{"id":"9791b5f9-747b-4387-be53-b55dfc44ab5c","arxiv_id":"1908.05220","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Among all nonempty subsets of {0,...,k}, the interval {1,...,k} maximizes the number of additive divisors, and the all-ones number maximizes lunar divisors in every base.","lead":"This paper proves which finite sets of numbers have the most ways to be written as a sum of two other sets, resolving conjectures about 'lunar arithmetic', a strange digit-based multiplication. The results connect a classic question in additive number theory to a playful arithmetic system and give the first proofs of three open conjectures.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 28's proof of the general-base conjecture misapplies Corollary 12 with an off-by-one; as written, Theorem 29 does not follow, though a shifted-index repair appears available.","rationale":"The paper's binary-section machinery is mostly sound: the promotion argument (Theorem 8) and the Schroeppel bijection (Theorem 11) give real support for the 0-rooted and binary claims, and the hidden monotonicity in Theorem 17 is true by a simple injection (increase the first part of a headstrong composition). I therefore do not object to the binary conclusions. The general-base claim is where the text is weakest. Theorem 28's proof contains a concrete misapplication of Corollary 12: it indexes headstrong compositions by n instead of n+1 and uses H(n+r,m) in the target divisor count, so the displayed inequalities compare quantities that are not the ones in the theorem. The small example A={1,2}, b=2, k=3 shows the proof's own formulas fail to dominate d(f). Because Theorem 29 is stated as resolving Conjecture 12 and depends directly on Theorem 28, this is the most load-bearing issue. The error appears repairable — the correct bound follows from Corollary 26, Corollary 21 iterated, and monotonicity of d([m]_b) — so a conditional verdict rather than rejection is appropriate. The reader's weakest_assumption (monotonicity in Theorem 17) is genuine but less severe; the Theorem 28 off-by-one is the decisive gap.","tokens_in":21640,"tokens_out":18005,"duration_ms":155044,"concrete_test":"Recompute Theorem 28's chain for A = {1,2}, b = 2, k = 3 using Corollary 27 and Theorem 25: d(f) = 12 and d([3]_2) = 34, whereas the proof's formulas yield 4 and 6, exhibiting the off-by-one. Then verify the repair: replace the H(n,m) sums by H(n+1,m), bound d(f) by (r+1)d([n]_b) via Corollary 26, iterate Corollary 21 to get (r+1)d([n]_b) < d([n+r]_b), and use monotonicity d([m]_b) ≤ d([m+1]_b) to reach d([k]_b); if this repaired chain checks for all r,n ≥ 1 and b ≥ 2, the theorem is recovered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is in the proof of Theorem 28, on which Theorem 29 (Conjecture 12) rests. After setting n = max(A) − r, the proof claims \"By Corollary 12 we have d(f) = (r+1)Σ_{m=1}^n H(n,m)b^m\". Corollary 12 says the number of divisors of [n] of cardinality m is H(n+1,m), not H(n,m), and it applies to the full interval [n], not to an arbitrary 0-rooted set A−{r}. The correct statement is d(f) ≤ (r+1)Σ_{m=1}^{n+1} H(n+1,m)b^m, with equality only for A−{r} = [n]. The displayed expression d([k]_b) = Σ_{m=1}^k H(n+r,m)b^m is also wrong: d([k]_b) = Σ_{m=1}^{k+1} H(k+1,m)b^m, and the comparison must involve P(n+r+1), not P(n+r). This is not cosmetic: for A = {1,2}, b = 2, k = 3, the proof's first formula gives d(f) = 2·2 = 4, while Corollary 27 gives d(f) = 2(2+4) = 12 and d([3]_2) = 34, so the displayed chain 4 ≤ 6 cannot establish the required 12 ≤ 34. The underlying theorem may still be true — one can repair the argument by shifting indices and using Corollary 26 followed by Corollary 21 iterated — but as written the proof of the general-base conjecture is incomplete.