{"id":"f290e39d-7ffd-42e3-be7b-cb6a1ad4fa63","arxiv_id":"2607.21589","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper proves the Frankl–Tokushige product conjectures for r-cross-intersecting uniform and biased families, with the common 1-star attaining the sharp bound.","lead":"An extremal-set-theory conjecture by Frankl and Tokushige is proved: for r cross-intersecting families of k_i-subsets of an n-set, the product of normalized sizes is at most the product of k_i/n, and the analogous biased-measure inequality follows. The proof combines a random-partition coupling, an upper-shadow inequality, and a new analytic optimization theorem.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Most load-bearing concern: Lemma 3.3's endpoint cases (k=1, ℓ=n−1) are handled by an unproven pointwise limit; if (3) fails there, Lemma 3.4(5) and Theorem 1.1 collapse.","rationale":"The reader's weakest assumption identifies Lemma 3.3, and I agree. I examined the proof of Lemma 3.3 and found the interior cases logically coherent; the principal gap is the endpoint passage, which is not rigorous as written. I also scanned Theorem 4.1, especially Lemma 6.1 and Claim 1; the arguments are intricate but no internal contradiction surfaced. Corollary 1.2's lifting argument is standard. The weak-hypothesis claim in Remark 3.2 is unsupported but not used in the main theorem. Since the central claim depends on Lemma 3.3, a concrete endpoint verification is the most economical way to settle the remaining uncertainty. The verdict remains ACCEPT with moderate confidence; the concern is significant enough to warrant verification but not, on the current evidence, to change the verdict.","tokens_in":18776,"tokens_out":50041,"duration_ms":365190,"concrete_test":"Implement a high-precision numerical checker for Lemma 3.3. For n=3 to 15, all 1≤k<ℓ<n (specifically including k=1 and ℓ=n−1), and a dense grid of (x,y) with 0≤x≤y≤1, evaluate the left-hand side and right-hand side of (3) using the definition of Φ with endpoint conventions (2), and assert LHS≤RHS with a small tolerance. Also verify symbolically (e.g., with sympy) the two identities in Case 3: (1−p)y+px=(1−p)[(1−p0)+p0(A+C)] and (1−p1)(A+C−1)=((1−p)y+px−(1−p))/p. Then test the pointwise-limit claim by computing (3) along sequences k→1+ and ℓ→(n−1)−. If no violation appears, run a randomized version for larger n (up to 50) to increase confidence; any violation would pinpoint a counterexample and identify where the induction breaks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 3.3 is the two-point recombination inequality used as the induction step in the proof of the star-calibrated upper-shadow comparison (Lemma 3.4). Lemma 3.4, in turn, provides the pivot estimate (7) in Theorem 1.1: it bounds μ_{qn}(∂_{k_j→qn}A_j)−q by the minimum of two star-calibrated terms. If Lemma 3.3 fails in any parameter range, Lemma 3.4(5) fails, and the analytic hypothesis needed for Theorem 4.1 is not established. The proof of Lemma 3.3 explicitly assumes 1<k<ℓ<n−1 and then states that the cases k=1 and ℓ=n−1 are 'pointwise limits' in (2). This is plausible but not demonstrated. The endpoint profiles Φ_{0,t} and Φ_{s,1} are discontinuous at z=1, and those boundary cases do occur in the application: the pivot family may have k_j=1, and the target level qn may equal n−1. A failure of (3) at a boundary would break the induction in Lemma 3.4 and invalidate the proof of Theorem 1.1. The manuscript gives no rigorous ε-δ or monotone-convergence argument for the limit passage, and the Case 3 algebraic identities (e.g., (1−p)y+px=(1−p)[(1−p0)+p0(A+C)]) are stated without derivation, so a subtle sign error could easily be missed. All other parts of the proof, including the critical coupling (Lemma 3.1) and the analytic Theorem 4.1, are detailed and internally consistent; this makes Lemma 3.3 the most fragile link.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves the Frankl–Tokushige product conjectures for r-cross-intersecting families. Theorem 1.1 states that for r ≥ 2, 0 ≤ k_i ≤ (r−1)n/r, and r-cross-intersecting families F_i ⊆ C([n], k_i), one has ∏ μ_{k_i}(F_i) ≤ ∏ k_i/n, with equality attained by the corresponding levels of a common 1-star. Corollary 1.2 transfers the statement to p_i-biased product measure for p_i ≤ (r−1)/r. The proof has three main ingredients: an ordered-partition coupling giving sharp additive inequalities at every critical level vector (Lemma 3.1); a star-calibrated upper-shadow comparison (Lemma 3.4) proved by induction on n, with the induction step reduced to a two-point inequality (Lemma 3.3); and an analytic optimization theorem (Theorem 4.1) that converts the