{"id":"f9c6220a-9a3c-4462-aa81-25ca8c189a2e","arxiv_id":"2611.17197","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For dense regular uniform hypergraphs satisfying new codegree and intersection conditions, the logarithm of the matching partition function is determined up to the first sublinear correction, yielding sharp typical matchings and several enumeration refinements.","lead":"This paper proves a sharp formula for the number of matchings in many regular hypergraphs, including the first correction below the leading exponential term. It then derives statistical and enumeration consequences for Latin squares, Steiner systems, and matroids of fixed rank.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The proof that the greedy counting error is o(nd^{-1/k}) relies on the Bennett–Bohman estimate holding uniformly for every regular completion, but the completions can have super-polynomial order and the estimate's hypothesis requires L_2 = o(D/(log N)^5); the paper controls log N_i = O(log^2 n)…","rationale":"The reader identified the weakest assumption as the uniform validity of the Bennett‑Bohman estimate through all regular completions, but did not pinpoint the specific gap: the pair codegree of the completed hypergraph is not bounded.  The paper controls the order and degree of the completions but does not bound the pair codegree of the new vertices, which can be as large as the degree itself.  This is a concrete technical gap that could invalidate the iteration.  The test is feasible and would settle whether the gap is real.  The verdict should be CONDITIONAL because the central claim is plausible but depends on a missing verification; if the test shows the hypothesis is satisfied, the paper would be correct.  The reader's verdict was ACCEPT with MODERATE confidence; this concern raises the correctness risk to HIGH unless resolved.","tokens_in":55658,"tokens_out":848,"duration_ms":193794,"concrete_test":"For a concrete test, construct a small example: take a 3‑uniform linear hypergraph H on n vertices with degree d = n^{1/2}.  Apply the completion Lemma 4.2 to a residual hypergraph after one greedy stage (N/2 vertices, degree ≈ 2^{-(k-1)}d).  Compute the maximum pair codegree of the completed hypergraph K_P.  If it exceeds D_{i+1} / (log N_i)^5, then the Bennett–Bohman hypothesis is violated and the iteration cannot be applied as claimed.  The test should be repeated for several random prefixes to see whether the violation is typical.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central lower bound (Proposition 6.2) depends on Theorem 4.1, which iterates the short‑greedy‑stage estimate (Theorem 4.3) through T stages.  At each stage the residual hypergraph is completed to an exactly regular hypergraph K_P via Lemma 4.2.  The Bennett–Bohman estimate (Theorem 4.3) requires that the maximum pair codegree L_2 of the hypergraph to which it is applied satisfy L_2 = o(D/(log N)^5).  The paper controls the order of the completed hypergraph (Claim 4.5: log N_i = O(log^2 n)) and the degree D_i, but it does not bound the pair codegree of the completed hypergraph K_P.  Lemma 4.2 only asserts Δ_j(K) ≤ max{Δ_j(F), 1} for j ≥ 2, but the completion adds new vertices and edges; a pair consisting of one old vertex and one new vertex can have codegree as large as D_{i+1} (the new vertex is adjacent to every copy of the old vertex in the auxiliary construction).  If D_{i+1} is large (it is at least d p_i^{k-1} and can be as large as 2d), then L_2 for K_P could be Θ(D), violating the hypothesis L_2 = o(D/(log N)^5).  The paper does not address this, and the iteration may therefore apply Theorem 4.3 outside its stated hypotheses.  If the estimate fails for some completion, the geometric accumulation of errors in Claim 4.6 collapses and the lower bound R_λ ≥ na_λ – o(na_λ) is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper claims to determine the first sublinear correction to the logarithm of the matching partition function for a class of n-vertex d-regular k-uniform hypergraphs, uniformly for activities in each fixed compact interval of (0,∞). The main theorem (Theorem 1.6) gives an asymptotic formula for log Z_H(λ) with a correction term n(λd)^{-1/k}. The proof combines a random greedy counting argument with regular completions and an entropy method based on intersection patterns. Applications include partial transversals in Latin squares, linear 3-uniform hypergraphs, and matroids of fixed rank.","tokens_in":56123,"tokens_out":3844,"duration_ms":34078,"significance":"If correct, the result provides the first determination of the sublinear correction to the matching partition function for