{"id":"19d89465-9e3f-4f2f-ac8a-0950ab573784","arxiv_id":"2608.05335","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For sufficiently large n, ordinary latin squares span the whole space of fuzzy latin squares, and the paper nearly classifies four-term vanishing and six-term non-vanishing examples.","lead":"The authors prove that for large orders, ordinary Latin squares span the entire space of 'fuzzy' Latin squares, a relaxed version in which matrix summands may come from shorter permutations. They also classify the smallest fuzzy Latin squares, which appear in the study of quasirandom permutation patterns.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.1's proof hinges on the o(1/n) error in Proposition 2.2 (imported from [9], only sketched in Appendix A). If the true error were Θ(1/n) with the wrong sign, θ_{1^n} would vanish and dim L_n would drop, so this external bound is the most load-bearing assumption.","rationale":"The central claim is Theorem 2.1: dim L_n = n! - (n-1)^2 for all sufficiently large n. The proof diagonalizes MM^T and needs every eigenvalue θ_μ positive except the standard representation. For the alternating character μ=1^n, equation (2.7) makes positivity equivalent to κ(A_n)/κ(S_n) = 1/2 + o(1/n). The PSD of MM^T alone only gives ≥ 1/2 - 1/(2(n-1)), so the o(1/n) error term is exactly what excludes the boundary case θ_{1^n}=0. Proposition 2.2 supplies that term but is imported from the unpublished preprint [9] and only sketched in Appendix A. I checked the sketch's random-switch uniformity step and found it can be justified (the marginal of L' is uniform because each L' has the same total weight 2^{s(L')}·2^{-s(L')}=1), but the appendix still rests on three substantial lemmas from [9] that are neither proved nor precisely credited, and the error-accounting is compressed into a single sentence. The near-linear character cases μ=2 1^{n-2} and (n-2,2) do contain unsupported inferences (citing (2.6) when the minimum character ratio only gives θ≥0), but explicit character formulas show those eigenvalues are strictly positive, so those are minor proof gaps, not correctness threats. The computer-assisted classifications in Sections 3–4 also depend on the thesis [3] and repository [4] without a snapshot, which is a reproducibility issue but secondary to the dimension theorem. Overall the reader's CONDITIONAL verdict is appropriate: the claim is plausible and likely true, but should not be accepted as fully verified until Proposition 2.2 is either proven in the paper or shown to follow from a citable, checked source with the o(1/n) error. My concern does not move the verdict; it reinforces it.","tokens_in":14973,"tokens_out":30377,"duration_ms":234689,"concrete_test":"Verify Proposition 2.2 by extracting the explicit error bound from [9, Theorem 6.4] and checking whether the template-expander and row-mixing failure probabilities in Appendix A are each o(1/n), so their sum is also o(1/n). If [9] only gives 1/2+O(1/n) or if the Appendix's 'closely follow' step cannot be made rigorous, recompute θ_{1^n} from (2.7) with the best available bound; if θ_{1^n} can be zero, the proof of Theorem 2.1 fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (2.7) gives θ_{1^n} = ℓ_n/(n-2)! · (1/(n-1) + 2κ(A_n)/κ(S_n) − 1). Since θ_{1^n} ≥ 0 by positive semidefiniteness of MM^T, positivity for Theorem 2.1 requires κ(A_n)/κ(S_n) = 1/2 + o(1/n); a merely Θ(1/n) error, even one consistent with PSD, could make θ_{1^n}=0 and increase the nullity beyond (n−1)^2. Proposition 2.2 is therefore load-bearing. The paper's support for it is (i) a citation to the unpublished preprint [9] and (ii) a sketch in Appendix A that restates Lemmas A.1–A.3 from [9] and asserts, in one paragraph, that random stable-intercalate switching makes the first two rows have equal parity with probability 1/2. The appendix does not prove the three lemmas, does not justify the o(1/n) accumulation from the stated o(1/n) failure probabilities, and does not reconcile its derivation with the exact statement of [9, Thm 6.4]. The other exceptional character cases in §2.4 (μ=2 1^{n−2}, (n−2,2)) contain hand-wavy inferences that are repairable with elementary character formulas, so they are not the central risk; Proposition 2.2 is.