 shots) rather than from the exact trace. If the high-overlap test accuracies collapse toward chance — in particular if the two-qubit Werner 0.89 and three-qubit GHZ 0.87 figures cannot be reproduced — the noiseless-simulation claim is isolated to an ideal limit and the practical detection claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the generalized eigenvalue problem used to compute the optimal projection vector."},{"cited_title":"Classification of mixed three-qubit states,","cited_arxiv_id":null,"evidence_quote":"Defines the three-qubit mixed-state families and PPT entangled states that are used as test cases."},{"cited_title":"Detect- ing some three-qubit mub diagonal entangled states via nonlinear opti- mal entanglement witnesses,","cited_arxiv_id":null,"evidence_quote":"Supplies the alternative three-qubit PPT state representation tested in the classification experiments."},{"cited_title":"Entanglement detection with quantum support vector machine (qsvm) on near-term quantum devices,","cited_arxiv_id":null,"evidence_quote":"Provides the two-qubit circuit and concurrence-based entanglement parameterization used to generate labeled training data."}],"review_version":2}