regular hypergraphs, resolving an averaged version of Kahn's conjecture and giving refined counts for several important combinatorial objects. The techniques introduced (short greedy stages with regular completions, intersection pattern coefficients) are likely to be influential.","major_comments":[{"comment":"The proof of the lower bound (Proposition 6.2) depends on Theorem 4.1, which iterates the short-greedy-stage estimate (Theorem 4.3) through T stages. At each stage the residual hypergraph is completed to an exactly regular hypergraph K_P via Lemma 4.2. The Bennett–Bohman estimate (Theorem 4.3) requires that the maximum pair codegree L_2 of the hypergraph to which it is applied satisfy L_2 = o(D/(log N)^5). The paper controls the order of the completed hypergraph (Claim 4.5: log N_i = O(log^2 n)) and the degree D_i, but it does not bound the pair codegree of the completed hypergraph K_P. Lemma 4.2 only asserts Δ_j(K) ≤ max{Δ_j(F), 1} for j ≥ 2, but the completion adds new vertices and edges; a pair consisting of one old vertex and one new vertex can have codegree as large as D_{i+1} (the new vertex is adjacent to every copy of the old vertex in the auxiliary construction). If D_{i+1} is large (it is at least d p_i^{k-1} and can be as large as 2d), then L_2 for K_P could be Θ(D), violating the hypothesis L_2 = o(D/(log N)^5). The paper does not address this, and the iteration may therefore apply Theorem 4.3 outside its stated hypotheses. If the estimate fails for some completion, the geometric accumulation of errors in Claim 4.6 collapses and the lower bound R_λ ≥ na_λ – o(na_λ) is unsupported.","section":"§4, Lemma 4.2, Theorem 4.3, Claim 4.5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Claim 4.5 bounds log N_i = O(log^2 n) for i ≤ T, but this bound is derived assuming the completion uses a_i = O(D_{i+1}^2) copies and w_i/a_i ≤ (k-1)N_i/2. However, Lemma 4.2 states that a = O_k(D^2) copies are used, but this constant may depend on k and the choice of prime Q. The bound a_i = O_k(D_{i+1}^2) is plausible, but the paper does not verify that the number of copies a_i is polynomial in D_{i+1} with a constant that does not depend on the specific residual hypergraph. Moreover, the bound w_i/a_i ≤ (k-1)S/D is given, but S is the total deficit, which could be as large as O(N_i D_i). The paper then asserts w_i/a_i = O_k(N_i ℓ sqrt(L_2^{(i)}/D_i)) (equation (4.7)), but this relies on the deficit being O(N_i sqrt(D_i L_2^{(i)} ℓ)). This deficit estimate itself depends on the degree estimate from Theorem 4.3, which is only guaranteed to hold with high probability for the original hypergraph, not for the completed one. The argument appears circular: to bound the deficit we need the degree estimate, but to apply the degree estimate we need to bound the codegree of the completed hypergraph, which is not done.","section":"§4.3, Claim 4.5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Lemma 5.4 (random-order exposure) bounds H(P) for a random perfect matching P of H[U]. The proof uses the Radhakrishnan entropy argument and assumes that for each vertex v, the number of available edges through v given the history is at most 1 + g_v(P) y^{k-1} + sum b_{v,π}(P) y^{m(π)-1}. However, this bound counts all edges through v that are disjoint from previously revealed matching edges, but does not account for the fact that some of these edges may not be extendable to a perfect matching of the remaining vertices. The entropy bound (3.2) uses the support size, not the number of edges that can be completed to a perfect matching. The paper implicitly assumes that every edge through v that is disjoint from the revealed matching can be extended to a perfect matching of the remaining vertices, which is not justified. This is a known issue with the Radhakrishnan method for hypergraphs (see Luria's paper, which the paper cites, but Luria's application is for perfect matchings in complete hypergraphs where the condition holds by symmetry). Here the hypergraph is arbitrary, and the bound may overcount the support, potentially invalidating the entropy upper bound.","section":"§5, Lemma 5.4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Claim 6.2 bounds H(W) - E[S_W]/(kd) using Shearer's inequality and a weighting argument. The application of Shearer's inequality (Lemma 3.3) requires that every vertex belongs to at least r members of the family A. Here the family consists of the edges of H plus a_v copies of each singleton {v}, where a_v = d - d_H(v). The paper states 'Add a_v copies of {v} to the edge family, so every coordinate belongs to exactly d members.' However, the a_v copies are not edges of the hypergraph; they