{"id":"c65a1f87-103b-4958-9ca0-1227c7b96999","arxiv_id":"2501.16285","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For the Ulam sequence, this paper proves a_n <= 1.454^n for large n and proves that min_{1<=k<=n} a_{k+1}/a_k <= 1 + 7 log n / n for large n.","lead":"This math paper proves a new upper bound on how fast the Ulam sequence can grow, and proves that the sequence must contain some relatively small gaps. It gives rigorous tools for a famously irregular sequence where almost nothing has been proven since the 1970s.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's constant 1.454 rests entirely on an unreported length-15 word enumeration; the reported extremal word has length 13, so the computation is neither reproducible nor internally consistent.","rationale":"The reader's diagnosis matches my own: the small-gap theorem (Theorem 2) has a self-contained, plausible counting proof, while Theorem 1 depends on a finite computation that is reported without code or detailed tables. I additionally notice that the displayed extremal word has length 13, not 15, which strengthens the concern that the computation may contain a typo or be internally inconsistent. This does not change the appropriate verdict: the paper's main new result is conditional on reproducing or supplying the missing enumeration. The structural reduction to a constrained joint-spectral-radius-type problem is legitimate, and Theorem 2's proof, despite a minor notation overloading of X as potentially including the original pairs, can be read as bounding alternative representations and is arithmetically sound. Therefore the reader's CONDITIONAL verdict should stand unchanged.","tokens_in":7259,"tokens_out":27588,"duration_ms":228304,"concrete_test":"Ask the authors to supply, or independently run, an exhaustive enumeration of all admissible words of length 15 over {T1, T2, T3} with no adjacent T3, computing the spectral norm of every product and taking the 1/15-power maximum. Verify that this maximum is at most 1.4539 and identify the maximizing word(s). If the reported W = (T3T1)^3(T1T3)^3T2 has length 13, check whether it is the length-13 maximizer or a typo; in particular, confirm whether the true length-15 maximum still gives the same bound and whether the final constant 1.454 remains valid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central growth bound of Theorem 1, an ≤ 1.454^n for all sufficiently large n, is justified in Section 2 by the sentence: 'Using this with L = 15 and checking all admissible words gives C ≤ 1.4539 . . . with the extremal word being W = (T3T1)^3(T1T3)^3T2.' No enumeration procedure, code, or table of operator norms is provided, and the displayed word W has 13 letters, not the announced L = 15. Since the proof of Theorem 1 reduces the problem exactly to a finite maximum over admissible length-15 words, the numerical conclusion is not independently verifiable from the manuscript as written. This is distinct from Theorem 2, whose counting argument appears self-contained and correct. The structural reduction to the three matrices and the no-33 admissibility condition is sound, but the claimed constant 1.4539 is load-bearing and currently unsupported by any checkable artifact. A typo in the reported extremal word would be harmless if the actual computed maximum is correct, but the absence of code or data makes that impossible to confirm from the text.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the Ulam sequence a_1=1, a_2=2, ... defined by adding the smallest integer that is uniquely representable as the sum of two distinct earlier terms. It claims two results: Theorem 1 improves the known growth bound to a_n ≤ 1.454^n for all sufficiently large n, down from the previous 1.466^n bound; Theorem 2 shows that for some c>0, min_{1≤k≤n} a_{k+1}/a_k ≤ 1 + c log n / n, with c=7 admissible for large n. The proof of Theorem 2 is a self-contained counting argument: assuming all ratios up to n exceed 1+δ, the paper counts many candidate sums a_n+a_i lying in a short interval and bounds the number of alternative representations, forcing δ ≤ c log n/n. The proof of Theorem 1 reduces the growth problem to bounding products of three 4×4 matrices T1, T2, T3 with the admissibility restriction that T3 cannot appear