{"id":"a6a50c84-6b55-4315-8af3-729b7b22f5be","arxiv_id":"2504.18163","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The authors use a support vector machine to construct entanglement witnesses for small multipartite systems, claiming they are optimal and can detect bound entangled states.","lead":"This paper claims that a support vector machine can construct optimal entanglement witnesses for two, three, and four qubit states. The method is demonstrated on Bell-diagonal and GHZ states with white noise, but the optimality claims are not backed by rigorous proof or reproducible code.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SVM outputs are never validated against the full separable set; a finite-sample separator may fail Definition 2, so the claimed optimality and complete tangency are ungrounded.","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: the SVM decision boundary is assumed to be an EW without verification on the full separable set. This is the right concern because it undermines both the witness validity and the optimality claim. The paper's own proofs only exhibit selected tangent product states; they do not certify nonnegativity over the convex separable region. The SVM objective has no mechanism to enforce constraints on unseen separable states, so finite-sample separation is insufficient. Given that no code or data is provided, the numerical results cannot be independently reproduced or audited. Our proposed SDP check would settle the validity of the two-qubit witness definitively. Since the reader already rejected the paper and our analysis supports that rejection, the verdict should remain unchanged.","tokens_in":17703,"tokens_out":4034,"duration_ms":47657,"concrete_test":"For the two-qubit witness W1 in Eq. (13), solve the semidefinite program: minimize Tr(W1 ρ) subject to ρ ≥ 0, Tr(ρ) = 1, and ρ^{T_B} ≥ 0. Since PPT is necessary and sufficient for separability in 2⊗2, a negative optimum proves W1 is not an EW. If the optimum is nonnegative, repeat the check against a dense parameterization of pure product states for the three- and four-qubit witnesses in Eqs. (19) and (25), minimizing Tr(W |ν⟩⟨ν|) over the product-state angles.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the SVM-derived hyperplane is an optimal entanglement witness. Section 2 identifies the SVM decision function of Eqs. (5)-(9) with an EW, but the optimization in Eq. (4) only constrains the training points. A valid EW must satisfy Tr(Wρ_sep) ≥ 0 for every separable state, not just the sampled ones. No such full-set verification is provided. The optimality 'proofs' in Section 3 check Tr(W|ν⟩⟨ν|)=0 only on a handful of product states (Tables 1-3); at best this shows local contact, not that W is an EW, nor that it is tangent to the whole separable region. The subsequent argument that no positive operator can be subtracted because ⟨ν_i|ω⟩=0 for a few states is also incomplete: it does not establish the required Lewenstein criterion for all positive operators. Thus the abstract's assertion that 'when the algorithm succeeds, the EWs are optimal and are completely tangent to the separable region' is unsupported: a soft-margin SVM can return a hyperplane that is not a witness at all.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a support vector machine (SVM) method to construct entanglement witnesses (EWs) for two-, three-, and four-qubit systems. The authors train an SVM on separable and entangled Werner states (and on a three-qubit PPT entangled family), map the decision hyperplane to an operator W via a Pauli-basis expansion, and claim that the resulting EWs are optimal and completely tangent to the separable region. They also report a non-decomposable witness that detects a PPT entangled state.","tokens_in":18008,"tokens_out":10515,"duration_ms":97483,"significance":"If the central claims were correct, the paper would offer a practical, scalable numerical route to optimal EWs for multipartite systems, including PPT-entangled detection, which is of genuine interest to the quantum information community. The paper usefully provides an explicit algorithmic pipeline (Algorithm 1) and explicit numerical witnesses. However, the key assertions—that the SVM hyperplane is a valid EW and that it is optimal—are neither rigorously proven nor, in the two-qubit case, consistent with the reported expectation values. The concrete failure of the two-qubit witness to detect all entangled Werner states undermines the claimed central result.","major_comments":[{"comment":"The two-qubit witness W1 fails to detect all entangled Werner states and therefore cannot be optimal. From the entries of W1 in Eq. (13), using ρ_W(p) = p|φ00><φ00| + (1-p)/4 I, one obtains Tr(W1ρ_W(p)) = 0.25 − 0.703p; this gives a detection threshold p ≈ 0.355, whereas ρ_W(p) is entangled for p > 1/3. The paper's statement that W1 'recognizes all the entangled states ... for all p's' is thus contradicted by its own numerical witness.","section":"§3.1, Eqs. (12)–(13)"},{"comment":"The optimality