asymmetric additive constraint into the desired product bound. The paper also recovers the equal-level theorem of Frankl–Tokushige and discusses extensions to r-cross t-intersection.","tokens_in":19227,"tokens_out":39822,"duration_ms":312066,"significance":"If the proof is completed, this settles two long-standing conjectures in full generality, and the method is attractive: Lemma 3.1 isolates essentially all combinatorial content in a sharp linear constraint, the upper-shadow comparison is a slice analogue of Bellman-function arguments, and the analytic theorem is a nontrivial replacement for the symmetric AM–GM step. The equality cases and parameter ranges are discussed carefully, and the biased theorem follows from a clean lifting argument. The result would be a major contribution to extremal set theory. However, the proof has a load-bearing gap in the endpoint cases of the two-point inequality, as detailed below; once that is repaired, the paper should be acceptable.","major_comments":[{"comment":"The proof of Lemma 3.3 begins by saying that the cases k=1 and ℓ=n−1 are 'the pointwise limits in (2)' and that the inequalities below pass to those limits. This is not a proof. The endpoint profiles Φ_{0,t} and Φ_{s,1} are discontinuous at z=1, and the arguments inside the Φ terms in (3) can equal 1 in the boundary cases, so pointwise convergence alone does not automatically preserve the inequality. These endpoint cases are not decorative: in the induction in Lemma 3.4, a pivot family may have k_j=1, and the target level may be qn=n−1, so Lemma 3.3 is invoked exactly at k=1 or ℓ=n−1. If (3) fails at any boundary point, Lemma 3.4(5), and hence the pivot estimate (7) used in Theorem 1.1, is unproved. Please supply a direct verification of the two endpoint cases, or state and prove Lemma 3.3 for real k,ℓ with a rigorous limit argument that handles the discontinuities at z=1 explicitly.","section":"§5, Lemma 3.3; used in §3.4, Eq. (5) and §4, Eq. (7)"}],"minor_comments":[{"comment":"The 'more importantly' statement that Theorem 1.1 remains valid under a weaker hypothesis (the critical-sum constraint replacing cross-intersection) is asserted without proof. It is not used later, but as written it is a substantive claim. Either provide a proof or explicitly label it as a conjecture/expected consequence.","section":"§3.2, Remark 3.2"},{"comment":"Several algebraic identities are stated as 'direct substitution' without derivation, e.g. the two displayed identities following the Case 3 split and the penultimate equality in Case 2. They are verifiable, but a line or two of explanation would help the reader and reduce the risk of a hidden sign error.","section":"§5, proof of Lemma 3.3, Case 3"},{"comment":"The definition p = 1−q + (j−2)/r is easy to misread as 1−q + j − 2/r. Use a displayed equation with parentheses: p = 1 − q + (j−2)/r.","section":"§4, proof of Theorem 4.1"},{"comment":"Typo: 'level vectors level vectors' is duplicated. Also, the phrase 'the exact bound is attained by the corresponding levels of a common 1-star' is repeated in the abstract and introduction; this is fine but should be checked for redundancy.","section":"§1, first paragraph"},{"comment":"The endpoint profiles Φ_{0,t} and Φ_{s,1} are defined by pointwise limits but the limits are not computed. Since these profiles are used in the shadow induction, a short explanation of the convergence and the interpretation as empty/full sections would be helpful.","section":"§2, Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The main architecture of the proof appears sound and the result is very significant. The one genuinely load-bearing issue is the unproved endpoint passage in Lemma 3.3; it is likely repairable by a direct case check or a careful continuity argument, but it must be fixed before publication. The unproved weaker-hypothesis claim in Remark 3.2 should also be resolved by either proof or downgrading to a remark. I would be willing to look at a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper settles the uniform and biased Frankl–Tokushige product conjectures for r-cross-intersecting families in full generality. That is the headline, and it is a real result. The novelty is not in doubt: the unequal-level uniform bound and the unequal-bias measure bound are new, and the old proofs for equal levels or equal biases did not contain the asymmetric pivot construction or the analytic Theorem 4.1. The proof strategy is coherent: a random-partition coupling converts cross-intersection into a sharp additive constraint, a star-calibrated upper-shadow comparison controls densities level by level, and a separate analytic theorem turns the resulting asymmetric bound into the product bound. I read the central line carefully and did not find a load-bearing error. Lemma 3.1 is elegant and clearly correct; Lemma 3.4 is detailed; Theorem 4.1 is intricate but the optimization argument checks out as far as I pushed it.