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies additive divisors: for finite A ⊆ N, d(A) is the number of sets B for which there exists C with A = B+C. The main results are that among 0-rooted subsets of [k] the interval [k] uniquely maximizes d(·) for k≠1,2 (Theorem 8); among all nonempty subsets of [k] the shifted interval {1,...,k} uniquely maximizes d(·) except for k=1,3 (Theorem 19); and that in any base b≥3 the all-ones b-ary number uniquely maximizes the number of lunar divisors among k-digit numbers (Theorem 29), resolving Conjectures 12, 13, and 14 Part I of LeBrun et al. The proofs introduce a k-promotion procedure, a sumset-to-lunar multiplication correspondence, and a multiset-array model for base-b lunar arithmetic.","tokens_in":22020,"tokens_out":4622,"duration_ms":42861,"significance":"The paper connects two previously separate worlds: additive decomposability of finite sets and lunar (dismal) arithmetic. If the identified gaps are repaired, the results would resolve three conjectures from the lunar arithmetic literature and provide a clean, largely self-contained proof mechanism that re-derives Schroeppel's theorem and Lemma 15 of LeBrun et al. rather than quoting them. The headstrong-composition triangle and its recurrence (Theorem 20) are useful in their own right. The central claims are concrete and falsifiable, and the proofs are mostly constructive, which makes the paper a solid contribution to the combinatorics of sumsets and to the theory of lunar arithmetic.","major_comments":[{"comment":"The proof of Theorem 28 contains an off-by-one error in the application of Corollary 12. Corollary 12 states that the number of divisors of [n] with cardinality m is H(n+1,m), not H(n,m), and it applies to the full interval [n], not to an arbitrary 0-rooted set A−{r}. Consequently the displayed identity d(f) = (r+1)Σ_{m=1}^n H(n,m)b^m is unjustified. The subsequent comparison d([k]_b) = Σ_{m=1}^k H(n+r,m)b^m is also incorrect; the correct expression involves H(k+1,m). A concrete check shows the failure: for A={1,2}, b=2, k=3, the proof's formula gives d(f)=4, whereas Corollary 27 gives d(f)=2(2+4)=12 and d([3]_2)=34, so the displayed chain 4 ≤ 6 cannot establish 12 ≤ 34. Since Theorem 29 depends on this comparison, the proof of Conjecture 12 is incomplete as written. A shifted-index repair using Corollary 26 followed by Corollary 21 iterated appears plausible, but it is not present in the manuscript.","section":"§7, proof of Theorem 28"},{"comment":"The induction step of Theorem 17 uses the inequality d([k−j−1]) ≤ d([k−j]) without proof or citation. Specifically, in the chain (j+2)d([k−j−1]) ≤ (j−1)d([k−j]) + 2d([k−j]) the first term requires d([k−j−1]) ≤ d([k−j]). This monotonicity is true — every additive divisor of an interval maps to an additive divisor of a larger interval — but it is not stated or proved anywhere in the paper, and the induction step of Theorem 17, which underpins Theorem 18 and hence Theorem 19, depends on it. The gap is easily filled by a one-line composition argument or a short lemma, but as written the proof is incomplete.","section":"§5, proof of Theorem 17"}],"minor_comments":[{"comment":"There are several typos and name inconsistencies: 'cojecture' in §1, 'LeBrun'/'Le Brun' used inconsistently, 'Appelgate' in the references should be 'Applegate', 'Wirsig' should be 'Wirsing', and 'Fibonacci Quaterly' should be 'Fibonacci Quarterly'.","section":"Throughout"},{"comment":"The convention S+∅ = ∅ in multisumset addition is a deliberate choice, but it deserves more discussion because it departs from the usual set-plus-empty-set convention and it is essential for the descending-chain representation to be a monoid. A reader might reasonably expect S+∅ = S for all S under the set-array interpretation.","section":"§7, Definition preceding Theorem 22"},{"comment":"The notation M_b, M_k, and M^b_k is introduced but the compound M^b_{\\le k} used in Theorems 28 and 29 is not formally defined; it should be defined explicitly as the union of M^b_{\\ell} for ℓ ≤ k.","section":"§7, notation"},{"comment":"The phrase 'for all k ≥ min(4,j)' in the induction statement is confusing. Since the induction is on j, the