{"id":"fd6bdf8b-8c48-42b5-9da0-065d902d6577","arxiv_id":"1908.05358","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The moments of weighted Cantor measures obey an infinite-product generating function, decay no faster than a power when the largest digit has positive weight, and can be estimated to uniform error epsilon in near-linear time.","lead":"This paper derives explicit product formulas, decay rates, and a fast approximation algorithm for the moments of weighted Cantor measures, a family of fractal probability distributions on the unit interval. The results give researchers a faster way to compute moment sequences and to know how quickly those sequences shrink.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The advertised O(log log(1/ε) m log m) algorithm bound is not uniform in m: Theorem 3.6 forces k=O(log(m^{3/2}/ε)), adding an O(log log m) factor to the loop count.","rationale":"The paper's main moment results are mostly well supported: Theorem 3.6's Cauchy-integral proof checks out, Theorem 3.9's tail-mass lower bound appears correct, and the infinite product formula in Theorem 3.4 is sound. The weakest load-bearing point is the algorithm complexity claim in Remark 3.8, because it is one of the three advertised contributions and the stated asymptotic is simply not extractable from Theorem 3.6 when m varies. The reader's weakest_assumption identifies exactly this nonuniformity, so I agree with the reader's diagnosis. One additional in-scope issue is that Theorem 2.9 appears false: for β = (1/2+δ, 1/2-δ) converging to α = (1/2,1/2), each μ_β is singular with respect to Lebesgue measure, so the total variation distance to μ_α remains 1; the proof's Riemann-Stieltjes approximation step is also dimensionally invalid. However, Section 3 does not rely on Theorem 2.9, so I do not treat that as the load-bearing objection to the central moment and algorithm claims. The correct response is to keep the verdict conditional: the moment estimates and lower bound can stand, but the complexity statement and the Section 2 overclaim must be corrected or qualified.","tokens_in":17222,"tokens_out":18810,"duration_ms":198701,"concrete_test":"Fix ε = 10^{-6} and N = 2. For m = 2^{10}, 2^{20}, and 2^{30}, compute the required truncation level k = ceil(log_N(e m sqrt(m-1)/ε)) from Theorem 3.6 and count the number of FFT polynomial multiplications in Remark 3.8, namely ceil(log_2 k). If this count grows with m (approximately as log log m) while the claimed O(log log(1/ε)) remains constant, the bound is not uniform in m. Analytically, substitute k = O(log(m^{3/2}/ε)) into the pseudocode loop count; the resulting complexity is O(log log(m/ε)·m log m), which reduces to O(log log(1/ε)·m log m) only for bounded m.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Remark 3.8 claims that the first m moments can be estimated to uniform error ε in O((log log(1/ε))·m log m). To ensure |I_m - I_{m;k}| ≤ ε, Theorem 3.6 requires k ≥ log_N(e m sqrt(m-1)/ε), i.e. k = O(log(m^{3/2}/ε)). The pseudocode in Remark 3.8 performs log_2(k) truncated degree-m polynomial multiplications, each costing O(m log m) via FFT. The resulting complexity is therefore O((log log(1/ε) + log log m)·m log m), not O(log log(1/ε)·m log m), unless m is held fixed while ε tends to 0. The abstract does not state such a fixed-m regime; as a two-variable asymptotic claim, the stated bound is not justified. The underlying error estimate in Theorem 3.6 and the moment approximation itself appear sound, but the complexity contribution as advertised is overstated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies α-weighted Cantor measures μ^α supported on [0,1], defined by the IFS φ_n(x)=(x+n)/N with weights α∈Δ_N. It derives a recurrence for the moments I_m (Theorem 3.2), an infinite product formula for the Laplace and moment generating functions (Theorem 3.4), an explicit error bound for moment approximations from truncated products (Theorem 3.6), an algorithm for estimating the first m moments (Remark 