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces fuzzy permutation matrices P_σ^{↑n} and fuzzy latin squares as Q-linear combinations of such matrices equal to a constant matrix. It computes the dimension of the space F_n of fuzzy latin squares on S_n, proves that ordinary row-normalized latin squares span all of F_n for sufficiently large n (Theorem 2.1), gives a partial classification of four-term vanishing fuzzy latin squares, and reports a computer-assisted classification of six-term non-vanishing fuzzy latin squares. The proof of Theorem 2.1 uses the representation theory of S_n, an asymptotically equidistributed parity result for derangements in random latin squares imported from the unpublished preprint [9], and several auxiliary character computations.","tokens_in":15281,"tokens_out":10586,"duration_ms":83103,"significance":"If Theorem 2.1 is correct, it closes the dimension question for fuzzy latin squares and shows that no additional fuzzy degrees of freedom appear at large order; this would be a genuine advance connecting latin square enumeration with permutation pattern statistics. The paper also contributes a substantial computational census with explicit data and source-code references, and the arguments are parameter-free rather than fitted. However, the main theorem is not yet established at the required level of rigor: its positivity argument for the alternating character rests on an external unpublished result that is only sketched, and the treatment of several exceptional characters contains gaps or incorrect quantitative claims.","major_comments":[{"comment":"Proposition 2.2 is load-bearing for the conclusion θ_{1^n} > 0, but the appendix does not prove it. Lemmas A.1–A.3 are restated from [9] without proofs, and the paragraph following Lemma A.3 merely asserts that independent stable-intercalate switching produces a uniformly random latin square in which the first two rows have equal parity with probability exactly 1/2. No argument is given for the parity-mixture claim, for the preservation of uniformity under the switching process, or for the accumulation of the stated o(1/n) error from the individual failure probabilities. Since [9] is an unpublished preprint, this is not a minor omission.","section":"Appendix A and §2.4, Eq. (2.7)"},{"comment":"The text says that when n is even, χ_μ(n) = 1, when n is odd χ_μ(3,2^{(n-3)/2}) = 1, and then 'It follows from (2.6) that θ_μ > 0.' This does not follow: with χ_μ(1^n) = n-1, the displayed value gives max |χ_μ(λ)|/χ_μ(1^n) = 1/(n-1), so the lower bound in (2.6) is only nonnegative, not positive. A separate estimate for the contribution of the extremal conjugacy class in (2.4) is needed and is not supplied.","section":"§2.4, case μ = 2 1^{n-2}"},{"comment":"The assertion that χ_μ(2^{n/2}) > 0 implies θ_μ > 0 via (2.6) is not justified. Positivity at one conjugacy class does not control the maximum of |χ_μ(λ)|/χ_μ(1^n) over all derangement classes, which is what (2.6) requires. No bound on that maximum is given for this character, so the conclusion θ_μ > 0 is unsupported.","section":"§2.4, case μ = (n−2,2)"},{"comment":"The split-merge argument contains a quantitative error. The text claims that for n ≥ 8, |C_{λ'}| > n |C_λ| for λ' = 4 1^{n−4} and λ = 2^{n/2}, and then uses Lemma 2.3 to conclude κ(λ') ≥ n κ(λ)/2. The class-size claim is false in the relevant parameter range: for n = 14, |C_{2^7}| = 135135 while |C_{4 1^{10}}| = 6006, so the opposite inequality holds. Consequently the displayed lower bound for θ_μ in this case is unsupported, and the proof of positivity for μ = 2^2 1^{n−4} is incomplete. Since Lemma 2.5 itself is also delegated to the unpublished thesis [3] without proof, this entire exceptional case needs a revised argument.","section":"§2.4, case μ = 2^2 1^{n−4}, after Lemma 2.5"}],"minor_comments":[{"comment":"The table labelled 'Table 4' is referenced twice for two different objects; the second reference, which lists irreducible six-term fuzzy latin squares, should point to a separate table (probably Table 5).","section":"§4, Table 4/Table 5"},{"comment":"The text says 966 length lists were enumerated, but Table 4 only displays totals for n = 4, 5, 6; it would be helpful to state explicitly how many candidate lists survived the Lemmas 4.1 and 4.2 filters for each n.","section":"§4, search description"},{"comment":"In the first case of the proof, 'We may therefore assume wlog that c_1, c_2 > 0 and c_3 < 0' requires more justification: scaling by a negative constant and permuting terms does not by itself put the coefficients into this sign pattern, and the subsequent support-size argument depends on that pattern.","section":"§3.2, Theorem 3.6 proof"},{"comment":"The notation in the line '1̸∈λ⊢n' is nonstandard; writing