intended meaning is that the strict inequality holds for all k ≥ 4 (with small j handled separately); the formula should be rewritten for clarity.","section":"§5, proof of Theorem 17"}],"recommendation":"major_revision","confidential_remarks":"The off-by-one issue in Theorem 28 is genuine and was independently identified in the stress-test scenario. It is not a matter of taste: the displayed formula contradicts the paper's own Corollary 27 and Lemma 15. I believe the theorem is true and repairable, but the current manuscript cannot claim to resolve Conjecture 12 without the repair. The missing monotonicity in Theorem 17 is a smaller fix but should be addressed explicitly rather than left implicit. Overall the paper is promising and the core ideas are sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper proves something real: the binary lunar divisor conjectures from LeBrun et al. are settled, and the sumset-to-lunar correspondence is a genuinely useful dictionary. The promotion procedure in Section 2 is the key new idea and it works. On my reading, Theorems 8 and 19 stand up; the reasoning is clear and the headstrong-composition bijection gives a nice handle on the counts.\n\nThe soft spots are in the general-base portion. Theorem 28, which is the load-bearing step for Conjecture 12, has a real off-by-one. Corollary 12 says the number of divisors of [n] of cardinality m is H(n+1,m), not H(n,m). In the proof of Theorem 28, after setting n = max(A) − r, the paper claims d(f) = (r+1)Σ_{m=1}^n H(n,m)b^m. That's not right even for A−{r} = [n]; the correct sum is over H(n+1,m). The expression for d([k]_b) is also off: it should be Σ_{m=1}^{k+1} H(k+1,m)b^m, not Σ_{m=1}^k H(n+r,m)b^m. The example A={1,2}, b=2, k=3 makes this concrete: the proof's formula gives 4, true d(f)=12, d([3]_2)=34. So Theorem 29 does not follow as written. The repair is in reach: use d(f) ≤ (r+1)Σ_{m=1}^{n+1} H(n+1,m)b^m and then iterate Corollary 21 with shifted indices. But the correction is not cosmetic, and the referee should ask for it.\n\nThere's a smaller gap in Theorem 17: the step d([k-j-1]) ≤ d([k-j]) is used without proof. That monotonicity is true (headstrong compositions of m embed into those of m+1 by increasing the first part), so it is a minor omission, not a flaw.\n\nWho is this for? People working on additive combinatorics of finite sets, and fans of lunar/dismal arithmetic. The binary results are solid and deserve to be in the literature. The general-base theorem is probably true, but the proof as written is incomplete. I'd recommend engaging with it, and sending it to review, but the referee should insist on fixing the off-by-one before acceptance.","headline":"Solid binary-case results; the general-base proof has a repairable off-by-one.","tokens_in":22484,"tokens_out":10591,"would_cite":true,"duration_ms":77022,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["11P70","05A17"],"pacs":[],"model":"deepseek-v4-flash","headline":"The interval $\\{1,\\ldots,k\\}$ uniquely maximizes the number of additive divisors among all nonempty subsets of $\\{0,\\ldots,k\\}$, except at $k=1$ and $k=3$.","keywords":["additive divisors","sumsets","lunar arithmetic","headstrong compositions","restricted compositions","generalized Fibonacci numbers","inverse additive number theory","multisets"],"falsifier":"Brute-force count $d(A)$ for every nonempty subset $A$ of $\\{0,1,\\ldots,6\\}$: if any set other than $\\{1,\\ldots,6\\}$ has $d(A)\\ge d(\\{1,\\ldots,6\\})$, then Theorem 19 is false. This is a finite computation, since $d(A)$ counts the distinct $B\\subseteq\\mathbb{N}$ for which some $C\\subseteq\\mathbb{N}$ satisfies $A=B+C$.","tokens_in":21435,"feed_emoji":"🌙","tokens_out":11888,"duration_ms":99816,"temperature":0.7,"pith_summary":"This paper asks which finite set of integers is the most reducible: among all nonempty subsets of $\\{0,\\ldots,k\\}$, which one can be written as a sumset $A+B$ in the greatest number of ways? The paper proves that the shifted