\n\nThere are two soft spots, and I think they are both real but minor. First, Remark 3.2 asserts without proof that Theorem 1.1 remains valid under a much weaker measure hypothesis, and it is presented as an \"in fact\" claim. That is stronger than a passing comment; either it should be proved or the wording should be softened to a conjecture. It is not used in the main proof, so it does not threaten the result, but a referee should not let it stand as stated.\n\nSecond, the proof of Lemma 3.3, the two-point inequality that underpins the star-calibrated shadow comparison, handles the endpoint cases k=1 and ℓ=n−1 by saying they are \"pointwise limits\" of the profiles in (2). The profiles Φ_{0,t} and Φ_{s,1} are discontinuous at z=1, and the manuscript gives no epsilon-delta argument or monotone-convergence justification for passing the inequality to those limits. The endpoint cases genuinely occur in the application. I believe the inequality is still true there, and the omission is probably fixable, but as written it is a gap in the most fragile link of the proof. The rest of Section 5 is a long case analysis with several algebraic identities stated without derivation; I did not find an error, but it is exactly the kind of place where a subtle sign mistake could hide.\n\nCitation pattern is fine. The self-citation to [7] is for the equal-bias case, which is genuinely prior and directly relevant. The equal-level theorem [20] is properly credited. No fitted constants or circular normalizations appear.\n\nWho gets value from this paper? Anyone working on intersection problems or product extremal bounds. It deserves a serious referee, not a desk rejection, and I would send it out with a request to prove or qualify the Remark 3.2 claim and to supply a rigorous endpoint argument in Lemma 3.3. After those repairs, the proof should be publishable in a strong journal.","headline":"A serious, largely convincing resolution of both Frankl–Tokushige product conjectures; the main chain is sound, but two small gaps (unproved weaker-hypothesis claim in Remark 3.2 and the endpoint passage in Lemma 3.3) should be fixed before publication.","tokens_in":19675,"tokens_out":1782,"would_cite":true,"duration_ms":20934,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05D05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves the Frankl–Tokushige product conjecture: for r-cross-intersecting uniform families with levels at most (r−1)n/r, the product of normalized sizes is at most the product of the levels over n, attained by a common 1-star.","keywords":["r-cross-intersecting families","Frankl–Tokushige product conjecture","upper shadow","star extremal configuration","biased product measure","ordered-partition coupling","extremal set theory","product bound"],"falsifier":"Search all pairs 0≤x≤y≤1 for small adjacent slice parameters (e.g., n=5, k=2, ℓ=3, and n=4, k=1, ℓ=2) and test whether the two-point inequality (3) holds; a single violation would invalidate Lemma 3.3 and the induction. Independently, an exhaustive check over all r-cross-intersecting uniform families for small n (say n=6, r=3, k_i∈{0,1,2,3,4}) would test the product bound directly.","tokens_in":18692,"feed_emoji":"⭐","tokens_out":8800,"duration_ms":82811,"temperature":0.7,"pith_summary":"This paper establishes the Frankl–Tokushige product conjecture for r-cross-intersecting uniform families. The statement is that whenever one selects r families—each family consisting of k_i-element subsets of an n-element set, with every choice of one set from each family sharing at least one element—the product of their normalized sizes is at most (k_1/n)⋯(k_r/n). The corresponding levels of a common one-point star attain this bound, so the star is the extremal shape over the full allowed range k_i≤(r−1)n/r. The same sharp bound is transferred to biased product measures on all subsets, for biases at most (r−1)/r. A sympathetic reader should care because this closes a conjecture that had only been settled in equal-level and special asymmetric cases, and it does so with a method that reduces the intersection condition to a single additive coupling bound.","feed_headline":"Common star maximizes product in r-cross-intersecting families","feed_subtitle":"One set from each family always shares a point, and their normalized product cannot beat the common star's value.","key_machinery":"The key machinery is the