{"id":"a8b708e3-8da8-4ee2-913f-37728ea0341f","arxiv_id":"2506.15326","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Exact and near-exact 1-, 2-, and infinity-norm bounds, condition-number bounds, and eigenvalue and numerical-range regions are derived for Bai's Wasserstein-1 metric matrices in one and two dimensions.","lead":"The paper derives tighter norm and condition-number bounds for the Wasserstein-1 metric matrix, a structured matrix used in optimal-transport computations. These bounds sharpen earlier estimates by Bai, but many are routine consequences of the matrix's simple Toeplitz form.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1(v)-(x), the paper's main bounds, are never proved; part (vii) is false for n=1, and the 'tight' upper bound in (v) is not the exact row-sum value.","rationale":"The central assertions of the paper are the bounds in Theorem 3.1(v)-(x) and their two-dimensional analogues. My reading is that these inequalities are mostly true for n≥2: the 1/∞ norms and the condition number can be bounded using the row-sum formula together with Bai's inverse bound, and numerical checks support them. However, the paper as written does not prove these parts of the theorem: the proof stops after part (iv), and the 'tight' upper bound is not the natural row-sum bound, which is both simpler and tighter. More seriously, Theorem 3.1(vii) is false at n=1 for every λ∈(0,1); this shows the statement lacks a necessary hypothesis. These issues are fixable, so the result should be CONDITIONAL on supplying a complete proof and an explicit n≥2 hypothesis. I agree with the reader's weakest point that the inverse-norm bounds are under-supported, and I also flag the false edge case and the overclaim of tightness.","tokens_in":13873,"tokens_out":20962,"duration_ms":181801,"concrete_test":"Check the n=1 case: compute Q=[1] and evaluate Theorem 3.1(vii); the inequality fails for all λ∈(0,1), so the theorem statement requires n≥2. Then, for a fixed instance such as n=5, λ=0.5, compute all row sums of Q and compare ||Q||∞ (=2.5) with the bound in (v) (=2.875). This would show that the claimed 'tight' bound is not the exact row-sum bound and would force either a corrected statement or an explicit proof of (v).","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Theorem 3.1(v)-(x) are the paper's main contribution, yet the proof stops after part (iv). No argument is supplied for the 1/∞-norm upper bound, the inverse-norm bounds, or the condition-number bounds; parts (vii)-(x) silently rely on the factorization (4) from Bai [3] and Bai's inverse bound, without deriving or isolating them. This is not merely cosmetic: as stated, part (vii) is false. For n=1, Q=[1], so ||Q^{-1}||_1=1, while the claimed lower bound is (1−λ)/((1+λ)(1−λ)^2)=1/(1−λ^2)>1 for every λ∈(0,1). The theorem needs an explicit n≥2 hypothesis and a worked proof. Additionally, the 'tight' upper bound in (v) is not tight in the natural sense: the exact ∞-norm is the largest row sum, which for n=5, λ=0.5 is 2.5, while the bound gives 1+2λ(1−λ^{n-1})/(1−λ)=2.875. The paper never computes the obvious row-sum bound, so the claimed sharpness is unsubstantiated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the one- and two-dimensional Wasserstein-1 metric matrices Q defined by q_ij = λ^{|i-j|} (and their Kronecker-product analogues). It claims lower and upper bounds for the 1-, 2-, and ∞-norms of Q, Q^{-1}, and the associated condition numbers, asserting that these bounds are much sharper than those of Bai [2,3]. It also gives eigenvalue and numerical-range inclusion regions, determinant formulas, Hadamard-product decompositions, and numerical-radius-based condition-number estimates. The main results are Theorem 3.1 (one-dimensional case) and Theorem 3.9 (two-dimensional case), with supporting computations in Theorems 3.3 and 3.7 and numerical tables.","tokens_in":14088,"tokens_out":11439,"duration_ms":104225,"significance":"If the stated bounds are correct, they would provide a useful refinement of Bai's norm and conditioning estimates for a matrix class that arises in entropy-regularized optimal transport. A strength of the paper is that the bounds are explicit functions of the regularization parameter λ and the