{"id":"9bf3e724-23f7-4eb1-a534-c1a0f4a2b726","arxiv_id":"2506.01896","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A new family of U sets gives the best known lower bound theta >= 1.173077 for the sums and differences of finite integer sets.","lead":"The paper builds specially crafted sets of integers where the difference set grows much faster than the sum set, yielding a new record lower bound of 1.173077 for the sum-difference exponent. The improvement is tiny, but it demonstrates a new large-deviation technique for counting constrained simplexes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Binomial factor estimate (5) is false as stated: C(m,k) is not W(m,k,1), and for ar>1/2 the claimed rate log2−I(ar,1) overstates the true entropy H(ar); Theorem 1 does not exclude this regime.","rationale":"The reader identified exactly the false equality C(m,k)=W(m,k,1), and I agree that this is a real defect in the proof. I mark agreement as partial because the equality is not automatically fatal: for ar<1/2 the rate log2−I(ar,1) equals H(ar), so the binomial estimate can be repaired by a short monotonicity argument. The load-bearing issue is that the manuscript never proves or even states that its numerical maximizer operates in the ar<1/2 regime, and no optimizing parameters are reported. My own hand checks of the displayed formula suggest there are admissible parameters (e.g., B=5, r≈0.82, a=0.5, giving ar=0.41) for which the formula yields a larger value than the paper's table, so the claim may well be true; but the proof as written cannot be checked at the exact point used for the headline constant. The MATLAB output is also not a rigorous certificate, so the appropriate verdict remains CONDITIONAL, as the reader concluded. I would not move to REJECT because the mathematical strategy is plausible and the specific flaw appears repairable; I would not move to ACCEPT because the central estimate (5) is false as stated and the numerical supremum is uncertified. Hence no change to the reader's verdict.","tokens_in":5753,"tokens_out":48255,"duration_ms":472587,"concrete_test":"Patch the MATLAB code in §4 so the first addend in term(r,B,a,eps) is the binary entropy H(a*r), the exact rate of log C(m,floor(a*r*m))/m, instead of logW(a*r,1); then record the optimizing r,a for each B and recompute Table 1. If the B=5 optimum has a*r<1/2 and still attains θ−1 ≥ 0.173077, the false equality in (5) is repairable and the headline constant is safe. If the optimizer moves to a*r≥1/2, or the corrected value drops below 0.173077, Theorem 1 as written does not establish the claimed bound. A stronger version would also print an explicit (B,r,a) certificate and verify the lower bound with interval arithmetic or rational estimates.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 1 lower-bounds d(U) by one term containing the binomial coefficient C(m,k). To estimate that coefficient, the manuscript asserts: in the set W(m,k,1), exactly k of the m coordinates are 1, and the rest are 0, so C(m,k)=W(m,k,1). This equality is false: W(m,k,1) is the cumulative set of binary vectors with at most k ones, so |W(m,k,1)| = sum_{j≤k} C(m,j). As a result, inequality (5) claims liminf log C(m,floor(ar m))/m ≥ log2 − I(ar,1), but the true rate is the binary entropy H(ar). The two agree when ar≤1/2; for ar>1/2, the manuscript's estimate is log2, which is strictly larger than H(ar). The proof imposes no restriction ar<1/2 and does not report the optimizing r,a used for the headline B=5 table, so the derivation of (7) and the claimed bound θ≥1.173077 are not supported as written. A repair exists if the numerical maximizer stays in ar<1/2: for c<ar<1/2, C(m,k) ≥ C(m,cm) ≥ |W(m,cm,1)|/(cm+1), yielding the same liminf. But this replacement is absent, and the required condition ar<1/2 is never verified. The uncertified MATLAB supremum is a secondary concern; the false counting equality is the concrete mathematical defect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims a new lower bound θ ≥ 1.173077 for the sums and differences of sets problem, improving on the previous best θ = 1.173050. The construction takes U to be the image of W(m, L, B) = {x ∈ [0,B]^m : Σ x_i ≤ L} under the base-(2B+1) encoding, and uses large-deviation estimates (Cramér's theorem) to asymptotically count |U+U| and |U−U|. The resulting lower bound is then optimized numerically over parameters B, r, and a using MATLAB routines fminbnd and fminunc. A table reports optimized values for B = 3,...,10, with the B = 5 entry supporting the headline constant 1.173077.","tokens_in":6117,"tokens_out":6873,"duration_ms":72948,"significance":"If Theorem 1 were established rigorously, the paper would give the best known lower bound for the exponent θ in the sums and