ordered-partition coupling (Lemma 3.1), which turns r-cross-intersection into a family of additive inequalities over all level vectors summing to (r−1)n, together with a star-calibrated upper-shadow comparison (Lemma 3.4) whose induction step is a two-point inequality (Lemma 3.3). The comparison profile Φ_{s,t}(z) is a piecewise power function interpolating between z^{log s/log t} and its complement; the two-point inequality says that after restricting a family along one coordinate and recombining the two sections, the profile is preserved. This gives a smooth directed isoperimetric estimate normalized so that a 1-star is the extremal shape, and the final analytic st","core_discovery":"The central claim is Theorem 1.1: for r≥2 and 0≤k_i≤(r−1)n/r, any r-cross-intersecting families F_i⊆binom([n],k_i) satisfy the product inequality ∏μ_{k_i}(F_i)≤∏k_i/n, with equality attained by the corresponding levels of a fixed 1-star. The argument has three linked parts. First, an ordered random partition of [n] shows that for any target levels ℓ_i with sum (r−1)n, the densities satisfy ∑μ_{ℓ_i}(F_i)≤r−1; this is the only place the cross-intersection hypothesis enters. Second, a star-calibrated upper-shadow comparison, proved by induction through a two-point inequality, controls the density of a family raised to a target level relative to the star density. Third, an analytic theorem for o","pith_inferences":["Editorial extension: the proof's reduction suggests a transfer principle—any families satisfying the level-sum constraints, even without an intersection condition, should obey the same product bound; one could test this by constructing non-intersecting families that satisfy the constraints.","Editorial extension: a stability theorem should hold near the star: families whose normalized product is close to (k_1⋯k_r)/n^r should be close to a common 1-star in normalized symmetric difference, since the shadow comparison is strict away from the star.","Editorial extension: the ordered-prefix pivot used to choose target levels might be adapted to other unbalanced product problems in extremal set theory, where unequal parameters block the simple AM–GM step."],"forward_implications":["The common 1-star is extremal over the full uniform range k_i≤(r−1)n/r; no mixed construction can beat the product of the star's levels.","The biased version holds: for any r-cross-intersecting families of all subsets with p_i≤(r−1)/r, the product of their p_i-biased measures is at most p_1⋯p_r.","Both parameter ranges are best possible—crossing the threshold (r−1)n/r or (r−1)/r makes the full level or large-subset family violate the inequality.","The proof reproduces the earlier equal-level Frankl–Tokushige theorem as a special case, using only the additive coupling.","The coupling extends formally to r-cross t-intersecting families for t≥2, with a missing shadow-comparison step identified as the open part."],"fun_headline_variants":["Common star wins product bound for r-cross-intersecting families","For r-cross families, one common point gives maximum product","Sharp product bound proven for r-cross-intersecting families","Frankl-Tokushige conjectures settled by star extremal product"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is the two-point comparison inequality (Lemma 3.3) used in the shadow induction; if it failed for any adjacent slice parameters, including the endpoint cases k=1 and ℓ=n−1, the star-calibrated upper-shadow bound and the analytic hypothesis of the final theorem would no longer be available.","fun_headline_variants_meta":{"raw":{"variants":["Common star wins product bound for r-cross-intersecting families","For r-cross families, one common point gives maximum product","Sharp product bound proven for r-cross-intersecting families","Frankl-Tokushige conjectures settled by star extremal product"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001255,"raw_usage":{"total_tokens":5064,"prompt_tokens":916,"completion_tokens":4148,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":4077}},"tokens_in":660,"tokens_out":4148,"duration_ms":28761,"temperature":1.0,"reasoning_tokens":4077,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T06:59:11.795897+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Search all pairs 0≤x≤y≤1 for small adjacent slice parameters (e.g., n=5, k=2, ℓ=3, and n=4, k=1, ℓ=2) and test whether the two-point inequality (3) holds; a single violation would invalidate Lemma 3.3 and the induction. Independently, an exhaustive check over all r-cross-intersecting uniform families for small n (say n=6, r=3, k_i∈{0,1,2,3,4}) would test the product bound directly.","supporting_citations":[],"review_version":1}