λ with no part equal to 1 as '1̸∈λ' is understandable but should be defined explicitly the first time it is used.","section":"§2.4, Eq. (2.6)"}],"recommendation":"major_revision","confidential_remarks":"The main theorem is interesting and the computational census is valuable, but the proof of Theorem 2.1 depends on (i) an unpublished external result with an insufficient sketch and (ii) several exceptional-character arguments that are either incomplete or contain a numerically false class-size bound. These are load-bearing for the paper's central claim. I recommend asking the authors for a complete proof of Proposition 2.2 or a restatement of Theorem 2.1 conditional on [9], and for a corrected and fully detailed treatment of the cases μ = 2 1^{n−2}, (n−2,2), and 2^2 1^{n−4}."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the main theorem is new and likely correct, but the proof as written has a load-bearing gap in the alternating-character case. I'd send it to a good referee, with a request for revisions.\n\nWhat's actually new: Theorem 2.1 (dim L_n = n! - (n-1)^2 for large n) is a genuine step beyond [5], and the eigenvalue proof via MM^T is the right framework. The byproduct — kappa(A_n)/kappa(S_n) ≥ 1/2 - O(1/n), improving the 1/4 - o(1) in [2] — is neat. The four-term vanishing classification and the six-term census also extend the known list, and the tables will be useful to people working on quasirandom forcing.\n\nThe soft spots are concentrated in the proof of Theorem 2.1. First, the display after (2.6): for mu = 2 1^{n-2} the text says theta_mu > 0 follows from (2.6), but with chi(1) = n-1 and the possibility of a character ratio of magnitude 1/(n-1), (2.6) only gives theta_mu >= 0. You need a separate estimate to rule out equality. This is repairable with elementary character computations, but it is not in the paper.\n\nThe bigger issue is Proposition 2.2. The estimate kappa(A_n)/kappa(S_n) = 1/2 + o(1/n) is load-bearing: through (2.7) it's what forces theta_{1^n} > 0. If the true error were Theta(1/n) with the wrong sign, the alternating eigenvalue would vanish and the theorem would not go through. Appendix A sketches the Kwan-Petrova-Sawhney argument but does not prove the three lemmas, does not justify the o(1/n) accumulation, and does not reconcile the sketch with the exact statement of [9, Thm 6.4]. Since [9] is itself an unpublished preprint, this is the weakest link in the chain.\n\nThe rest of the paper is in better shape. The split-merge estimate for mu = 2^2 1^{n-4} is clever, the small-n verification (n <= 11) gives real evidence, and the census seems thorough. Reproducibility could be improved: the six-term solutions live on GitHub with no snapshot in the arXiv listing, so a commit hash or ancillary file would help.\n\nWho should read it: anyone in latin squares, permutation patterns, or quasirandom forcing. It deserves a serious referee. The referee should insist on either a self-contained proof of Proposition 2.2 or a citation to a published version of [9], and should ask for the missing character estimates in the near-linear cases. With those repairs, I'd accept; as is, it's conditional.","headline":"Genuinely new dimension result for fuzzy latin squares, but the alternating-character case leans on an external o(1/n) estimate that is only sketched; patch that and it's a solid paper.","tokens_in":15845,"tokens_out":4952,"would_cite":true,"duration_ms":41895,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05B15","20C30","05A05"],"pacs":[],"model":"deepseek-v4-flash","headline":"For sufficiently large n, ordinary row-normalized Latin squares span the entire space of fuzzy Latin squares, whose dimension is n! - (n-1)^2.","keywords":["fuzzy latin squares","permutation matrices","latin square dimension","quasirandom permutations","symmetric group characters","vanishing sums","pattern statistics","derangement parity"],"falsifier":"Enumerate, for as large an n as feasible, the parity distribution of the second row relative to the first in uniformly random row-normalized Latin squares; if for some n the ratio κ(A_n)/κ(S_n) deviates from 1/2 by an amount that is not o(1/n), Proposition 2.2 is false and Theorem 2.1 no longer follows by this route. Alternatively, exhibit for a specific large n a fuzzy Latin square that is provably not in the span of ordinary Latin squares, which would directly contradict dim L_n = n! - (n-1)^2 for that order.","tokens_in":14731,"feed_emoji":"🧩","tokens_out":6078,"duration_ms":45935,"temperature":0.7,"pith_summary":"The