proof is invalid. The condition Tr(W1|ν><ν|)=0 does not place |ν> in the kernel of W1, since W1 is an indefinite operator; the Lewenstein criterion for optimality requires product vectors |ν> with W1|ν>=0. Moreover, the argument that no |ω> is orthogonal to the four listed |ν_i> does not rule out all positive operators P, and if the four |ν_i> were indeed in the kernel and spanned the full space, they would force W1=0, which is false. The support of the claim 'we have proved that all the EW's are completely tangent ... and are therefore optimal' is therefore missing.","section":"§3.1, Table 1 and Eq. (17)"},{"comment":"No evidence is provided that any of the SVM-derived operators is a valid entanglement witness. The SVM optimization in Eq. (4) only enforces the margin constraints on the finite training set; a valid EW must satisfy Tr(Wρ_sep) ≥ 0 for every separable state ρ_sep. The manuscript does not verify this condition for W1, WGHZ, the PPTES witness, or the four-qubit witness, so the identification of the SVM hyperplane with an EW (Eqs. (5)–(9)) is unsubstantiated.","section":"§2, Definition 2 and §3"},{"comment":"The non-decomposable witness claim rests on a single numerical example (a=0.3525, b=0.3196, c=0.81642, Tr(Wρ_ppt) = −0.0129). To establish that W is a non-decomposable EW, the authors must show both that W is a valid witness (nonnegative expectation on all separable states) and that it detects a PPT entangled state; the latter is demonstrated, but the former is not, and the single example does not constitute a proof.","section":"§3.2.1, Eqs. (22)–(23)"}],"minor_comments":[{"comment":"The tensor product factor should read σ_{i1}⊗σ_{i2}...⊗σ_{iN}, not σ_{i1}⊗σ_{i1}...⊗σ_{iN}.","section":"Eq. (5)"},{"comment":"The text cites 'matrix 18' for the two-qubit Werner state; the correct reference is Eq. (12).","section":"§3.1"},{"comment":"Rows 2 and 4 are identical, so only three distinct tangent states are listed despite the text referring to four product pure states.","section":"Table 1"},{"comment":"The bias term b is defined in Eqs. (4) and (8) but is absent from the classifier in Eq. (9); the role of the constant term in the EW expectation value should be clarified.","section":"§2, Eqs. (4), (8), (9)"},{"comment":"The Werner parameter is restricted to 0.6 ≤ p ≤ 1 without explanation, while the two-qubit case used 0 ≤ p ≤ 1; the choice of training ranges should be justified.","section":"Eq. (18)"},{"comment":"There are numerous typographical errors, including 'seperable', 'classifid', 'matrice', and 'Hillbert', and the abstract's 'We drive and implement' should be 'We derive and implement'.","section":"General"}],"recommendation":"reject","confidential_remarks":"The reader's assessment aligns with my own reading. The central technical claims are unsupported: the two-qubit witness provably fails to detect all entangled Werner states, the optimality proof conflates zero expectation with kernel membership, and the witness condition is never verified on the full separable set. These are load-bearing errors that cannot be fixed by local revisions within the scope of the present manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Mahdian and Mousavi claim SVM-generated hyperplanes are optimal entanglement witnesses for multipartite systems. The central assertion doesn't hold up on the evidence. The decision boundary is trained on finite samples; no verification on the full separable set, so the resulting operator may not satisfy Definition 2. The optimality 'proofs' only show zero expectation on a handful of product states (the SVM support vectors), which is local contact, not the Lewenstein criterion. The two-qubit witness itself fails: Tr(W1ρ) = -0.703p + 0.25 gives threshold p ≈ 0.355, missing entangled states with 1/3 < p < 0.355. So the abstract's 'when the algorithm succeeds, the EWs are optimal' is unsupported.\n\nThat said, the paper does something useful. It takes a concrete supervised-learning approach, gives explicit 4x4, 8x8, and 16x16 Hermitian operators for two, three, and four qubits, and shows an operator that detects the Acín PPT state — that's a stable, checkable result. The idea of extracting a witness from the SVM hyperplane in local Pauli coordinates is a reasonable one, and the numerical work is transparent enough to be reproduced with effort. The self-citation pattern is not unusual.\n\nThe soft spots are not minor. First, a finite-sample separator may cut through the separable set; a valid witness needs nonnegative expectation on all separable states. Second, the circularity: the tangent states are the support vectors, so the optimality argument goes in a circle. Third, no comparison with known analytical witnesses for the Werner families, making it hard to see what is gained. Fourth, no code or data, so the numbers cannot be independently checked.