{"id":"11e51a59-798d-4fb5-b440-f19648a40b97","arxiv_id":"1908.02428","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"If A is normal and A and B are traceless with ||A||_F^2+||B||_F^2=1/d, then the two largest squared singular values of A⊗I+I⊗B are bounded by (3d-4)/d^2.","lead":"This paper proves a matrix inequality that is a step toward deciding whether certain noisy quantum states can be distilled, a long-standing open problem in quantum information. The result covers the case where one of the two matrices in A⊗I + I⊗B is normal, generalizing an earlier proof that required both to be normal.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 26's p=0 case misidentifies the non-normal block B1 as anti-symmetric; the omitted subcase breaks the reduction to Theorem 8 and leaves Theorem 10 unproved.","rationale":"The reader's weakest_assumption identifies precisely the point I would flag. The central claim is conditional on Lemma 26 covering all maximizers of the reduced optimization problem, and the p=0 branch is not reduced correctly. The error is not stylistic: the matrix B1 with p=0 is not anti-symmetric unless its diagonal vanishes, and it is not normal unless qz=0, so the invocation of the normal-normal Theorem 8 is invalid in the excluded subcase. No alternative argument in the manuscript covers p=0 with q,z nonzero; the later g(v) analysis explicitly assumes p,q != 0. The numerical evidence and the attainable example in (16)-(19) support the inequality itself but do not repair the proof. Therefore the reader's REJECT verdict is appropriate, and I would not change it. I also note the reader's separate concern about Lemma 21, but the p=0 omission in Lemma 26 is the most load-bearing single defect.","tokens_in":23610,"tokens_out":25987,"duration_ms":275916,"concrete_test":"Set d=5 and write the KKT system for the optimization problem (197) with the additional constraint p=0 and q z != 0, i.e. b12=-b21 with b11-b22 and b12 both nonzero, keeping the trace and norm constraints. Symbolically or with high-precision numerics, determine whether a feasible stationary point exists in this submanifold. If such a point exists, evaluate h=σ1^2+σ2^2 there and compare with (3d-4)/d^2 = 11/25; this distinguishes an omitted proof case from an actual counterexample to Conjecture 5. If no stationary point exists, the gap is still present unless the proof shows that the maximizer cannot lie there.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Appendix C, Lemma 26 is the final reduction proving Lemma 20 for B in P, and Lemma 21 and Theorem 18 depend on it. After the change of variables b12=p+q, b21=p-q, the proof of case (1), pq=0, argues that p=0 gives b12=-b21 and hence 'B is anti-symmetric, and thus normal', so Theorem 8 applies. This is false. For the 2x2 block B1=[[w+z,q],[-q,w-z]] with q=b12, anti-symmetry would require w=z=0, and normality fails exactly when q z != 0: B1B1^T - B1^T B1 has off-diagonal entries -4qz. The manuscript gives no argument that a maximizer of the constrained problem (197) with p=0 must satisfy q=0 or z=0. Consequently the subcase p=0, q,z both nonzero is not covered by Theorem 8 and is not handled elsewhere in the proof. Since Lemma 26's bound is the bridge from the reduced form (169)-(171) to the general B in P, Theorem 10 is not established by the presented argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims to prove Theorem 10, namely that Conjecture 5 holds when one of the two matrices A, B in C^{d×d} (d > 4) is normal and the other is arbitrary: for traceless A, B with ||A||_F^2 + ||B||_F^2 = 1/d, the matrix X = A⊗I + I⊗B satisfies σ1^2(X) + σ2^2(X) ≤ (3d-4)/d^2. The proof strategy is to (i) reduce, via local unitary similarity and a real/imaginary splitting, to the case of real diagonal A; (ii) prove the real version (Theorem 18) by splitting into the case where the two largest singular values lie in one block (Lemma 19) and the case where they come from two different blocks (Lemma 21); (iii) reduce general B to block-diagonal form B ∈ P (Lemma 