are artificial. Shearer's inequality as stated applies to a family of subsets of the random vector's coordinates. The paper does not verify that the inequality holds when the family includes repeated singletons. Moreover, the entropy H(W∩e) for a singleton {v} is just H(1_{v∈W}), which is fine, but the inequality's factor 1/r requires that each coordinate appears in at least r members of the family. Here each vertex v appears in d_H(v) edges and a_v singleton copies, total d. So r = d. But the inequality then gives H(W) ≤ (1/d) [ sum_e H(W∩e) + sum_v a_v H(1_{v∈W}) ]. This is valid if Shearer's inequality holds for families with repetitions, which is true (the standard proof via subadditivity of entropy works). However, the paper then uses this to bound H(W) - E[S_W]/(kd). The derivation of (6.9) from (6.8) and Shearer's inequality needs to be checked carefully: S_W = k sum_e 1_{e∩W≠∅} - d|W| + sum_{v∉W} a_v. The expectation of the last term is sum_v a_v (1-p_v). The paper's expression (6.9) seems to omit this term or absorb it into the np/k term. The derivation is not fully explicit, and the cancellation in Claim 6.2 may rely on an algebraic manipulation that is not clearly justified.","section":"§6.2, Claim 6.2","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The statement of Theorem 1.6 includes the condition b_π(H) ≤ C d^{m(π)/k} n^{-ε} for π∈Π_k. The definition of Π_k excludes the partition (k) (which corresponds to the edge itself) and (1,1,...,1) (which corresponds to an edge meeting k different matching edges). The paper states that these are excluded, but it is not entirely clear why (1,1,...,1) is excluded; later in the proof, the term g_v(P) y^{k-1} accounts for edges meeting k blocks. The definition of b_π(H) should be clarified to explicitly state which partitions are included.","section":"§1, Theorem 1.6","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The proof of Lemma 4.5 uses Freedman's inequality to control the martingale M_t. The bound on the conditional variance is given as O_k(L_j L_{j+1}/N). However, the derivation of this bound is sketchy: it uses E[(ΔY_S(i))^2 | F_i] ≤ k L_{j+1} E[-ΔY_S(i) | F_i]. This inequality relies on the fact that each removal of a vertex destroys at most L_{j+1} edges counted by Y_S, and each step removes at most k vertices. The factor k L_{j+1} is an upper bound on the maximum possible squared change, but the conditional variance bound should be more carefully justified. Also, the paper uses the bound E[-ΔY_S(i) | F_i] = O_k(L_j/N), which comes from the drift calculation. The product L_j L_{j+1}/N is then used. This seems plausible but the argument is compressed.","section":"§4.1, Lemma 4.5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The statement of Lemma 4.6 includes the term O_{k,C_0}(ℓ^2 sqrt(L_2/D)). The proof uses the bound w_P/(a_P N) = O_{k,C_0}(ℓ sqrt(L_2/D)) from (4.7). However, the derivation of (4.7) relies on the deficit estimate S = O_{k,C_0}(N sqrt(D L_2) ℓ). This deficit estimate itself depends on the degree estimate from Theorem 4.3, which is only guaranteed for good trajectories. The lemma assumes that good trajectories have probability at least 1/2, but the bound on w_P/(a_P N) should be uniform over all P in P. The paper should argue that for every good prefix, the deficit is bounded as claimed, which it does implicitly, but the dependence on the constant C_1 in the definition of D' should be made explicit.","section":"§4.2, Lemma 4.6","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The integral estimate in Lemma 5.2 uses the bound (5.3) for ∫_0^1 log(1 + x y^{-q}) dy. For q=1, the bound is O(x log(e/x)). The paper then applies this to x = b_π(H)/d and q = k - m(π). For q=1, this requires m(π) = k-1. The paper states that b_π(H)/d = O(d^{-1/k} n^{-ε}) under the hypotheses, so the contribution is O(d^{-1/k} n^{-ε} log n). This is o(d^{-1/k}) because n^{-ε} log n → 0. However, the paper also uses the bound for q≥2, where the contribution is O((b_π(H)/d)^{1/q}). For q=2, this is O((d^{-1/k} n^{-ε})^{1/2}) = O(d^{-1/(2k)} n^{-ε/2}), which is o(d^{-1/k}) only if d^{-1/(2k)} = o(d^{-1/k}), i.e., d^{1/(2k)} → ∞, which holds. But the paper should check that the exponent 1/q is indeed ≤ 1/2 for q≥2, which is true. The argument is sound but the presentation could be clearer.","section":"§5.1, Lemma 5.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The proof of Lemma 6.1 uses Claim 6.1 to bound f(p, θ(1-p)) in terms of u = (log T)/k - g. The claim requires u > 0, which is ensured by the hypothesis g < (log T)/k. The bound then gives (1-p)g + (p/k) log T - f(p, θ(1-p)) ≥ g + θ(1-p) log(1 - T^{-1/k} e^g). The paper then drops the factor (1-p) in the logarithm term, using that the logarithm is negative and 1-p ≤ 1. This is valid but the inequality becomes weaker; the final bound is still sufficient. The step is correct but could be explained more clearly.","section":"§6.1, Lemma 6.