\n\nWho is this for? Someone looking for a template of how ML can be used to construct witnesses, and a reminder that finite-sample classifiers need out-of-sample certification. The paper is not ready for publication as is. I would not send it to a top journal. But if an editor thinks a referee can help the authors see what is missing, a single round of major revision might be defensible. My own verdict: do not accept; the central claim is not established.","headline":"The central claim of optimal SVM entanglement witnesses is unproven: the decision boundary is only validated on finite samples, so the operators need not be witnesses at all.","tokens_in":18404,"tokens_out":2551,"would_cite":false,"duration_ms":26065,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P40","68T05","81P68"],"pacs":[],"model":"deepseek-v4-flash","headline":"SVM-trained hyperplanes become optimal entanglement witnesses","keywords":["entanglement witness","support vector machine","optimal entanglement witness","non-decomposable entanglement witness","Bell-diagonal states","GHZ states","PPT entangled states","machine learning entanglement detection"],"falsifier":"For each constructed witness, compute the minimum of $\\mathrm{Tr}(W\\rho)$ over all fully separable states: for two-qubit Bell-diagonal states use the exact octahedron separability condition, and for three- and four-qubit states use dense sampling of product states. A single separable state with negative expectation would show that the SVM hyperplane is not an entanglement witness.","tokens_in":17543,"feed_emoji":"⚛️","tokens_out":8784,"duration_ms":85714,"temperature":0.7,"pith_summary":"The paper claims that a support-vector-machine classifier, trained on the local Pauli expectation values of known separable and entangled states, produces a Hermitian operator that is an entanglement witness: nonnegative on every separable state and negative on the entangled states it targets. The authors construct such witnesses for two-, three-, and four-qubit Bell-diagonal and Werner/GHZ mixed states, and argue that whenever the SVM training succeeds the resulting hyperplane is optimal, being completely tangent to the convex set of separable states. They further report a three-qubit non-decomposable witness that detects a positive-partial-transpose entangled state, a type that the PPT criterion alone cannot identify. If the claim holds, the method is a numerical recipe for optimal entanglement witnesses that extends to N-qubit systems without requiring an analytic construction.","feed_headline":"SVM-trained hyperplanes become optimal entanglement witnesses","feed_subtitle":"For two-, three-, and four-qubit states, the boundary touches the separable region and can detect bound entanglement.","key_machinery":"The load-bearing object is the maximum-margin hyperplane of a soft-margin SVM in the space of local Pauli expectation values. The optimization minimizes $\\frac{1}{2}\\|w\\|^2 + C\\sum_i \\xi_i$ subject to $y_i(w^T x_i + b)\\ge 1-\\xi_i$; the resulting coefficients $c_{\\vec i}$ define $W$ by Eq. (5), and the bias $b$ places the boundary $\\langle W\\rangle=0$. Max-margin geometry is what makes the boundary tangent to the separable region, and the paper's optimality certificate is the vanishing of $\\mathrm{Tr}(W|\\nu\\rangle\\langle\\nu|)$ on product states plus the claim that no positive operator lies in the kernel of $W$.","core_discovery":"On its own terms, the central discovery is that the SVM decision boundary can be read directly as an entanglement witness. With features $x_{\\vec i} = \\mathrm{Tr}(\\sigma_{i_1}\\otimes\\cdots\\otimes\\sigma_{i_N}\\rho)$ and decision function $y = \\sum_{\\vec i} c_{\\vec i} x_{\\vec i} - b$, the weight vector becomes the witness $W = \\sum_{\\vec i} c_{\\vec i}\\sigma_{i_1}\\otimes\\cdots\\otimes\\sigma_{i_N}$; the separable class has $y\\ge 0$ and the entangled class $y<0$. For the Werner families studied, the trained witnesses are reported to be completely tangent to the separable region: explicit product states $|\\nu\\rangle$ satisfy $\\mathrm{Tr}(W|\\nu\\rangle\\langle\\nu|)=0$, and the absence of any positive operator that could be subtracted from $W$ is taken as proof of optimality. In the three-qubit case, a witness built for the PPT-entangled density matrix of the paper's reference [34] has negative expectation on that state, which the authors use to conclude that the witness is non-decomposable.","pith_inferences":["A strict check of the claim would verify nonnegativity of $\\mathrm{Tr}(W\\rho)$ over the entire separable set, not just the training sample; the paper's tangency points are evidence, not a full certificate, for states far from those points.","For other state families, the same recipe should be expected to work only if the training sample densely covers the separable-entangled boundary; sparse or biased samples could produce a hyperplane that cuts into the separable region.","Because the witness is a sum of local Pauli terms, the trained coefficients translate directly into a measurement protocol, so the method is testable in the laboratory.","The approach could be benchmarked against exact separability criteria where they exist, for instance