20, with Lemmas 23–27 in Appendix C); and (iv) solve the resulting low-dimensional optimization problem in Lemma 26, using the normal–normal theorem of Pankowski et al. (Theorem 8) as the base case. An explicit example attaining the bound is given in Eqs. (16)–(18).","tokens_in":23852,"tokens_out":40759,"duration_ms":364387,"significance":"If Theorem 10 were correct, the paper would provide real progress on a long-standing open problem: Conjecture 5 generalizes Conjecture 4, which is equivalent to the two-copy undistillability of 4×4 Werner states, and the present result would extend the normal–normal theorem of Pankowski et al. to the much broader setting of one normal and one arbitrary matrix. The reduction architecture is well organized; the algebra up to the reduced optimization problem is mostly explicit and checkable; the attainability example (16)–(18) is correct; and the proof is non-circular, relying on an independent published base theorem rather than on the conjecture itself, with no fitted parameters. These strengths are genuine, but they are contingent: the two load-bearing gaps identified below mean the central theorem is not established by the present argument.","major_comments":[{"comment":"The proof of case (1) with p = 0 asserts that b12 = -b21 makes B anti-symmetric and hence normal, so that Theorem 8 applies. Under the change of variables (189)–(191), the 2×2 block is B1 = [[w+z, q], [-q, w-z]] with q = b12, and anti-symmetry would require w = z = 0, which is not established. A direct computation gives B1B1^T - B1^T B1 = [[0, -4qz], [-4qz, 0]], so B1 is normal if and only if qz = 0; the manuscript gives no argument that a maximizer of the constrained problem (197) with p = 0 must satisfy q = 0 or z = 0. Hence the subcase p = 0, q ≠ 0, z ≠ 0 is covered neither by Theorem 8 nor by any other argument in the paper. Since Lemma 26 is the bridge from the reduced form (169)–(171) to Lemma 20 (via Lemma 27), and since Lemma 21 and Theorem 18 depend on Lemma 20, Theorem 10 is not established by the presented proof.","section":"Appendix C, Lemma 26 (case (1), p = 0)"},{"comment":"The displayed objective h is not the exact value of σ1^2(X) + σ2^2(X) under the stated normalization. After fixing φ = e1 and ψ = cosθ e1 + sinθ e2 and assuming only b1j = 0 for j > 4 and b2j = 0 for j > 5, the expansion of φ^T Y1 φ + ψ^T Y2 ψ contains the additional terms (1+cos^2θ)b14^2, sin^2θ b25^2, and 2 sinθ cosθ b14 b24, none of which appears in (62)–(64). A stronger normalization using the remaining orthogonal freedom in span{e3,...,ed} could force b14 = b25 = 0 and thereby repair the formula, but that is not what the manuscript states. As written, b14 and b25 lie in the set B (Eq. (65)) that the proof declares 'not involved in the objective function h', and the β-rescaling argument that justifies setting those variables to zero depends on exactly that claim, which is false for the true objective. The reduction to the optimization problem (71) and the subsequent application of Lemma 16 therefore do not bound the actual objective.","section":"Section III-B, Lemma 21, Eqs. (62)–(64)"}],"minor_comments":[{"comment":"'LOOC' should be 'LOCC', and 'validness' in the abstract should be 'validity'.","section":"Definition 1"},{"comment":"In the statement, σ2(X) = σ2(a2I2 + B1) should presumably be σ1(a2I2 + B1) to match the setting of Lemma 20; the proof also writes the second block's Gram matrix as (a2I2 + B1)(a1I2 + B1)^T, where the second factor should be (a2I2 + B1)(a2I2 + B1)^T.","section":"Appendix C, Lemma 26"},{"comment":"In case (1), the bullet for q = 0 contains a typo: it states p = (b12 - b21)/2 = 0, but it should read q = (b12 - b21)/2 = 0; the conclusion b12 = b21 is correct.","section":"Appendix C, Lemma 26"},{"comment":"The conclusion refers to 'Corollaries 18 and 13'; Corollary 11 and Corollary 13 appear to be meant.","section":"Conclusion"},{"comment":"In the proof of Lemma 16, the weights are denoted ω_i in the statement but w_i in the proof, and then the ratio ξ_i/ω_i is used; the notation should be made consistent.","section":"Appendix A"},{"comment":"There are numerous typographical slips ('Morevoer', 'matric', 'defied', 'eay to see', 'Conjecture's') that should be corrected in a revision.