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The proof of Proposition 7.1 uses the exponential moment estimate (7.1) to derive tail bounds via Markov's inequality. The choice of s depends on x, and the paper uses a common o(n(λd)^{-1/k}) error bound for all |s| ≤ s_0. This is valid because the error in (7.1) is uniform over the interval. However, the paper then uses this error bound in the tail estimate for both upper and lower tails, and then integrates to get the second moment bound. The integration step uses the tail bound (7.2) for x ≥ η n(λd)^{-1/k}. The derivation of (7.2) from Claim 7.1 uses the fact that o(n(λd)^{-1/k}) ≤ (cη/2) x for x ≥ η n(λd)^{-1/k} and sufficiently large n. This is correct because o(n(λd)^{-1/k}) is o(x) for such x. The argument is sound.","section":"§7.1, Proposition 7.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The proof for k = t+2 and k = t+3 relies on Claims 8.1 and 8.2, which verify the hypotheses of Theorem 1.6. The bounds on b_π(H) are derived by counting the number of candidate blocks B that intersect a given matching P in a specified pattern. The arguments are combinatorial and appear correct, but they are somewhat intricate and the paper does not provide full details for all cases. For example, for k = t+3 with h ≥ 2 parts of size t+1, the bound is O(n) (not O(n^2)). The justification is that after fixing the first two matching blocks and their intersections, the candidate B is determined. This relies on the fact that two (t+1)-subsets of B intersect in exactly t-1 points, and that the span of these two subsets is B. The paper should verify that this holds for all configurations, which it does briefly. The argument is plausible but the reader may need to fill in some steps.","section":"§8.2, Proof of Theorem 2.5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The proof of Proposition 8.1 (maximum-degree upper bound) is given in Appendix C. The proof follows the same entropy argument as in Section 5, but with the simplification that the hypergraph is linear, so all b_π vanish. The paper uses the tangent bound and integral estimates to obtain the result. The proof seems correct, but it relies on Lemma 5.4, which itself has the issue mentioned in major comment 3. The validity of Proposition 8.1 is therefore contingent on fixing that issue.","section":"§8.3, Proposition 8.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The concluding remarks mention several open questions and limitations. The paper acknowledges that the pair-codegree range may be extended, that the variance asymptotic is not determined, and that the pointwise vacancy question remains open for linear hypergraphs. This is appropriate and shows the authors are aware of the scope.","section":"§9, Concluding remarks","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper states that ChatGPT carried out the technical calculations in Section 4 and contributed to the development of the entropy method. This is unusual for a serious journal submission. While the paper states that the author has independently verified all AI-generated results, the reliance on AI for core technical calculations raises concerns about reproducibility and the depth of verification. The referee cannot independently verify that the AI-generated calculations are correct without redoing them from scratch. This is a meta-concern that may affect the journal's policy.","section":"Statement of AI use","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Hyunwoo Lee has written a substantial paper. The main result—determining the first sublinear correction n(λd)^{-1/k} to the matching partition function for a broad class of regular k-uniform hypergraphs—is a genuine advance. The paper also gives the first sublinear correction to several important enumeration formulas (Latin squares, Steiner systems, matroids) and establishes an averaged version of Kahn's conjecture. The proof combines a short-stage regularized greedy iteration with an intersection-pattern entropy method. The paper is well-structured, the hypotheses are explicit, and the author is honest about limitations (e.g., the pair-codegree range is not optimal, the variance asymptotic is not resolved). The applications are nontrivial and the verification of the structural conditions for each application is carried out in detail.