paper studies a fractional relaxation of Latin squares in which the permutation-matrix summands are replaced by fuzzy permutation matrices that smear a short permutation σ ∈ S_k onto an n×n frame. A fuzzy Latin square is a rational linear combination of these fuzzy matrices equal to a constant matrix. The paper's main result is that for all sufficiently large n, the subspace L_n spanned by ordinary row-normalized Latin squares equals the whole space F_n of fuzzy Latin squares, with dimension n! - (n-1)^2. This says the fuzzy relaxation introduces no genuinely new degrees of freedom at large order: every constant-line-sum combination of fuzzy permutation matrices is a rational combination of ordinary Latin squares. The proof combines representation theory of the symmetric group with a parity-equidistribution input about two rows in a random Latin square, and the paper also classifies four-term vanishing squares and six-term non-vanishing fuzzy Latin squares.","feed_headline":"Ordinary latin squares span fuzzy ones for large n","feed_subtitle":"Dimension equals n! - (n-1)^2, so fuzzy Latin squares add no new degrees of freedom asymptotically.","key_machinery":"The central object is the eigenvalue decomposition of the incidence Gram matrix M M^T, where M is the n! × ℓ_n matrix of permutations versus row-normalized Latin squares. Its eigenvalues are character sums θ_μ = Σ_σ ℓ_n(id, σ) χ_μ(σ)/χ_μ(1^n). The zero eigenspace for the standard representation μ = (n-1,1) has dimension (n-1)^2; the whole proof is showing every other θ_μ > 0 for large n. The deciding input for the alternating character μ = 1^n is Proposition 2.2, the parity-equidistribution estimate κ(A_n)/κ(S_n) = 1/2 + o(1/n), imported from work on parities in random Latin squares.","core_discovery":"In the paper's own terms, Theorem 2.1 establishes that dim L_n = n! - (n-1)^2, hence L_n = F_n for sufficiently large n. Here L_n is the subspace of Q[S_n] spanned by the symbol sets of row-normalized Latin squares of order n, and F_n is the space of fuzzy Latin squares whose constituent permutations all have length exactly n. The proof shows that the Gram matrix M M^T, indexed by permutations against Latin squares, has all eigenvalues strictly positive except the zero eigenvalue of multiplicity (n-1)^2 coming from the standard representation; the alternating-character eigenvalue is forced positive by the asymptotic κ(A_n)/κ(S_n) = 1/2 + o(1/n), and the remaining characters are handled by Larsen–Shalev bounds and split-merge estimates. For n ≤ 11 the dimension formula is verified computationally.","pith_inferences":["If L_n = F_n for large n, then any positive-coefficient fuzzy Latin square is asymptotically a limit of ordinary Latin squares in the vector-space sense; a testable consequence is that the quasirandom-forcing threshold, currently six patterns, cannot be lowered by using fuzzy mixed-length statistics alone.","The algebraic positive-semidefinite argument already gives κ(A_n)/κ(S_n) ≥ 1/2 - Θ(1/n), so a direct combinatorial proof of the matching o(1/n) upper error rate would make Theorem 2.1 self-contained rather than dependent on imported intercalate-switching machinery.","The exceptional vanishing length profiles at small orders suggest that any failure of the large-n picture lives at bounded order; checking n = 12 and n = 13 with the appendix data of reference [2] would pinpoint the smallest order at which the dimension formula holds.","A fuzzy analogue of orthogonality for Latin squares, floated in the conclusion, becomes more natural now that F_n and L_n coincide asymptotically: orthogonal pairs would correspond to complementary decompositions of the all-ones matrix within a common span."],"forward_implications":["For all sufficiently large n, every fuzzy Latin square of order n is a rational linear combination of ordinary row-normalized Latin squares; the fuzzy relaxation does not enlarge the space of balanced permutation statistics asymptotically.","The dimension formula dim L_n = n! - (n-1)^2 now holds for every n ≤ 11 and for all large n; only finitely many orders are left unresolved.","Since quasirandom-forcing statistics must be fuzzy Latin squares, the six-term classification delimits which six-permutation pattern statistics can force quasirandomness; a next step suggested by the paper is to test whether the expression (5.1) is the unique six-term forcing expression.","Vanishing fuzzy Latin squares with four terms can only have length profiles (n,n,n,n), (n,n,n-1,n-1), or