dimension n, with no fitted parameters, and several identities (determinant, Hadamard-inverse norms) are exact. However, the significance is currently undermined by the absence of proofs for the main norm and condition-number bounds in Theorem 3.1(v)-(x), by a concrete counterexample to part (vii) in the n=1 case, and by the use of an invalid general numerical-radius identity. These issues must be corrected before the claims can be accepted; they appear fixable within the scope of the manuscript.","major_comments":[{"comment":"The proof of Theorem 3.1 stops after part (iv); parts (v)-(x), which are the paper's headline norm and condition-number bounds, are asserted without any proof. Since these bounds are the main contribution, the authors must supply a complete derivation for each part, or at least reduce each bound explicitly to the factorization (4)-(6) from Bai [3] with all hypotheses stated. The two-dimensional claims in Theorem 3.9(v)-(x) inherit this gap, since their proofs appeal directly to the corresponding unproved parts of Theorem 3.1.","section":"Theorem 3.1"},{"comment":"Theorem 3.1(vii) is false as stated for n=1. For n=1, Q=[1], so ||Q^{-1}||_1=1, while the claimed lower bound is (1-λ)/((1+λ)(1-λ)^2)=1/(1-λ^2)>1 for every λ∈(0,1). Parts (iv) and (vi) also fail for n=1: for Q=[1] the claimed numerical-range upper bound becomes 0 and the claimed 2-norm upper bound becomes 0. The theorem must include an explicit n≥2 hypothesis, and Theorem 3.9 must include an analogous n,m≥2 hypothesis.","section":"Theorem 3.1(vii)"},{"comment":"The proof of Theorem 3.1(ii) uses the identity ω(I+A)=1+ω(A), listed as property (P2) in Section 2. This identity is false in general (for example A=-I or A=[[0,1],[1,0]]). It is valid for A=L+L^T because this matrix is nonnegative and symmetric, but the manuscript does not state or justify that condition. The background property (P2) should be corrected to a qualified statement, and the proof of (ii) should mention the applicable hypothesis.","section":"Theorem 3.1(ii), Section 2 (P2)"},{"comment":"The upper bound in Theorem 3.1(v) is called 'tight' in the text preceding the theorem, but it is not tight in the natural sense. For n=5, λ=0.5, the exact ∞-norm of Q is the largest row sum, which equals 2.5, while the claimed bound gives 1+2λ(1-λ^{n-1})/(1-λ)=2.875. The exact row-sum value is easy to compute, so the paper should either give the exact maximum row sum or substantially soften the claim of tightness; as it stands, the title's 'tight bounds' claim is unsupported.","section":"Theorem 3.1(v)"}],"minor_comments":[{"comment":"Property (P2) should be replaced by a correct statement, for example that ω(I+A)=1+ω(A) holds when A is Hermitian and positive semidefinite, or more generally when A is nonnegative and symmetric; this prevents the reader from relying on a false general identity elsewhere.","section":"Section 2, property (P2)"},{"comment":"In the proof of Theorem 3.9, part (i) says 'Similar to (1)', which should read 'Similar to (i)'; this is a typographical error that should be corrected.","section":"Theorem 3.9 proof"},{"comment":"Reference [9] is incomplete: 'Application of the Wasserstein metric seismic signal, (2013), arXiv:1311.4581v1' lacks the authors' initials and a proper venue; it should be completed.","section":"Reference [9]"},{"comment":"There are several typographical issues, including 'Wassetrstein matrix' in the sentence before Theorem 3.3. A careful proofreading pass is needed.","section":"General"},{"comment":"The bound ||Q||_2 ≤ n from the Hadamard-product argument is much weaker than the bounds already given in Theorem 3.1(vi) and is not used later; consider removing it or noting its limited role.","section":"Theorem 3.3(i)"},{"comment":"The statement that the Cayley transform C(Q) can never be positive definite should explicitly mention the known criterion that the Cayley transform of a positive definite matrix is positive definite if and only if all eigenvalues lie in (0,1); the current wording leaves the reasoning implicit.","section":"Remark 3.5(2)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript contains a promising set of explicit bounds, but in its current form the main theorem is not proved (Theorem 3.1(v)-(x) are