interval $\\{1,\\ldots,k\\}$ is the unique winner for every $k$ except $k=1$ and $k=3$, where the maximum is also attained elsewhere. Because sumset addition on binary digit sets corresponds exactly to lunar multiplication (the carry-free digitwise-min multiplication), the same theorem says that the binary number $2^k-2$ has more lunar divisors than every other $k$-digit binary number, uniquely except for $n=2$ and $n=4$. The argument is then lifted to arbitrary bases by replacing sets with multisets, proving that in every base $b \\geq 3$ the all-ones repunit $(b^k-1)/(b-1)$ uniquely maximizes the number of lunar divisors among $k$-digit numbers. These results resolve the conjectured maxima for binary lunar divisors and for lunar divisors in every base $b\\ge 3$, connecting an inverse question in additive number theory to a digitwise arithmetic system.","feed_headline":"Interval {1,…,k} wins the additive-divisor race over {0,…,k}.","feed_subtitle":"In binary lunar arithmetic this says 2^k − 2 has the most divisors among k-digit numbers, and the same holds in all bases.","key_machinery":"The argument is carried by two linked objects. The first is $k$-promotion: for a set $A\\subseteq[k]$ containing $0$, every factorization $A=B+C$ is transformed into a factorization of the full interval $[k]$, either by appending the missing elements of $[k]$ to $C$ or by shifting them by $\\max(B)$. This produces disjoint families of divisors, so $[k]$ has at least as many divisors as any $0$-rooted set $A$, and the construction is strict enough to force uniqueness. The second object is the bijection between divisors of $[k]$ and headstrong compositions of $k+1$, compositions whose first part is at least as large as every other part; these are counted by the generalized Fibonacci numbers $F(n,k)$. Lemmas 13 and 14 convert properties of these numbers into the inequalities $2d([k-1])\\geq d([k])$ and $3d([k-2])\\leq 2d([k-1])$ that let $\\{1,\\ldots,k\\}$ dominate all other subsets. For general bases the machinery widens to arrays of nested subsets: a multiset with maximum multiplicity $b$ is written as $b$ nested sets, multisumset addition is coordinatewise sumset addition, and a divisor of a plain set $A$ contributes $b^{\\mathrm{card}\\,B}$ divisors of the corresponding multiset.","core_discovery":"The central discovery, stated on the paper's own terms, is that the interval $[k+]=\\{1,\\ldots,k\\}$ maximizes the additive-divisor count $d(A)=\\#\\{B\\subseteq\\mathbb{N}:\\exists C,\\ A=B+C\\}$ among all nonempty subsets of $[k]=\\{0,\\ldots,k\\}$, and is the unique maximizer except for $k=1$ and $k=3$. In binary lunar arithmetic this is the statement that $n=2^k-2$ has strictly more divisors than any other $k$-digit binary number, the unique maximum outside $\\{2,4\\}$. The route is a two-stage reduction: first, $k$-promotion shows that the full interval $[k]$ is the unique maximum among sets containing $0$; second, the identity $d(A)=(\\min A+1)d(A-\\{\\min A\\})$ reduces sets without $0$ to sets with $0$, and inequalities on headstrong compositions show $2d([k-1])>d([k])$, so $\\{1,\\ldots,k\\}$ outscores every $0$-rooted set. The same machinery, with multisets in place of sets, proves the base-independent statement: in every base $b \\geq 3$ the repunit $(b^k-1)/(b-1)$ uniquely maximizes the number of base-$b$ lunar divisors among $k$-digit numbers.","pith_inferences":["The paper leaves the runner-up problem open: for odd $k$-digit binary numbers, which number has the second-largest divisor count? The headstrong-composition triangle in Section 6 gives a plausible route, since ranking compositions by number of parts is exactly what controls the divisor counts.","The multiset correspondence suggests a testable extension: define additive divisors directly for finite multisets of integers and ask whether every multiset of height $b$ is dominated by the $b$-truncated interval; this would recast the base-$b$ theorem as a statement about all multisets.","The $k$-promotion construction is tailored to the interval $[k]$, so it is not obvious whether the maximizing set for an arbitrary ambient interval $\\{a,\\ldots,b\\}$ is still a translate of an initial interval; a small computational experiment for a few asymmetric intervals would show whether the paper's strongest conclusion survives outside the symmetric setting."],"forward_implications":["For every $k\\ne 1,3$, $d(\\{1,\\ldots,k\\})$ is strictly larger than $d(A)$ for every other nonempty $A\\subseteq\\{0,\\ldots,k\\}$.","In binary lunar arithmetic, $2^k-2$ is the unique $k$-digit number with the maximal number of lunar divisors whenever $k\\ne 1,3$.","In every base $b\\ge 3$, the all-ones number $(b^k-1)/(b-1)$ is the unique $k$-digit maximum for the lunar divisor count.","Among odd $k$-digit binary numbers, $111\\ldots111$ (that is, $2^k-1$) is the unique maximizer of $d_2$, as an immediate corollary of the $0$-rooted case."],"supporting_citations":[{"why":"Introduces lunar arithmetic, formulates the conjectures on maximal divisors, and supplies the lemma $d(A)=(r+1)d(A-\\{r\\})$ that reduces non-rooted sets.","marker":"[1]"},{"why":"Provides independent proofs and combinatorial interpretations of the correspondence between $d_2(111\\ldots1)$ and restricted compositions, supporting Theorem 11.","marker":"[5]"},{"why":"Supplies the generating functions and asymptotic estimates for headstrong compositions used to connect divisor counts to generalized Fibonacci numbers in Sections 4 and 6.","marker":"[10]"}],"fun_headline_variants":["Additive divisors: {1,…,k} outranks {0,…,k}","Lunar arithmetic: {1,…,k} wins the divisor race","The set {1,…,k} tops {0,…,k} in divisor count","Maximal additive divisors: {1,…,k} beats {0,…,k}","In all bases, {1,…,k} has the most lunar divisors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of Theorem 17 assumes without proof that enlarging an interval by one element cannot decrease its number of additive divisors; if $d([k-j-1])$ ever exceeded $d([k-j])$, that induction step would collapse.","fun_headline_variants_meta":{"raw":{"variants":["Additive divisors: {1,…,k} outranks {0,…,k}","Lunar arithmetic: {1,…,k} wins the divisor race","The set {1,…,k} tops {0,…,k} in divisor count","Maximal additive divisors: {1,…,k} beats {0,…,k}","In all bases, {1,…,k} has the most lunar divisors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00088,"raw_usage":{"total_tokens":3883,"prompt_tokens":1102,"completion_tokens":2781,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":718,"completion_tokens_details":{"reasoning_tokens":2679}},"tokens_in":718,"tokens_out":2781,"duration_ms":20614,"temperature":1.0,"reasoning_tokens":2679,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:22:46.082533+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Brute-force count $d(A)$ for every nonempty subset $A$ of $\\{0,1,\\ldots,6\\}$: if any set other than $\\{1,\\ldots,6\\}$ has $d(A)\\ge d(\\{1,\\ldots,6\\})$, then Theorem 19 is false. This is a finite computation, since $d(A)$ counts the distinct $B\\subseteq\\mathbb{N}$ for which some $C\\subseteq\\mathbb{N}$ satisfies $A=B+C$.","supporting_citations":[{"cited_title":"Dismal Arithmetic","cited_arxiv_id":null,"evidence_quote":"Introduces lunar arithmetic, formulates the conjectures on maximal divisors, and supplies the lemma $d(A)=(r+1)d(A-\\{r\\})$ that reduces non-rooted sets."},{"cited_title":"On the sequence A079500 and its c ombinatorial interpretations","cited_arxiv_id":null,"evidence_quote":"Provides independent proofs and combinatorial interpretations of the correspondence between $d_2(111\\ldots1)$ and restricted compositions, supporting Theorem 11."},{"cited_title":"Compositions with parts c ontrained by the leading summand","cited_arxiv_id":null,"evidence_quote":"Supplies the generating functions and asymptotic estimates for headstrong compositions used to connect divisor counts to generalized Fibonacci numbers in Sections 4 and 6."}],"review_version":1}