one of the three sporadic profiles (3,3,3,2), (3,3,3,1), and (4,4,3,2)."],"supporting_citations":[{"why":"Supplies the Birkhoff–von Neumann theorem used to identify the image of the fuzzy matrix map with the space H_n of constant-line-sum matrices, fixing the dimension of F_n.","marker":"[1]"},{"why":"Supplies the split-merge inequalities (Lemma 2.3) and small-n counts of ℓ_n(id, δ) used to control eigenvalues and to verify the dimension formula for n ≤ 11.","marker":"[2]"},{"why":"Introduces fuzzy permutation matrices, the fuzzy-Latin-square condition, and the prior classification of fuzzy Latin squares with at most five terms that the six-term search extends.","marker":"[5]"},{"why":"Supplies the second family of four-term vanishing squares (Example 3.2), used in the vanishing-square classification.","marker":"[8]"},{"why":"Provides the parity equidistribution estimate κ(A_n)/κ(S_n) = 1/2 + o(1/n) (Proposition 2.2), the key input for positivity of the alternating-character eigenvalue.","marker":"[9]"},{"why":"Provides the Larsen–Shalev character bounds used to show positivity of θ_μ for all but a finite list of partitions.","marker":"[10]"}],"fun_headline_variants":["Fuzzy Latin squares are just ordinary ones in disguise","Dimension formula collapses fuzzy Latin squares to classical","Large-n theorem: fuzzy Latin squares span no new space","Ordinary Latin squares generate all fuzzy ones for large n","Dim L_n = n! - (n-1)^2 settles fuzzy span"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of Theorem 2.1 relies on the parity-equidistribution estimate κ(A_n)/κ(S_n) = 1/2 + o(1/n) for two rows in a random Latin square, stated as Proposition 2.2 and only sketched via intercalate switching; if that error rate fails, the positivity of the alternating-character eigenvalue, and with it the equality L_n = F_n, breaks.","fun_headline_variants_meta":{"raw":{"variants":["Fuzzy Latin squares are just ordinary ones in disguise","Dimension formula collapses fuzzy Latin squares to classical","Large-n theorem: fuzzy Latin squares span no new space","Ordinary Latin squares generate all fuzzy ones for large n","Dim L_n = n! - (n-1)^2 settles fuzzy span"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1352,"prompt_tokens":928,"completion_tokens":424,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":342}},"tokens_in":544,"tokens_out":424,"duration_ms":4418,"temperature":1.0,"reasoning_tokens":342,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T15:05:56.837977+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Enumerate, for as large an n as feasible, the parity distribution of the second row relative to the first in uniformly random row-normalized Latin squares; if for some n the ratio κ(A_n)/κ(S_n) deviates from 1/2 by an amount that is not o(1/n), Proposition 2.2 is false and Theorem 2.1 no longer follows by this route. Alternatively, exhibit for a specific large n a fuzzy Latin square that is provably not in the span of ordinary Latin squares, which would directly contradict dim L_n = n! - (n-1)^2 for that order.","supporting_citations":[{"cited_title":"Birkhoff, Tres observaciones sobre el algebra lineal, Univ","cited_arxiv_id":null,"evidence_quote":"Supplies the Birkhoff–von Neumann theorem used to identify the image of the fuzzy matrix map with the space H_n of constant-line-sum matrices, fixing the dimension of F_n."},{"cited_title":"Cavenagh, C","cited_arxiv_id":null,"evidence_quote":"Supplies the split-merge inequalities (Lemma 2.3) and small-n counts of ℓ_n(id, δ) used to control eigenvalues and to verify the dimension formula for n ≤ 11."},{"cited_title":"Crudele, P.J","cited_arxiv_id":null,"evidence_quote":"Introduces fuzzy permutation matrices, the fuzzy-Latin-square condition, and the prior classification of fuzzy Latin squares with at most five terms that the six-term search extends."},{"cited_title":"Forcing quasirandomness with 4-point permutations","cited_arxiv_id":"2407.06869","evidence_quote":"Supplies the second family of four-term vanishing squares (Example 3.2), used in the vanishing-square classification."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the parity equidistribution estimate κ(A_n)/κ(S_n) = 1/2 + o(1/n) (Proposition 2.2), the key input for positivity of the alternating-character eigenvalue."},{"cited_title":"Larsen and A","cited_arxiv_id":null,"evidence_quote":"Provides the Larsen–Shalev character bounds used to show positivity of θ_μ for all but a finite list of partitions."}],"review_version":1}