differences problem, a small but genuine improvement over the recent construction of Gerbicz. The use of large-deviation estimates to count the constrained integer partitions W(m,L,B) is a natural and potentially useful addition to the existing framework, and the manuscript includes its MATLAB code for reproducibility. However, the proof as written contains a false combinatorial counting step that invalidates the key binomial estimate, and the numerical maximization is not presented as a rigorous computation. The overall approach appears salvageable, but the central claim is not supported in the current form.","major_comments":[{"comment":"The assertion 'in the set W(m,k,1), exactly k of the m coordinates are 1, and the rest are 0, so C(m,k) = W(m,k,1)' is false: by the definition W(m,L,B), the set W(m,k,1) consists of all binary vectors with at most k ones, not exactly k ones. The correct identity is C(m,k) = |W(m,k,1)| − |W(m,k−1,1)|. As a consequence, the claimed asymptotic lower bound (log C(m,k))/m ≥ log 2 − I(ar,1) is only valid for ar ≤ 1/2; for ar > 1/2 the true rate is the binary entropy H(ar), which is strictly smaller than log 2. The proof imposes no restriction ar ≤ 1/2 and does not report the optimizing parameters, so the derivation of Eq. (7) and the conclusion θ ≥ 1.173077 are not supported as written. A repair exists: for c < ar ≤ 1/2 one can use C(m,k) ≥ C(m,⌊cm⌋) ≥ |W(m,⌊cm⌋,1)|/(⌊cm⌋+1), which yields the same rate; this replacement and the condition ar ≤ 1/2 must be made explicit.","section":"Section 3, Table 1"},{"comment":"The final numerical constant 1.173077 is produced by applying fminbnd and fminunc with finite tolerances (ϵ down to 10^{-10}) to a nonlinear optimization problem. The manuscript provides no interval-arithmetic bounds, no certified global maximum, and no reporting of the maximizing triple (B,r,a). Even if the counting error in Eq. (5) were repaired, this numerical computation would still not constitute a rigorous proof of the displayed inequality. To make Theorem 1 rigorous, the authors should either exhibit one explicit triple (B,r,a) for which the expression exceeds 1.173077 and evaluate it with a certified method, or provide rigorous enclosures for the supremum.","section":"Section 3, Table 1"},{"comment":"In the case r < B/2, the proof states lim_{m→∞} (1/m) log P(x_1 + ... + x_m ≤ mB/2) = −I(r,B). This is not the event defining W(m,⌊rm⌋,B); the correct statement should be P(x_1 + ... + x_m ≤ mr). As written, the displayed identity is false unless r = B/2. This appears to be a typographical slip, but Lemma 1 is used throughout the proof of Theorem 1, so the intended statement and its derivation must be corrected.","section":"Section 2, Lemma 1 proof"}],"minor_comments":[{"comment":"The line 'Adding (5), (4) and (6) we get lim inf_{m→∞} d(U)/m ≥ ...' should read 'lim inf_{m→∞} (log d(U))/m ≥ ...'; as written the dimensions do not match the preceding estimates.","section":"Section 3, Eq. (7)"},{"comment":"In the function theta(B), the fourth tolerance is set to 1e-4, but the text and Table 1 indicate it should be 1e-10.","section":"Section 4, MATLAB code"},{"comment":"The code comment describes a function logW(c,B,eps), but the actual function definition is 'function ret = I(c, B, eps)'. Since the main routine calls logW, the code as listed would not run unless an additional alias is defined.","section":"Section 4, MATLAB code"},{"comment":"The citation for Cramér's theorem appears as '[ ?]' and should be '[3]'.","section":"Section 2, proof of Lemma 1"},{"comment":"Typos: 'supreme' should be 'supremum'; 'yetrelevant' should be 'yet relevant'; the phrase 'marked in blue' refers to colors not present in the printed text.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The false counting step is a concrete mathematical error that is likely fixable with a standard binomial estimate, and the paper's asymptotic framework is plausible. The more serious barrier to acceptance is the lack of a rigorous numerical certification for the headline constant; journals in this area typically expect either an explicit parameter choice or a certified computation. The manuscript would also benefit from reporting the optimizing parameters (B,r,a) for the final value."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real news is that you get a slightly better lower bound for the sum-difference exponent, 1.173077, building on the prior record of 1.173050. The move is to add individual coordinate bounds to the earlier construction and then use large-deviation estimates to count the constrained simplexes. That is a genuinely new ingredient for this problem, and the paper ships its MATLAB code, so the numerical part is at least reproducible. The framework is not circular: the parameters B, r, a are chosen after the fact to maximize a legitimate lower bound, which is fine.