twice consecutively, and then invokes an exhaustive check of admissible words of length 15 to obtain the numerical constant.","tokens_in":7485,"tokens_out":12816,"duration_ms":118526,"significance":"If correct, Theorem 1 is a genuine if modest improvement over the Eggleton bound, and the reduction to an admissible-word joint-spectral-radius problem is an interesting technique. Theorem 2 is the more substantial contribution: it gives the first rigorous proof that arbitrarily small ratio gaps must occur somewhere among the first n terms, and the argument is elementary, explicit, and fully checkable. The credit for the paper rests mainly on Theorem 2, whose counting proof appears sound, and on the clean structural reduction in Section 2. The main weakness is that the numerical constant in Theorem 1 is asserted on the basis of a finite enumeration that is neither described nor supplied, so Theorem 1 as written is not reproducible.","major_comments":[{"comment":"The claim 'Using this with L = 15 and checking all admissible words gives C ≤ 1.4539 . . .' is load-bearing for Theorem 1, but the manuscript provides no enumeration algorithm, no code, no table of computed norms, and no count of words checked. Moreover, the displayed extremal word W = (T3T1)^3(T1T3)^3T2 has length 13, not the announced L = 15. As written, the numerical constant 1.454 cannot be independently verified. Please supply a reproducible enumeration (code or a complete table/case analysis) and correct the length discrepancy, or explicitly state that the computation was performed with L = 13.","section":"Section 2, proof of Theorem 1"},{"comment":"The step 'We require # {1 ≤ k < n : a_n + a_k ∈ [a_n, a_n + δ/2 a_n]} ≤ X' is stated without justification. The inequality is in fact a consequence of the definitions, since each candidate sum (n,i) corresponds to a distinct pair (i,n) counted in X; however, the proof does not explain this and the sentence reads like an assumption. Please clarify the subset/injection argument so that the subsequent contradiction is transparent.","section":"Section 3, proof of Theorem 2"}],"minor_comments":[{"comment":"There is a typo in 'for some j < k ≤ n = 1'; this should presumably read 'j < k ≤ n'.","section":"Section 3"},{"comment":"The text refers to 'Figure 2', but only one figure ('Figure 1. Clumps followed by a big jump') is included; please renumber or add the missing figure.","section":"Section 3"},{"comment":"The passage from block-length estimates to 'C ≤ max_admissible_word ||...||^{1/L}' glosses over the additive constant for products whose length is not a multiple of L; the statement should be phrased as a limsup or with an explicit constant that is absorbed for large n.","section":"Section 2"},{"comment":"The statement of Theorem 2 says 'for all n ∈ N_{≥2}', while the proof only shows that c = 7 works for n sufficiently large; the finitely many small n should be checked explicitly or the theorem statement adjusted.","section":"Section 1.2 and Section 3"},{"comment":"In the proof of Lemma 1, the sentence 'we are only left with i = n − 3 which is too small' compresses the case analysis; adding a line explaining that j ≤ n−2 implies a_i + a_j ≤ a_{n−3} + a_{n−2} ≤ a_{n−1} + a_{n−2} would improve readability.","section":"Section 2, Lemma 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's main qualitative result, Theorem 2, is sound and fits the journal well. The only substantive obstacle is the unverifiable finite enumeration behind the numerical constant in Theorem 1. If the authors supply the code or a complete data table and correct the reported extremal word, I would support acceptance. If they prefer, they could present Theorem 1's numerical value as conditional on the finite check, but the current wording overstates reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The two theorems are both genuine advances. Theorem 1 improves the longstanding 1.466^n upper bound to 1.454^n, and the matrix-product reduction with the no-33 admissibility condition is a clever way to exploit the fact that Eggleton's bound cannot hold twice in a row. Theorem 2 is a clean pigeonhole/counting argument that gives an explicit c=7, and the proof appears self-contained and correct. The paper is honest about its limits: the method cannot prove a bound below 1.4146, and the small-gap statement does not rule out exponential growth.