asserted without proof), and part (vii) is false for n=1. The use of the numerical-radius identity (P2) as a general fact is also a substantive correctness issue. The 'tight' claim in the title and text should be revised unless exact norms are provided; the existing upper bounds are still sharper than Bai's for large n, which may suffice for the paper's stated contribution. I would ask the authors to supply complete proofs, add the dimension hypotheses, and correct the background lemma. The issues appear repairable, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a real, if modest, contribution—sharper norm and conditioning bounds for Bai's Wasserstein-1 matrices—but the main theorem is only partly proved, and one part is false at n=1. I'd send it back for a revision, not reject it.\n\nThe genuinely new pieces are the explicit upper bound 1 + 2λ(1−λ^{n−1})/(1−λ) for ||Q||_1, which improves Bai's (1+λ)/(1−λ), and the corresponding condition-number bound. The Kronecker extension in Theorem 3.9 follows cleanly from known properties of ⊗, and the Hadamard-inverse results (Theorem 3.7) are a nice add. The numerical-range inclusions are also new and are proven in part.\n\nNow the soft spots. The proof of Theorem 3.1 stops after (iv). Parts (v)–(x) are asserted with 'similarly' or no proof at all. That is not cosmetic: (vii) is false when n=1. For Q=[1], ||Q^{-1}||_1 = 1, while the claimed lower bound is 1/(1−λ^2) > 1. The theorem needs an explicit n≥2 hypothesis, and the reader cannot verify (v)–(x) without derivations. Also, the bound in (v) is called tight, but the exact ∞-norm is the largest row sum, which for n=5, λ=0.5 is 2.5, whereas the bound gives 2.875. So it is a valid upper bound, but tight is an overstatement. The proof of (ii) invokes ω(I+A) = 1 + ω(A) as a general property; that is false. It happens to be valid for this A because L+L^T is nonnegative symmetric, but the paper doesn't say so. Finally, (vii)–(x) lean on Bai's factorization (4); that's legitimate, but the paper should state it explicitly as the starting point and prove or clearly cite it.\n\nThe mathematics that is proved looks correct. The determinant and Hadamard-inverse formulas are standard and check out. The references to Bai are appropriate, and the paper doesn't overfit.\n\nWho is this for? Matrix analysts and people implementing fast Sinkhorn-type algorithms for Wasserstein-1. It's elementary but not worthless. I'd send it to a referee, but I'd expect a substantial revision: full proofs for (v)–(x), an n≥2 assumption, and a corrected numerical-radius statement.","headline":"Sharper bounds for Wasserstein-1 matrices, but the main theorem is half-proved and one part is false at n=1.","tokens_in":14673,"tokens_out":9714,"would_cite":false,"duration_ms":83356,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A18","15A60","15B05","65F05","65H10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims sharply improved norm and condition-number bounds for Wasserstein-1 metric matrices, extending to Kronecker-product two-dimensional cases.","keywords":["Wasserstein metric matrix","matrix norms","condition number","symmetric Toeplitz matrix","Kronecker product","numerical range","Hadamard product","positive definite matrix"],"falsifier":"Compute the exact 1-norm of $Q$ for a fixed $n$, say $n=4$ and $\\lambda=0.5$, by summing row entries and comparing it to the claimed upper bound $1 + 2\\lambda(1-\\lambda^{n-1})/(1-\\lambda)$; if the bound is violated for any $\\lambda \\in (0,1)$, Theorem 3.1(v) fails. More directly, symbolically expand $(I-\\lambda N^T)Q(I-\\lambda N)$ for $n=2$ and $n=3$ and check whether it equals $\\hat{D}_\\lambda$; a single counterexample to this identity for a diagonal entry would overturn the inverse and condition-number bounds.","tokens_in":13617,"feed_emoji":"📐","tokens_out":8925,"duration_ms":76688,"temperature":0.7,"pith_summary":"The paper establishes tight lower and upper bounds on the 1-, 2-, and infinity-norms of Wasserstein-1 metric matrices, their inverses, and their condition numbers, for both the one-dimensional matrix $Q$ with entries $\\lambda^{|i-j|}$ and its two-dimensional Kronecker-product analogue. The key results show that the 1-norm (equivalently the infinity-norm) is