\n\nThe main problem is at inequality (5). The paper asserts that C(m,k) equals W(m,k,1), on the grounds that W(m,k,1) contains exactly the vectors with exactly k ones. That is false. W(m,k,1) is the cumulative set of binary vectors with at most k ones. So the claimed rate log2 − I(ar,1) is only valid when ar ≤ 1/2; for ar > 1/2 the true rate is the binary entropy H(ar), and the paper's estimate is too large. The proof never verifies that its numerical maximizer stays in ar < 1/2 and does not report the optimizing r and a for the headline B=5 row. As written, the derivation of (7) and of Theorem 1 is not supported.\n\nThe good news is that this looks repairable. For any c < ar < 1/2, C(m,k) ≥ C(m,cm) and C(m,cm) is at least |W(m,cm,1)|/(cm+1), so the same limit holds with the necessary condition ar < 1/2. If the true maximizer satisfies that, the final constant survives. The secondary issue is the uncertified MATLAB supremum, but the stability across tolerances down to 1e-10 is encouraging. The proof of Lemma 1 has a small gap for the r ≥ B/2 case (the lim inf argument only gives one direction), but that can be fixed and is not load-bearing.\n\nThis paper deserves a serious referee. The core approach is plausible and the computational evidence is honestly presented. I'd ask for a rigorous treatment of the binomial factor, a statement of the verified parameter regime, and ideally interval-arithmetic bounds on the supremum. You might also gently point out that the MATLAB code has a typo in theta_eps's fourth tolerance call.\n\nIn short: read it, send it back for revision, and if the ar < 1/2 condition checks out, the result is a legitimate if incremental improvement.","headline":"The claimed bound is likely right, but the proof as written contains a false counting equality that must be repaired before the theorem can stand.","tokens_in":6597,"tokens_out":1238,"would_cite":false,"duration_ms":14357,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["11B75"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims a new lower bound θ ≥ 1.173077 for the sums-and-differences exponent, improving the previous best known value 1.173050.","keywords":["sumset","difference set","additive combinatorics","theta exponent","large deviations","constrained integer partitions","coordinate encoding"],"falsifier":"Evaluate the asserted identity at $m=5,k=2$: $\\binom{5}{2}=10$ whereas $|W(5,2,1)|=\\sum_{j=0}^{2}\\binom{5}{j}=16$, so the equality used in the proof is not true; checking whether any correct lower bound on the binomial factor of the same exponential strength exists would settle the proof of the stated constant.","tokens_in":5582,"feed_emoji":"➕","tokens_out":11369,"duration_ms":115894,"temperature":0.7,"pith_summary":"This paper claims a new lower bound, $\\theta \\geq 1.173077$, for the exponent that controls how large a difference set must be relative to a sumset for finite integer sets. The number improves on all earlier bounds, including the previous explicit value $1.173050$ and the earlier machine-search value $1.1584$. The route is to build explicit sets from coordinate-restricted vectors, count their sums and differences using large-deviation estimates rather than exact enumeration, and then optimize a parameter formula numerically. If the proof is correct, this is the current best known lower bound for the sums-and-differences problem.","feed_headline":"New sum-difference record: θ = 1.173077","feed_subtitle":"Large-deviation counting of restricted vectors improves the best lower bound for integer sumset growth.","key_machinery":"The central object is $W(m,L,B)$, the set of $m$-dimensional nonnegative integer vectors whose entries are each at most $B$ and whose sum is at most $L$. The proof uses the identities $|U+U|=|W(m,2L,2B)|$ and $|U-U|=\\sum_{k=0}^{\\min(m,L)} \\binom{m}{k}|W(k,L-k,B-1)|\\,|W(m-k,L,B)|$, together with the large-deviation rate function $I(c,B)=\\sup_t\\{tc-\\log((1+e^t+\\cdots+e^{Bt})/(B+1))\\}$, which controls the exponential growth of $|W(m,\\lfloor rm\\rfloor,B)|$. These ingredients convert the finite-set construction into the parameter formula that is optimized to yield $\\theta \\geq 1.173077$.","core_discovery":"The paper's central claim is Theorem 1: for the sums and differences of sets problem, the attainable exponent satisfies $\\theta \\geq 1.173077$. Following an established construction line, the author takes the set $W(m,L,B)$ of $m$-tuples of nonnegative integers with each coordinate at most $B$ and coordinate sum at most $L$, encodes the tuples as integers via a base-$(2B+1)$ expansion, and calls the image $U$. Known identities express $|U+U|$ and $|U-U|$ through counts of such