the octahedron condition for two-qubit Bell-diagonal states, to measure how close numerical witnesses come to optimality."],"forward_implications":["For the two-, three-, and four-qubit Werner/GHZ families studied, the trained SVM yields a single operator that detects every entangled state in the family above its threshold $p$.","The same local-measurement construction is claimed to generalize to $N$-qubit systems with equal local dimensions, giving numerical optimal witnesses without analytic guesses.","Non-decomposable witnesses produced this way can detect PPT-entangled (bound) states that escape the PPT criterion.","Whenever the SVM succeeds, the witnesses are automatically optimal in the standard sense, because the margin-maximizing boundary is completely tangent to the separable region."],"supporting_citations":[{"why":"Defines the optimality criterion, namely that no positive operator can be subtracted while keeping a valid witness, which the paper uses to prove optimality.","marker":"[14]"},{"why":"Supplies the three-qubit PPT-entangled density matrix whose detection marks the constructed witness as non-decomposable.","marker":"[34]"},{"why":"Provides the PPT criterion used to identify states whose entanglement requires non-decomposable witnesses.","marker":"[11]"},{"why":"Establishes the necessary-and-sufficient separability framework behind the geometric witness-as-hyperplane picture.","marker":"[12]"},{"why":"Previous machine-learning separability-entanglement classifier that this SVM approach extends.","marker":"[29]"},{"why":"Provides the soft-margin SVM formulation whose weight vector becomes the witness coefficients.","marker":"[31]"},{"why":"Supplies the SVM optimization background for the hyperplane decision boundary.","marker":"[32]"}],"fun_headline_variants":["SVM decision boundaries become optimal entanglement witnesses","SVM learns optimal witnesses tangent to separable region","SVM hyperplane witnesses detect bound entanglement","Multipartite entanglement witnesses from SVM hyperplanes","SVM-based witnesses optimal and detect bound entangled states"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole construction assumes that a hyperplane which fits the training data also keeps nonnegative expectation on every separable state the algorithm never saw, and the paper does not verify this on the full separable set.","fun_headline_variants_meta":{"raw":{"variants":["SVM decision boundaries become optimal entanglement witnesses","SVM learns optimal witnesses tangent to separable region","SVM hyperplane witnesses detect bound entanglement","Multipartite entanglement witnesses from SVM hyperplanes","SVM-based witnesses optimal and detect bound entangled states"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000691,"raw_usage":{"total_tokens":3120,"prompt_tokens":928,"completion_tokens":2192,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":2132}},"tokens_in":544,"tokens_out":2192,"duration_ms":14243,"temperature":1.0,"reasoning_tokens":2132,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:24:03.802382+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For each constructed witness, compute the minimum of $\\mathrm{Tr}(W\\rho)$ over all fully separable states: for two-qubit Bell-diagonal states use the exact octahedron separability condition, and for three- and four-qubit states use dense sampling of product states. A single separable state with negative expectation would show that the SVM hyperplane is not an entanglement witness.","supporting_citations":[{"cited_title":"Optimization of entanglement witnesses,","cited_arxiv_id":null,"evidence_quote":"Defines the optimality criterion, namely that no positive operator can be subtracted while keeping a valid witness, which the paper uses to prove optimality."},{"cited_title":"Classification of mixed three-qubit states,","cited_arxiv_id":null,"evidence_quote":"Supplies the three-qubit PPT-entangled density matrix whose detection marks the constructed witness as non-decomposable."},{"cited_title":"Separability criterion for density matrices,","cited_arxiv_id":null,"evidence_quote":"Provides the PPT criterion used to identify states whose entanglement requires non-decomposable witnesses."},{"cited_title":"Separability of mixed states: necessaryandsufficientconditions,","cited_arxiv_id":null,"evidence_quote":"Establishes the necessary-and-sufficient separability framework behind the geometric witness-as-hyperplane picture."},{"cited_title":"Separability-entanglementclassifierviamachinelearning,","cited_arxiv_id":null,"evidence_quote":"Previous machine-learning separability-entanglement classifier that this SVM approach extends."},{"cited_title":"Support vector machine,","cited_arxiv_id":null,"evidence_quote":"Provides the soft-margin SVM formulation whose weight vector becomes the witness coefficients."},{"cited_title":"Cristianini and J","cited_arxiv_id":null,"evidence_quote":"Supplies the SVM optimization background for the hyperplane decision boundary."}],"review_version":1}