3.8), and one-sided decay estimates for I_m (Theorem 3.9). Section 2 contains auxiliary results on the distribution function, symmetry, continuity of μ^α in the weight vector, and a two-term recursion for Legendre polynomials. The main moment-theoretic claims are derived from definitions and cited results, with no parameter fitting or circularity.","tokens_in":17332,"tokens_out":8877,"duration_ms":84006,"significance":"If the moment-theoretic results are correct, the paper provides a useful toolkit: the recurrence (8) is explicit, Theorem 3.4 gives a compact product representation of the MGF, Theorem 3.6 gives a quantitative convergence rate, and Theorem 3.9 shows how the largest weight α_{N-1} controls the polynomial decay of moments, matching a known Grabner–Prodinger asymptotic. The algorithm in Remark 3.8 is clearly described and would be a practical contribution. However, the paper currently contains a false theorem on total-variation continuity (Theorem 2.9) and an unsupported complexity claim in the abstract and Remark 3.8; these issues must be corrected before the paper can be published.","major_comments":[{"comment":"The proof of Theorem 2.9 is invalid and the statement is false. The displayed estimate uses sums of the form Σ_P (F_μβ(x_{j+1}) − F_μβ(x_j))Δx_j, which are not Riemann–Stieltjes sums for μβ; the measure of an interval is Σ_P (F(x_{j+1}) − F(x_j)), with no Δx_j factor. Consequently the derived bound |μβ(I_n) − μα(I_n)| ≤ ε λ(I_n) does not follow. More seriously, the claimed total-variation convergence is false: for N=2, take α=(1/2,1/2), so μα is Lebesgue measure, and β=(p,1−p) with p→1/2. For p≠1/2, μβ is singular with respect to Lebesgue (its binary digits have limiting frequency p), so ‖μβ−μα‖_TV = 2 for every p≠1/2. The correct and useful statement is weak convergence, which follows directly from Proposition 2.8 via uniform convergence of the CDFs. The theorem and its proof should be replaced by this corrected statement, or the theorem should be removed if not needed later.","section":"Theorem 2.9"},{"comment":"The complexity claim O((log log(1/ε))·m log m) is not justified as a two-variable asymptotic. Theorem 3.6 gives |I_m − I_{m;k}| ≤ e m√(m−1)/N^k. To guarantee uniform error ≤ ε one needs k = O(log(m^{3/2}/ε)) = O(log m + log(1/ε)). The algorithm in Remark 3.8 performs log_2(k) truncated products of degree-m polynomials, so the total cost is O((log log(1/ε) + log log m)·m log m). The stated bound in the abstract and Remark 3.8 therefore overstates the dependence on m. The authors should either state explicitly that the complexity claim is for fixed m as ε→0, or revise the claimed bound to include the log log m term.","section":"Remark 3.8 / Abstract"},{"comment":"The abstract's statement that the paper 'characterize[s] precisely when the moments I_m exhibit either polynomial or exponential decay' overstates Theorem 3.9. That theorem proves only one-sided bounds: an upper bound I_m ≤ ((N−1)/N)^m when α_{N−1}=0, and a lower bound I_m ≥ C(α)m^{−γ} otherwise. It does not provide matching lower bounds in the first case or upper bounds in the second, so the decay regime is not fully classified. Please rephrase the abstract to describe these one-sided estimates accurately.","section":"Abstract"}],"minor_comments":[{"comment":"The use of the 'left-endpoint approximation' as a lower bound relies on the integrand x^m being increasing; this monotonicity should be stated explicitly before Corollary 3.3, since the inequality 0 ≤ I_m − (approximation) is otherwise not immediate.","section":"Section 3, before Eq. (6)"},{"comment":"In the paragraph after Proposition 2.11, 'Theorem 2.11' should be 'Proposition 2.11'.","section":"End of Section 2"},{"comment":"There are several typographical errors: 'Gram Schmidt' should be 'Gram-Schmidt', 'immmmediately' should be 'immediately', and 'uniquess' should be 'uniqueness'.","section":"Throughout"},{"comment":"The text attributes Theorem 2.3 to 'Pei' and cites [7], but the reference list gives 'Hsu E. P.'; please