bounded above by $1 + 2\\lambda(1-\\lambda^{n-1})/(1-\\lambda)$ and the 1-condition number by $(1+\\lambda)(1-\\lambda+2\\lambda(1-\\lambda^{n-1}))/(1-\\lambda)^2$, both sharper than previous estimates. It also locates eigenvalues and numerical ranges in explicit intervals and provides decompositions of $Q$ as a Hadamard product and via the Cayley transform. For the Hadamard inverse, exact norm formulas are obtained. If correct, these bounds give near-exact cost and conditioning estimates for the matrix-vector multiplications central to Sinkhorn-type optimal transport algorithms.","feed_headline":"New bounds tighten Wasserstein matrix norm estimates","feed_subtitle":"It gives near-exact 1-norm and condition-number estimates for one- and two-dimensional Wasserstein-1 matrices.","key_machinery":"The central object is the symmetric Toeplitz matrix $Q = (\\lambda^{|i-j|})$, written as $Q = I + L + L^T$. The argument rests on four mechanisms: (1) the factorization $(I - \\lambda N^T) Q (I - \\lambda N) = \\hat{D}_\\lambda$, where $N$ is the subdiagonal nilpotent shift and $\\hat{D}_\\lambda$ is diagonal with entries $1-\\lambda^2$ except the last entry $1$, which reduces inverse-norm and conditioning bounds to diagonal scaling; (2) Gershgorin disks and Perron-Frobenius theory for eigenvalue and spectral-radius bounds; (3) numerical radius inequalities, including the relation $\\omega(I+A) = 1+\\omega(A)$ applied to the nonnegative symmetric matrix $A = L+L^T$, to bound eigenvalues; and (4) the Hadamard-product representation $Q = A \\circ A^T$, combined with a known spectral-norm bound for Hadamard products, to estimate the 2-norm.","core_discovery":"For the one-dimensional Wasserstein-1 metric matrix $Q = (\\lambda^{|i-j|})$ with $0<\\lambda<1$, the paper claims that its 1-norm and infinity-norm coincide and satisfy $1/(1+\\lambda)^2 \\leq \\|Q\\|_1 = \\|Q\\|_\\infty \\leq 1 + 2\\lambda(1-\\lambda^{n-1})/(1-\\lambda)$, and that the condition number $\\kappa_1(Q)$ satisfies $(1-\\lambda)/((1+\\lambda)^3(1-\\lambda^n)^2) \\leq \\kappa_1(Q) \\leq (1+\\lambda)(1-\\lambda+2\\lambda(1-\\lambda^{n-1}))/(1-\\lambda)^2$. These upper bounds improve on the earlier estimates $(1+\\lambda)/(1-\\lambda)$ and $((1+\\lambda)/(1-\\lambda))^2$. The same pattern is extended to the two-dimensional Kronecker product $Q = Q^{[2]} \\otimes Q^{[1]}$, where the bounds multiply. The paper further asserts that eigenvalues of $Q$ lie in $(0,\\, 1+2\\|L\\|_2 \\cos(\\pi/(n+1))]$ and that the numerical range is contained in $(0,\\, (1+\\lambda)(1-\\lambda^{n-1})/(1-\\lambda)]$, with corresponding two-dimensional analogues.","pith_inferences":["The upper bound in Theorem 3.1(v) is not achieved by any row of $Q$; the exact 1-norm is the maximum of the row sums, which is smaller than the geometric sum over all off-diagonal distances. A natural tightening would replace the sum by the middle-row sum.","The inverse-norm inequalities inherit any hidden assumptions in the factorization cited from earlier work; a direct proof of $(I - \\lambda N^T)Q(I - \\lambda N) = \\hat{D}_\\lambda$ for all $\\lambda \\in (0,1)$ and all $n$ would put the conditioning bounds on a self-contained footing.","The identity $\\omega(I+A) = 1+\\omega(A)$ is applied without stating its conditions; it holds for nonnegative symmetric $A$, which is true for $A = L+L^T$ here, but the proof should say so explicitly to avoid an invalid general inference.","The Hadamard-inverse results suggest that the entrywise reciprocal of $Q$, which is cheap to compute, has norms comparable to the inverse in closed form; this could be useful in kernel methods requiring entrywise operations."],"forward_implications":["The sharper bound for $\\|Q\\|_1$ gives a tighter worst-case cost estimate for the matrix-vector product $Qv$ used in Sinkhorn iterations for optimal transport.","The condition-number bound $(\\mathrm{ix})$ implies that linear systems $Qx = b$ are better behaved than earlier bounds suggested, especially for large $n$.","The eigenvalue and numerical-range inclusions can be used to design preconditioners or to certify convergence of iterative solvers for Wasserstein metric matrices.","In the two-dimensional case, the bounds scale as products of one-dimensional