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"My reading agrees with the stress-test note: the p = 0 subcase of Lemma 26 is genuinely uncovered, and the asserted anti-symmetry of B in that subcase is false and load-bearing. The formula issue in Lemma 21 is a second, independent gap in the proof of the main theorem. The result may well be true and the gaps may be repairable, but as submitted the central claim is not proved, so I recommend rejection. I found no circularity or citation-related problems; the external base theorem is used appropriately."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine attempt at a real open case, but the proof has a gap at a load-bearing point. The claimed result is new—no one had shown the inequality for one normal, one arbitrary matrix—and the setup is honest: reduce to real diagonal A, split same-block vs different-block cases, handle B in P through optimization, then extend. The extremal example and the use of the Pankowski normal-normal theorem are appropriate. Credit where due: this is not a crank paper, and the high-level strategy is reasonable.\n\nThe problem is in Lemma 26, Appendix C, in the pq=0 case. With p=0, b12=q and b21=-q, so the 2x2 block is [[w+z,q],[-q,w-z]]. The paper says B is anti-symmetric, hence normal, and invokes Theorem 8. That is false unless w=z=0. Normality for this block fails when qz≠0, and the proof never rules out a maximizer with p=0 and q,z both nonzero. Since Lemma 26 is the bridge from the reduced form (169)-(171) to the general B in P, the gap is load-bearing. The stress-test note is right, and the reader's rejection is proportionate.\n\nThere is also a smaller inconsistency in Lemma 21: the support reduction says b1j=0 for j>4 and b2j=0 for j>5, but the objective formula omits b14 terms, suggesting the intended support differs from the stated one. That may be repairable, but as written it is confusing.\n\nThe bottom line: I would not accept Theorem 10 as proved. The underlying inequality may still be true, and the error might be fixable—the reduction structure is plausible—but the presented argument does not close the case. That said, this is exactly the kind of paper that deserves a serious referee: it addresses a long-standing open problem, and the gap is specific and checkable rather than a vague handwave. I would send it to review with instructions to focus the referee on the pq=0 subcase. For my own work, I would not cite it as proof yet, but I would keep an eye on a revised version.","headline":"A genuine step toward Conjecture 5, but the main proof has a load-bearing gap in Lemma 26: the p=0 case misidentifies a non-normal 2x2 block as antisymmetric, so Theorem 10 is not established as written.","tokens_in":24386,"tokens_out":2641,"would_cite":false,"duration_ms":26441,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A42","15A45","81P40","81P45"],"pacs":["03.67.-a","03.67.Mn"],"model":"deepseek-v4-flash","headline":"The paper proves Conjecture 5, the top-two singular-value bound for $X=A\\otimes I+I\\otimes B$, whenever at least one of the traceless matrices $A,B$ is normal, and gives sharp examples attaining the bound.","keywords":["entanglement distillation","Werner states","bound entanglement","matrix inequality","singular values","normal matrices","Frobenius norm","quantum information theory"],"falsifier":"For $d=5$, evaluate the reduced six-variable objective $h(x,y,z,w,p,q)$ from equation (199) over the feasible set (196) restricted to $p=0$ and $z\\neq 0$. If the maximum exceeds $11/25$, the claimed bound fails in the excluded configuration; if it does not, the gap is a missing argument rather than a counterexample.","tokens_in":23384,"feed_emoji":"⚛️","tokens_out":14231,"duration_ms":132673,"temperature":0.7,"pith_summary":"This paper targets a matrix inequality that controls