\n\nThe soft spot is where the reader and the stress-test note land: the numerical constant 1.4539 in Section 2 is obtained by \"checking all admissible words\" of length 15, but no code, pseudocode, or table of operator norms is supplied. The displayed extremal word W has only 13 letters, not 15, which looks like a transcription slip but makes the whole computation impossible to verify from the text. Since that constant is load-bearing for Theorem 1, the theorem is not independently checkable as written. This is a specific, addressable gap, not a fundamental flaw. The structural argument is sound, and the counting proof in Theorem 2 stands on its own.\n\nI largely agree with the reader's conditional verdict. The paper deserves a serious referee, but the referee should ask for the enumeration script or a table of norms for all admissible length-15 words. That is a minor-to-moderate revision, not a rejection. The citation pattern looks fine; the prior work is cited appropriately, and the new contribution is clear.\n\nWho gets value from this? Specialists in Ulam sequences and people working on growth of s-additive sequences. It is not a breakthrough that changes the big picture, but it is a solid, honest step forward on a notoriously intractable problem. I would send it to peer review, and if the computational artifact is supplied, I would happily see it published.","headline":"A real but modest improvement on the Ulam growth bound plus a small-gap theorem, with Theorem 1's constant resting on an unreported finite computation that needs to be made reproducible.","tokens_in":7974,"tokens_out":1576,"would_cite":true,"duration_ms":15902,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["11B83","15A60"],"pacs":[],"model":"deepseek-v4-flash","headline":"The Ulam sequence is forced to contain small ratio gaps, and its growth is bounded by 1.454^n.","keywords":["Ulam sequence","s-additive sequences","small gaps","growth bound","joint spectral radius","pigeonhole counting","operator norms","combinatorial number theory"],"falsifier":"Independently enumerate all admissible words of length 15 over T1, T2, T3, compute the operator norm of each product (a finite, explicit computation), and check whether the maximum 15th root is indeed ≤ 1.4539; any larger value invalidates the proof of Theorem 1.","tokens_in":7043,"feed_emoji":"🔢","tokens_out":10240,"duration_ms":77963,"temperature":0.7,"pith_summary":"The Ulam sequence (1, 2, 3, 4, 6, 8, 11, ...) is built by repeatedly appending the smallest integer that can be written as a sum of two distinct earlier terms in exactly one way. This paper proves that, for every n, at least one of the first n ratios a_{k+1}/a_k is at most 1 + c log n / n, with c = 7 for large n. It also improves the known upper bound on the nth term from 1.466^n to 1.454^n. These results are steps toward the conjecture that the sequence grows only linearly, which remains open.","feed_headline":"Ulam sequence forced to contain small ratio gaps","feed_subtitle":"A counting argument shows some ratio drops below 1 + 7 log n / n; growth also falls to 1.454^n.","key_machinery":"For the growth bound, the central machinery consists of the three linear recurrences Type I (b_{n+1} = b_n + b_{n-3}), Type II (b_{n+1} = b_{n-1} + b_{n-2}), and Eggleton (b_{n+1} = b_n + b_{n-2}), written as 4×4 matrices T1, T2, T3. Submultiplicativity of the operator norm reduces the joint spectral radius over admissible words (where Eggleton cannot be repeated consecutively) to a finite check of all admissible words of length 15, yielding the constant 1.4539. For the small-gap theorem, the load-bearing mechanism is a pigeonhole count: the assumption that all early ratios exceed 1+δ forces many candidate sums a_n + a_i to fall into a short interval, while the number of alternative representations a_j + a_k that could block them is bounded by a δ-dependent quantity; these bounds contradict each other unless δ ≤ c log n / n.","core_discovery":"The paper establishes a small-gap theorem: there is a constant c (the proof gives c = 7 eventually) such that for every n one has min_{1≤k≤n} a_{k+1}/a_k ≤ 1 + c log n / n. It also improves the best growth