\n\nThe soft spots are real but not fatal. The stress-test concern about the Bennett-Bohman estimate holding uniformly for all regular completions is the most serious. The paper controls log N_i = O(log^2 n) and the degree D_i, but does not bound the pair codegree of the completed hypergraph K_P. Lemma 4.2 only asserts Δ_j(K) ≤ max{Δ_j(F), 1} for j ≥ 2, but the completion adds new vertices and edges; a pair consisting of one old vertex and one new vertex can have codegree as large as D_{i+1}. If D_{i+1} is large (it can be as large as 2d), then L_2 for K_P could be Θ(D), violating the hypothesis L_2 = o(D/(log N)^5) of Theorem 4.3. The paper does not address this, and the iteration may therefore apply Theorem 4.3 outside its stated hypotheses. If the estimate fails for some completion, the geometric accumulation of errors in Claim 4.6 collapses and the lower bound R_λ ≥ na_λ – o(na_λ) is unsupported. This is a genuine gap that a referee would need to check carefully. The paper also states that ChatGPT carried out the technical calculations in Section 4; without independent verification, this is a concern for soundness, though the author states they have verified all AI-generated results.\n\nWho is this paper for? Specialists in extremal combinatorics, probabilistic combinatorics, and design theory. It deserves a serious referee—the paper is important enough and the proof is detailed enough to warrant careful review. The stress-test concern is a genuine soft spot, but it is not obviously fatal; it may be fixable with additional argument. I would recommend sending it to peer review with a request to address the pair-codegree issue in the completion step.\n\nRecommendation: Accept for peer review, with a note that the referee should carefully check Claims 4.4 and 4.5 and the uniform validity of the Bennett-Bohman estimate through all regular completions.","headline":"A serious, technically deep paper that resolves the averaged form of Kahn's conjecture for a broad class of regular hypergraphs and gives the first sublinear correction to several important enumeration formulas.","tokens_in":56627,"tokens_out":740,"duration_ms":11969},"pith_extraction":{"msc":["05C65","05C70","05D40","05B35"],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Paper identifies first sublinear correction to matching partition function in regular hypergraphs","keywords":["matching partition function","regular hypergraph","random greedy matching","entropy method","vacancy probability","Kahn's conjecture","partial Latin squares","matroid enumeration"],"falsifier":"Construct a regular linear hypergraph of polynomial degree with d ≥ n^ε and Γ(H) = O($d^{{-1/k}}$ $n^{{-ε}}$) where the expected number of uncovered vertices deviates from (1+o(1)) n $d^{{-1/k}}$ by a factor larger than 1+o(1).","tokens_in":57116,"feed_emoji":"","tokens_out":1226,"duration_ms":15192,"temperature":0.7,"pith_summary":"This paper determines the first sublinear correction to the logarithm of the matching partition function for a broad class of regular k-uniform hypergraphs, uniformly for activities in each fixed compact interval of (0,∞). The result gives the sharp mean, concentration, and large deviations for the number of uncovered vertices, and refines counts for partial Latin squares, Steiner systems, subspace designs, and matroids of fixed rank.","feed_headline":"First sublinear correction found for hypergraph matchings","feed_subtitle":"New counting method determines the mean deficit and its fluctuations in regular hypergraphs, refining counts for Latin squares, Steiner…","key_machinery":"The proof combines a novel variant of random greedy counting, based on short stages and regular completions, with a new entropy method that uses intersection patterns to control reconstruction after deleting matching edges.","core_discovery":"For fixed k≥3 and ε>0, every simple n-vertex d-regular k-uniform hypergraph with d ≥ n^ε, Γ(H) ≤ C $d^{{-1/k}}$ $n^{{-ε}}$, and b_π(H) ≤ C $d^{{m(π)/k}}$ $n^{{-ε}}$ for all π satisfies log Z_H(λ) = (n/k)(log(λ d) - k + 1) + (1+o(1)) n (λ d)^{-1/k}, uniformly for λ in each compact subinterval of (0,∞). This determines the first sublinear correction to the matching partition function and, via activity shifts, gives the mean, concentration, and large deviations of the number of uncovered vertices.","pith_inferences":["The method suggests that the first sublinear correction to the partition function is universal for regular hypergraphs with sufficiently small higher codegrees, not just those satisfying the specific intersection pattern bounds.","The technique of short greedy stages with regular completions could be adapted to other counting problems where the natural process