restricted-vector sets; the new ingredient is a large-deviation analysis that gives the exponential growth rate of $|W(m,\\lfloor rm\\rfloor,B)|$ for fixed $B$ and varying $r$. Splitting the difference-set count into a binomial factor and two $W$-counts, and optimizing the resulting formula numerically, the paper obtains the stated constant, with the maximum occurring at $B=5$.","pith_inferences":["Extension: the flawed step identifying $\\binom{m}{k}$ with $|W(m,k,1)|$ might be repairable by a direct lower bound on the binomial coefficient from standard tail or entropy estimates; a correct replacement could preserve the constant, but the paper does not supply it.","Extension: the numerical optimization was restricted to $B=1,\\ldots,10$, $r\\in(0.5,2)$, and $a$ in a fixed interval, so a wider search over these parameters could push the constant higher; $1.173077$ is a lower bound for this construction family, not a proven maximum over all possible choices.","Extension: the same large-deviation counting of coordinate-bounded vectors should transfer to other additive-combinatorics problems where sums and differences of structured integer sets are measured, such as higher-dimensional analogues or sets defined by several linear constraints."],"forward_implications":["For any small $\\varepsilon>0$, the constructed sequence gives finite sets $A,B$ with $|A-B|\\geq c|A+B|^{1.173077-\\varepsilon}$ and $|A+B|$ arbitrarily large, subject to the growth constraint in the problem.","The value $1.173077$ supersedes the prior explicit bound $1.173050$ and all earlier bounds, making it the best known lower bound if the proof is correct.","The numerical table shows the maximum at $B=5$ is stable to fifteen decimal places as the optimization tolerance is tightened, so the reported constant is not an artifact of loose numerical search.","The construction yields a sequence of sets of growing size, so the improved exponent is obtained in the limit rather than from a single finite set."],"supporting_citations":[{"why":"Supplies the previous explicit construction at $1.173050$ and the identities for $|U+U|$ and $|U-U|$ that the proof reuses.","marker":"[1]"},{"why":"Defines the sums-and-differences problem and proves the exponent formula $\\theta \\geq 1+\\log(|U-U|/|U+U|)/\\log(2\\max(U)+1)$ that converts set-size counts into exponent bounds.","marker":"[2]"},{"why":"Provides the large-deviation principle invoked to compute the rate function $I(c,B)$ for sums of independent uniform variables.","marker":"[3]"},{"why":"Describes the earlier machine-discovered construction at $1.1584$ that this paper's bound improves upon.","marker":"[4]"}],"fun_headline_variants":["New θ record: 1.173077 via large-deviation counting","Sums and differences: θ bound now 1.173077","AlphaEvolve toppled: θ=1.173077 in sumset problem","Large-deviation trick lifts θ lower bound to 1.173077","Sum-difference sets: best θ yet at 1.173077"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of the final bound relies on treating the binomial coefficient $\\binom{m}{k}$ as equal to $|W(m,k,1)|$, the number of binary vectors with at most $k$ ones; that equality is false, and the paper supplies no substitute that gives the required lower bound on the binomial factor.","fun_headline_variants_meta":{"raw":{"variants":["New θ record: 1.173077 via large-deviation counting","Sums and differences: θ bound now 1.173077","AlphaEvolve toppled: θ=1.173077 in sumset problem","Large-deviation trick lifts θ lower bound to 1.173077","Sum-difference sets: best θ yet at 1.173077"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001012,"raw_usage":{"total_tokens":4202,"prompt_tokens":802,"completion_tokens":3400,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":418,"completion_tokens_details":{"reasoning_tokens":3302}},"tokens_in":418,"tokens_out":3400,"duration_ms":24301,"temperature":1.0,"reasoning_tokens":3302,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:32:58.563536+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the asserted identity at $m=5,k=2$: $\\binom{5}{2}=10$ whereas $|W(5,2,1)|=\\sum_{j=0}^{2}\\binom{5}{j}=16$, so the equality used in the proof is not true; checking whether any correct lower bound on the binomial factor of the same exponential strength exists would settle the proof of the stated constant.","supporting_citations":[{"cited_title":"Gyarmati, F","cited_arxiv_id":null,"evidence_quote":"Defines the sums-and-differences problem and proves the exponent formula $\\theta \\geq 1+\\log(|U-U|/|U+U|)/\\log(2\\max(U)+1)$ that converts set-size counts into exponent bounds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the earlier machine-discovered construction at $1.1584$ that this paper's bound improves upon."}],"review_version":1}