whether certain entangled states can be distilled. The conjecture is that for traceless $d\\times d$ matrices $A,B$ with $\\|A\\|_F^2+\\|B\\|_F^2=1/d$, the largest two singular values of $X=A\\otimes I+I\\otimes B$ satisfy $\\sigma_1^2(X)+\\sigma_2^2(X)\\le (3d-4)/d^2$ for $d>4$; settling it would establish two-copy undistillability of the one-copy-undistillable $4\\times 4$ Werner states (the standard rotationally symmetric family of bipartite states). The paper proves the inequality whenever at least one of $A,B$ is normal, leaving only the case where both matrices are non-normal. It also exhibits diagonal matrices that attain the bound, so the constant is sharp. The proof diagonalizes the normal matrix, splits $X$ into blocks, and reduces the hard case to a finite-parameter optimization.","feed_headline":"Singular-value bound proven for one normal matrix","feed_subtitle":"The remaining open case is a pair of non-normal matrices, one step from the Werner-state two-copy test.","key_machinery":"The central object is $X=A\\otimes I+I\\otimes B$ together with the sum $\\sigma_1^2(X)+\\sigma_2^2(X)$ of the squares of its two largest singular values. A matrix is normal when it commutes with its adjoint, equivalently when it is unitarily diagonalizable. The argument uses local unitary similarity to make the normal matrix diagonal, which turns $X$ into a direct sum $\\oplus_{i=1}^d(a_iI+B)$; the singular values of $X$ are then the singular values of these blocks. A recurring reduction is to the class $\\mathcal{P}$ of matrices locally unitarily similar to a direct sum of $1\\times 1$ and $2\\times 2$ blocks, with the claim that the extremal case can be assumed to lie in $\\mathcal{P}$. Two optimization lemmas do the heavy lifting: Lemma 16 shows that in a degree-two homogeneous maximization with extra squared variables under a quadratic constraint, the maximum is either $\\eta r$ or occurs when the extra variables vanish; Lemma 17 shows that a scalar-sum constraint forces auxiliary variables to be equal at the optimum. The real case splits according to whether the two largest singular values come from the same block or from two different blocks, and the complex case follows by applying the real result to $\\operatorname{Re}(X)$ and $\\operatorname{Im}(X)$ separately.","core_discovery":"On its own terms, the paper's central result is Theorem 10: for $d>4$ and traceless $A,B\\in \\mathbb{C}^{d\\times d}$ with $\\|A\\|_F^2+\\|B\\|_F^2=1/d$, if at least one of $A$ and $B$ is normal, then $\\sigma_1^2(X)+\\sigma_2^2(X)\\le (3d-4)/d^2$ for $X=A\\otimes I+I\\otimes B$. The proof begins by using local unitary similarity to assume the normal matrix is diagonal, so $X$ is a direct sum of $d$ blocks $a_iI+B$; it solves the real-matrix case first and then lifts the result to complex matrices by splitting $X$ into real and imaginary parts. As a corollary, the same bound holds in the scale-invariant form $\\sigma_1^2(X)+\\sigma_2^2(X)\\le \\frac{3d-4}{d}(\\|A\\|_F^2+\\|B\\|_F^2)$, and Theorem 12 shows the constant is optimal by constructing explicit diagonal $A,B$ that attain equality.","pith_inferences":["A concrete next step would be to test the excluded configuration numerically: search the feasible set with $p=0$ and $z\\neq 0$ for local maxima of the objective in Lemma 26. Finding none would suggest the exceptional case is not extremal, while finding one would pinpoint a real gap in the reduction.","The equality pattern in Lemma 17, where auxiliary variables coalesce at the optimum, hints at a wider extremal principle: for the fully non-normal problem, maximizers might concentrate the spectrum into repeated diagonal blocks, reducing the remaining open case to a finite-dimensional family.","In the language of the distillability problem, the remaining step is narrow: if the fully non-normal case also satisfies the inequality, the one-copy-undistillable Werner states with non-positive partial transpose in dimension $4$ are two-undistillable, removing a long-standing candidate for NPT bound