bounds, so sharpness carries over from $n$ to $nm$ dimensions.","The Hadamard decomposition $Q = A \\circ A^T$ yields a cheap upper bound for the spectral norm, and exact formulas for the Hadamard inverse give closed-form norms of the entrywise reciprocal matrix."],"supporting_citations":[{"why":"Defines the Wasserstein-1 metric matrix and establishes the earlier norm and condition-number estimates that this paper sharpens.","marker":"[2]"},{"why":"Provides the decomposition $Q = (I-\\lambda N)^{-1} + (I-\\lambda N^T)^{-1} - I$ and the factorization $(I-\\lambda N^T)Q(I-\\lambda N) = \\hat{D}_\\lambda$ on which the inverse-norm and conditioning bounds depend.","marker":"[3]"},{"why":"Supplies the Gershgorin disc theorem, Perron-Frobenius theorems, and spectral-radius results used for eigenvalue and numerical-range inclusions.","marker":"[13]"},{"why":"Provides the numerical-radius properties and the Hadamard-product spectral-norm bound used in the 2-norm estimates.","marker":"[14]"},{"why":"Supplies the numerical-range and numerical-radius foundations, including properties (P1)-(P5) cited in the proofs.","marker":"[10]"}],"fun_headline_variants":["Sharper norm bounds for Wasserstein metric matrices","Wasserstein matrix norms: now considerably tighter","Improved condition number bounds for Wasserstein matrices","Tight Wasserstein matrix norm estimates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The inverse and condition-number bounds rest on the factorization $(I - \\lambda N^T)Q(I - \\lambda N) = \\hat{D}_\\lambda$, which the paper cites from earlier work without proving; if that factorization is not valid for every $\\lambda \\in (0,1)$ and every $n$, those bounds do not follow. Additionally, the proof of the eigenvalue bound uses the numerical-radius relation $\\omega(I+A) = 1+\\omega(A)$ as a general fact, although it holds only when $A$ is nonnegative and symmetric, a condition that is met here by $A = L+L^T$ but is not stated.","fun_headline_variants_meta":{"raw":{"variants":["Sharper norm bounds for Wasserstein metric matrices","Wasserstein matrix norms: now considerably tighter","Improved condition number bounds for Wasserstein matrices","Tight Wasserstein matrix norm estimates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000555,"raw_usage":{"total_tokens":2689,"prompt_tokens":1036,"completion_tokens":1653,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":652,"completion_tokens_details":{"reasoning_tokens":1597}},"tokens_in":652,"tokens_out":1653,"duration_ms":13569,"temperature":1.0,"reasoning_tokens":1597,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:38:20.797348+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the exact 1-norm of $Q$ for a fixed $n$, say $n=4$ and $\\lambda=0.5$, by summing row entries and comparing it to the claimed upper bound $1 + 2\\lambda(1-\\lambda^{n-1})/(1-\\lambda)$; if the bound is violated for any $\\lambda \\in (0,1)$, Theorem 3.1(v) fails. More directly, symbolically expand $(I-\\lambda N^T)Q(I-\\lambda N)$ for $n=2$ and $n=3$ and check whether it equals $\\hat{D}_\\lambda$; a single counterexample to this identity for a diagonal entry would overturn the inverse and condition-number bounds.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Wasserstein-1 metric matrix and establishes the earlier norm and condition-number estimates that this paper sharpens."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the decomposition $Q = (I-\\lambda N)^{-1} + (I-\\lambda N^T)^{-1} - I$ and the factorization $(I-\\lambda N^T)Q(I-\\lambda N) = \\hat{D}_\\lambda$ on which the inverse-norm and conditioning bounds depend."},{"cited_title":"A.; Johnson, C","cited_arxiv_id":null,"evidence_quote":"Supplies the Gershgorin disc theorem, Perron-Frobenius theorems, and spectral-radius results used for eigenvalue and numerical-range inclusions."},{"cited_title":"A.; Johnson, C","cited_arxiv_id":null,"evidence_quote":"Provides the numerical-radius properties and the Hadamard-product spectral-norm bound used in the 2-norm estimates."},{"cited_title":"E.; Rao, D","cited_arxiv_id":null,"evidence_quote":"Supplies the numerical-range and numerical-radius foundations, including properties (P1)-(P5) cited in the proofs."}],"review_version":2}