loses regularity too quickly.","The entropy reconstruction bound using intersection patterns provides a general tool for converting deletion randomness into counting gains, potentially applicable to other hard-core models on hypergraphs.","The refined matroid count opens the possibility of determining the next correction for rank-3 matroids, where four-point hyperplanes contribute at order n^2."],"forward_implications":["The counting estimate implies an averaged version of Kahn's conjecture: under the uniform measure, the expected number of uncovered vertices is (1+o(1)) n d^{-1/k}, with concentration and large deviations on this scale.","Every Latin square of order n has (n/e^2)^n exp((3+o(1)) n^{2/3}) partial transversals, and almost all partial transversals have (1+o(1)) n^{2/3} missing cells.","There are (n/e^2)^{n^2/6} exp((1/2+o(1)) n^{5/3}) linear 3-uniform hypergraphs on [n], and almost all have n^2/6 - (1/6+o(1)) n^{5/3} edges.","For each fixed r ≥ 4, the number m(n,r) of rank-r matroids on [n] satisfies log m(n,r) = (n^{r-1}/r!)(log n - r + 1) + (1/((r-1)!)+o(1)) n^{r-1-1/r}, refining prior estimates."],"supporting_citations":[{"why":"Supplies the degree estimate for random greedy matching in the short greedy stages (Theorem 4.3).","marker":"[5]"},{"why":"Provides the leading estimate log|M(H)| = (n/k)(log d - k + 1) + o(n) under a small pair-codegree assumption, which the present paper refines.","marker":"[6]"},{"why":"Establishes sharp bounds for the matching partition function in graphs via occupancy fractions; the hypergraph extension motivates the present work.","marker":"[17]"},{"why":"Disproves Kahn's pointwise conjecture and motivates the averaged version proved here; also supplies the counterexample in Appendix A.","marker":"[47]"},{"why":"Gives leading estimates for partial Steiner systems and matroids that are refined by the present paper's correction.","marker":"[70]"}],"fun_headline_variants":["Hypergraph matchings get first sublinear correction","Matching formula refined for regular hypergraphs","New count reveals missing cells in Latin squares","Kahn's conjecture proved in averaged form","Sublinear term tames hypergraph matching entropy"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The proof depends on the Bennett-Bohman fixed-fraction degree estimate remaining valid with high probability for every regular completion in the iterative greedy counting, not just for the original hypergraph.","fun_headline_variants_meta":{"raw":{"variants":["Hypergraph matchings get first sublinear correction","Matching formula refined for regular hypergraphs","New count reveals missing cells in Latin squares","Kahn's conjecture proved in averaged form","Sublinear term tames hypergraph matching entropy"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.0001,"raw_usage":{"cost":0.0001003912,"is_byok":false,"cost_details":{"upstream_inference_cost":0.0001003912,"upstream_inference_prompt_cost":8.80152e-05,"upstream_inference_completions_cost":1.2376e-05},"total_tokens":1240,"prompt_tokens":1172,"completion_tokens":68,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":0}},"tokens_in":1172,"tokens_out":68,"cost_source":"openrouter_usage","duration_ms":2874,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":0,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T09:20:12.906604+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Construct a regular linear hypergraph of polynomial degree with d ≥ n^ε and Γ(H) = O($d^{{-1/k}}$ $n^{{-ε}}$) where the expected number of uncovered vertices deviates from (1+o(1)) n $d^{{-1/k}}$ by a factor larger than 1+o(1).","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the degree estimate for random greedy matching in the short greedy stages (Theorem 4.3)."},{"cited_title":"Binomial Random Matroids","cited_arxiv_id":"2603.10293","evidence_quote":"Provides the leading estimate log|M(H)| = (n/k)(log d - k + 1) + o(n) under a small pair-codegree assumption, which the present paper refines."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes sharp bounds for the matching partition function in graphs via occupancy fractions; the hypergraph extension motivates the present work."},{"cited_title":"Lee,Random matchings in linear hypergraphs, Forum Math","cited_arxiv_id":null,"evidence_quote":"Disproves Kahn's pointwise conjecture and motivates the averaged version proved here; also supplies the counterexample in Appendix A."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives leading estimates for partial Steiner systems and matroids that are refined by the present paper's correction."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}