entanglement."],"forward_implications":["Conjecture 5 is now known whenever at least one of $A,B$ is normal; the remaining open instance is exactly the pair of non-normal matrices.","Since Conjecture 4 for $d=4$ is a special case of Conjecture 5, the result moves the two-copy undistillability of the one-copy-undistillable $4\\times 4$ Werner states closer to resolution.","The explicit diagonal pair in equations (16)--(18) attains $(3d-4)/d^2$, so the constant cannot be improved in this setting.","Corollary 11 gives the scale-invariant bound $\\sigma_1^2+\\sigma_2^2\\le \\frac{3d-4}{d}(\\|A\\|_F^2+\\|B\\|_F^2)$ for all traceless pairs with one normal matrix, independent of the normalization."],"supporting_citations":[{"why":"Proved Conjecture 5 in the normal-normal case and connected the inequality to two-copy undistillability of Werner states; the present proof uses it as the base theorem for the normal block.","marker":"[21]"},{"why":"Provided the equivalence between the d=4 and d>4 formulations and the partial d=4 cases under special matrix shapes; used to justify reductions to locally unitarily similar and diagonal normal forms.","marker":"[38]"},{"why":"Showed that every NPT bipartite state can be locally converted to a Werner state and classified the one-distillable range, making the Werner-state case the target.","marker":"[24]"},{"why":"Established that Werner states with $1/d<\\alpha\\le 1/2$ are NPT and one-copy undistillable, delineating the regime where the two-copy question is open.","marker":"[25]"},{"why":"Supplied the KKT first-order conditions used to derive necessary conditions for all candidate maxima in the optimization reductions.","marker":"[41]"}],"fun_headline_variants":["One normal matrix secures the singular-value bound","Normal condition unlocks bound for tensor sum","Progress on distillability: bound proven for one normal matrix","Matrix inequality proven when A or B is normal","Key step: singular-value bound for normal matrix case"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of Lemma 26 assumes that when $p=0$ (meaning $b_{12}=-b_{21}$) the matrix $B$ is anti-symmetric and therefore normal, so the known normal-normal theorem applies; but a $2\\times 2$ block with nonzero off-diagonal entries and unequal diagonal entries is not normal, and the paper gives no argument excluding a maximizer of that shape.","fun_headline_variants_meta":{"raw":{"variants":["One normal matrix secures the singular-value bound","Normal condition unlocks bound for tensor sum","Progress on distillability: bound proven for one normal matrix","Matrix inequality proven when A or B is normal","Key step: singular-value bound for normal matrix case"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000922,"raw_usage":{"total_tokens":3996,"prompt_tokens":1030,"completion_tokens":2966,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":2894}},"tokens_in":646,"tokens_out":2966,"duration_ms":46430,"temperature":1.0,"reasoning_tokens":2894,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:47:10.757102+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For $d=5$, evaluate the reduced six-variable objective $h(x,y,z,w,p,q)$ from equation (199) over the feasible set (196) restricted to $p=0$ and $z\\neq 0$. If the maximum exceeds $11/25$, the claimed bound fails in the excluded configuration; if it does not, the gap is a missing argument rather than a counterexample.","supporting_citations":[{"cited_title":"On a matrix inequality related to the distillability problem,","cited_arxiv_id":null,"evidence_quote":"Provided the equivalence between the d=4 and d>4 formulations and the partial d=4 cases under special matrix shapes; used to justify reductions to locally unitarily similar and diagonal normal forms."},{"cited_title":"Minima of functions of several variables with inequalities as side conditions,","cited_arxiv_id":null,"evidence_quote":"Supplied the KKT first-order conditions used to derive necessary conditions for all candidate maxima in the optimization reductions."}],"review_version":1}