reconcile the author name in the text and the bibliography.","section":"Theorem 2.3 and References"},{"comment":"For the shifted measure ν^α, Remark 3.7 defines partial products H_{α;k}(s) but does not state an error bound analogous to Theorem 3.6. Since the algorithm is also claimed for the J_m, it would be helpful to state explicitly that the same kind of error estimate holds for the shifted moments.","section":"Remark 3.7"}],"recommendation":"major_revision","confidential_remarks":"The false Theorem 2.9 is a serious issue, but it appears isolated from the moment-theoretic core, so the paper is likely repairable by replacing that theorem with the correct weak-convergence statement. The complexity claim in Remark 3.8 also needs to be corrected, and the abstract should be adjusted. The paper fits the journal's scope and the main moment computations appear sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read it. First, the core results on moments—the error bound in Theorem 3.6, the fast product algorithm in Remark 3.8 (with one caveat below), and the polynomial lower bound in Theorem 3.9—are new and mostly correct. Second, Theorem 2.9 is not merely unproved; it is false as stated, and the error is not in a detail nobody will check.\n\nThe good part: the infinite product formula for the Laplace transform is standard but cleanly derived, and they are honest about crediting Jorgensen–Kornelson–Shuman for the moment recurrence. The error bound |I_m - I_{m;k}| <= e m sqrt(m-1)/N^k is a solid improvement over the naive truncation estimate, and the algorithm built on it is a real contribution to the orthogonal-polynomial side of the subject. The lower bound C(a) m^{-γ} is also a nice complement to Grabner–Prodinger, and the proof via optimizing f(k) is elementary and works.\n\nNow the soft spots, in order of seriousness.\n\n1. Theorem 2.9 says β→a implies convergence in total variation. That cannot hold for non-uniform weights: take N=2, a=(p,1-p), β=(q,1-q) with p≠q. The corresponding measures are supported on sequences of zeros and ones whose limiting frequency is p and q respectively, so by the strong law of large numbers they are mutually singular and the TV distance is 2 for all p≠q. The proof also uses an invalid Riemann–Stieltjes approximation—it mixes increments of the CDF with mesh lengths, which does not converge to the measure. Since nothing later uses Theorem 2.9, the clean fix is to delete the TV claim and keep the (true) weak-continuity statement in Proposition 2.8.\n\n2. The abstract promises a precise characterization of polynomial versus exponential decay. Theorem 3.9 gives only one-sided bounds: exponential upper bound when a_{N-1}=0, polynomial lower bound otherwise. That is not a precise characterization. Reword.\n\n3. The advertised O(log log(1/ε) m log m) complexity in Remark 3.8 is not uniform in m. Theorem 3.6 forces k = O(log(m^{3/2}/ε)), so the loop count is O(log k) = O(log log m + log log(1/ε)), and the stated bound holds only for fixed m. The stress-test note is right on this; it's an overstatement but easily corrected.\n\nWho's this for? People computing moments or running Gram–Schmidt on fractal L^2 spaces. They'll use the algorithm despite the wording issues. I'd send it to a referee—the Section 3 results deserve expert scrutiny, and the referee can demand the Section 2 fix.","headline":"Worth refereeing on the strength of the moment estimates, but only after the authors delete or fix the false total-variation continuity claim in Theorem 2.9.","tokens_in":17897,"tokens_out":7150,"would_cite":true,"duration_ms":66768,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["28A25","28A80"],"pacs":[],"model":"deepseek-v4-flash","headline":"For weighted Cantor measures, moments decay polynomially unless the last weight is zero, with rate set by $\\alpha_{N-1}$.","keywords":["Cantor measure","moments","orthogonal polynomials","generating function","iterated function system","fractal measure","moment decay","Laplace transform"],"falsifier":"Take $N=3$ and $\\alpha=(1/2,0,1/2)$, so $\\gamma=\\log_3 2$; compute the exact moments from the recurrence $I_m=(3^m-1)^{-1}\\sum_{n=0}^2\\alpha_n\\sum_{i=0}^{m-1}\\binom{m}{i}n^{m-i}I_i$ and check whether $\\liminf_{m\\to\\infty}m^\\gamma I_m$ is zero, which would contradict the theorem's lower bound $I_m\\ge C(\\alpha)m^{-\\gamma}$.","tokens_in":1852,"feed_emoji":"🧮","tokens_out":5714,"duration_ms":122694,"temperature":0.7,"pith_summary":"This paper studies self-similar probability measures on $[0,1]$ built from a base-$N$ subdivision and weights $\\alpha_0,\\ldots,\\alpha_{N-1}$. It determines exactly when the moments $I_m=\\int x^m\\,d\\mu^\\alpha$ decay exponentially and when they decay at a polynomial rate, and it gives an explicit infinite-product formula for their generating function together with a fast algorithm for estimating the first $m$ moments. A reader should care because these moments are what you need to construct orthogonal polynomial systems on the fractal support of such measures, and the polynomial decay rate matches earlier asymptotics for the classical Cantor distribution.","feed_headline":"Weighted Cantor moments: polynomial decay unless last weight vanishes","feed_subtitle":"The decay rate is set by the last weight, and a fast product formula estimates all first m moments.","key_machinery":"The central object is the moment generating function $G_\\alpha(z)=\\sum_{m\\ge0}I_m z^m/m!$, which Theorem 3.4 represents as the infinite product $\\prod_{r=1}^\\infty\\sum_{n=0}^{N-1}\\alpha_n\\exp(nz/N^r)$. Truncating this product at $k$ factors yields coefficients $I_{m;k}$, and the Cauchy estimate in Theorem 3.6 bounds the coefficient error by $e m\\sqrt{m-1}/N^k$; repeated squaring and truncation of the finite product give the computing algorithm. The decay theorem comes from measuring the tail mass $\\mu^\\alpha[1-1/N^k,1]=(\\alpha_{N-1})^k$ and optimizing the resulting lower bound $I_m\\ge(1-N^{-k})^m(\\alpha_{N-1})^k$ over $k$.","core_discovery":"Theorem 3.9: if every weight is less than $1$ and the last weight $\\alpha_{N-1}$ is zero, then $I_m\\le((N-1)/N)^m$; if $\\alpha_{N-1}>0$, there is a constant $C(\\alpha)>0$ with $I_m\\ge C(\\alpha)m^{-\\gamma}$ for all $m\\ge1$, where $\\gamma=\\log_N(1/\\alpha_{N-1})$. The paper also proves that the moment generating function is the entire infinite product $\\prod_{r=1}^\\infty\\sum_{n=0}^{N-1}\\alpha_n\\exp(nz/N^r)$, and that retaining only the first $k$ factors of this product gives coefficients differing from the true moments by at most $e m\\sqrt{m-1}/N^k$. For palindromic weights, the same machinery gives exponentially small moments for the measure shifted to $[-1/2,1/2]$.","pith_inferences":["An implication left implicit is that the advertised $O(\\log\\log(1/\\varepsilon)\\,m\\log m)$ complexity is valid when $m$ is fixed; if $m$ grows with $\\varepsilon^{-1}$, the truncation level required by Theorem 3.6 also grows with $m$.","A testable extension is to fit the observed decay rate of empirical moments and compare it with $\\log_N(1/\\alpha_{N-1})$ to identify an unknown self-similar weight vector.","Combining the moment estimates with the Legendre recursion of Proposition 2.11 gives a concrete way to compute orthonormal polynomial bases on non-spectral fractals.","The exponential bound for the shifted palindromic case suggests recentering improves the conditioning of moment-based computations on fractal measures."],"forward_implications":["If the last weight is zero, the support lies in $[0,(N-1)/N]$, so every moment is bounded by $((N-1)/N)^m$.","If the last weight is positive, the $m$-th moment cannot drop below an absolute constant times $m^{-\\log_N(1/\\alpha_{N-1})}$, the same power order found for classical Cantor distributions.","Truncating the moment generating function after $k$ factors gives uniform error $O(m^{3/2}/N^k)$ for the first moments, so each extra factor buys another factor of $N$ in accuracy.","The first $m$ moments can be estimated to uniform error $\\varepsilon$ in $O(\\log\\log(1/\\varepsilon)\\,m\\log m)$ time by repeated squaring and truncation.","For palindromic weights, the shifted measure has moments bounded by $2^{-m}$, making its orthogonal polynomial coefficients easier to control."],"supporting_citations":[{"why":"Supplies the existence and uniqueness theorem for the self-similar measure defined by the iterated function system, on which all later statements rest.","marker":"[1]"},{"why":"Earlier source of the moment recurrence relation that Theorem 3.2 reproves and that underlies the recursive computations.","marker":"[2]"},{"why":"Provides the differentiability, Hölder, and singularity properties of the distribution functions used to characterize the weighted Cantor measures in Section 2.","marker":"[7]"},{"why":"Supplies the regularity theorem for Borel measures used in proving total-variation continuity of the map from weights to measures.","marker":"[8]"},{"why":"Gives the asymptotic moment expansion for Cantor-type distributions whose leading power order Theorem 3.9 matches.","marker":"[10]"}],"fun_headline_variants":["Cantor moment decay: last weight decides polynomial vs exponential","Weighted Cantor moments: zero last weight speeds decay","Moment decay on Cantor sets tied to final weight","Fast moment bounds for weighted Cantor measures","Last weight rules Cantor moment decay: polynomial or exponential"],"cache_read_input_tokens":20096,"weakest_assumption_plain":"The algorithmic complexity claim assumes the truncation level $k$ can be fixed from the error tolerance $\\varepsilon$ alone, whereas the proved error bound ties $k$ to the largest moment index $m$; if $m$ and $\\varepsilon$ vary together, the stated speed is not uniform.","fun_headline_variants_meta":{"raw":{"variants":["Cantor moment decay: last weight decides polynomial vs exponential","Weighted Cantor moments: zero last weight speeds decay","Moment decay on Cantor sets tied to final weight","Fast moment bounds for weighted Cantor measures","Last weight rules Cantor moment decay: polynomial or exponential"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000526,"raw_usage":{"total_tokens":2583,"prompt_tokens":1031,"completion_tokens":1552,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":1474}},"tokens_in":647,"tokens_out":1552,"duration_ms":11104,"temperature":1.0,"reasoning_tokens":1474,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:17:09.537436+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $N=3$ and $\\alpha=(1/2,0,1/2)$, so $\\gamma=\\log_3 2$; compute the exact moments from the recurrence $I_m=(3^m-1)^{-1}\\sum_{n=0}^2\\alpha_n\\sum_{i=0}^{m-1}\\binom{m}{i}n^{m-i}I_i$ and check whether $\\liminf_{m\\to\\infty}m^\\gamma I_m$ is zero, which would contradict the theorem's lower bound $I_m\\ge C(\\alpha)m^{-\\gamma}$.","supporting_citations":[{"cited_title":"E., Fractals and self-similarity, Indian Univ","cited_arxiv_id":null,"evidence_quote":"Supplies the existence and uniqueness theorem for the self-similar measure defined by the iterated function system, on which all later statements rest."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier source of the moment recurrence relation that Theorem 3.2 reproves and that underlies the recursive computations."},{"cited_title":"P., A class of singular continuous functions, Elem","cited_arxiv_id":null,"evidence_quote":"Provides the differentiability, Hölder, and singularity properties of the distribution functions used to characterize the weighted Cantor measures in Section 2."},{"cited_title":"I., Measure theory","cited_arxiv_id":null,"evidence_quote":"Supplies the regularity theorem for Borel measures used in proving total-variation continuity of the map from weights to measures."},{"cited_title":"J., Prodinger H., Asymptotic analysis of the moments of the Cantor distribution, Statist","cited_arxiv_id":null,"evidence_quote":"Gives the asymptotic moment expansion for Cantor-type distributions whose leading power order Theorem 3.9 matches."}],"review_version":1}