bound to a_n ≤ 1.454^n for all sufficiently large n. The growth bound is obtained by showing that Eggleton's recurrence cannot be applied twice in a row, leaving three possible linear recurrences whose products are controlled through submultiplicativity of operator norms and a finite check over all admissible words of length 15. The small-gap result is obtained by a counting argument that compares candidate sums a_n + a_i with alternative representations a_j + a_k.","pith_inferences":["The counting argument for small gaps likely carries over to other 1-additive sequences with subexponential growth, provided their early terms satisfy similar density estimates.","Making the length-15 word enumeration reproducible with explicit code and tables would remove the one non-transparent step in Theorem 1.","An exact computation of the joint spectral radius of the three matrices might reveal an asymptotically periodic extremal word, which would suggest a route toward subexponential growth.","The small-gap theorem offers a quantitative obstruction to 'clump-and-gap' structures, which may be useful for classifying other Ulam-like sequences."],"forward_implications":["No product of admissible words from the three recurrences grows faster than 1.454^n, so any future improvement must go beyond these recurrences.","For every sufficiently large n, some ratio a_{k+1}/a_k is below 1 + 7 log n / n, so relative gaps of size O(log n / n) are unavoidable.","If the Ulam sequence grows linearly as conjectured, this small-gap bound is sharp up to the logarithmic factor.","The spectral-radius lower bound from the word T3 T1^2 shows the present method cannot prove a_n ≤ ρ^n for any ρ < 1.4146."],"supporting_citations":[{"why":"Supplies the Eggleton bound a_{n+1} ≤ a_n + a_{n-2}, the starting point for the growth argument.","marker":"[19]"},{"why":"Introduces the joint spectral radius framework used to bound products of the three matrices.","marker":"[20]"},{"why":"Gives the spectral radius formula used to derive the lower bound from the word T3 T1^2.","marker":"[10]"}],"fun_headline_variants":["Small ratio gaps forced in Ulam sequence","Growth bound for Ulam sequence cut to 1.454^n","Ulam sequence: small gaps and tighter growth bound","Proof forces small ratio gaps in Ulam sequence","Ulam sequence: small ratios, growth ≤ 1.454^n"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The numerical factor 1.4539 in Section 2 is obtained by 'checking all admissible words' of length 15 over the three matrices T1, T2, T3, and the paper does not specify the enumeration procedure, provide code, or list the full table of operator norms; if that finite computation contains an error, the stated constant 1.454 in Theorem 1 is not established.","fun_headline_variants_meta":{"raw":{"variants":["Small ratio gaps forced in Ulam sequence","Growth bound for Ulam sequence cut to 1.454^n","Ulam sequence: small gaps and tighter growth bound","Proof forces small ratio gaps in Ulam sequence","Ulam sequence: small ratios, growth ≤ 1.454^n"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000732,"raw_usage":{"total_tokens":3234,"prompt_tokens":861,"completion_tokens":2373,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":2293}},"tokens_in":477,"tokens_out":2373,"duration_ms":14259,"temperature":1.0,"reasoning_tokens":2293,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:33:56.995414+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Independently enumerate all admissible words of length 15 over T1, T2, T3, compute the operator norm of each product (a finite, explicit computation), and check whether the maximum 15th root is indeed ≤ 1.4539; any larger value invalidates the proof of Theorem 1.","supporting_citations":[{"cited_title":"Recaman, Research Problems: Questions on a Sequence of Ulam","cited_arxiv_id":null,"evidence_quote":"Supplies the Eggleton bound a_{n+1} ≤ a_n + a_{n-2}, the starting point for the growth argument."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the joint spectral radius framework used to bound products of the three matrices."},{"cited_title":"Gelfand, Normierte ringe, Rech","cited_arxiv_id":null,"evidence_quote":"Gives the spectral radius formula used to derive the lower bound from the word T3 T1^2."}],"review_version":1}