{"id":"7f7aae26-887a-43dd-87ad-b2045843e8c6","arxiv_id":"2608.22423","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A generalized Cantor function has local scaling exponent α at a point exactly when the accumulated log-mass to log-length quotient converges to α and extreme-digit runs grow sublinearly, which yields an exact dimension-distortion formula on a full-measure set.","lead":"Generalized Cantor functions are the distribution functions of self-similar Cantor measures. This paper gives a complete necessary and sufficient condition for such a function to scale like a power of distance at a support point, and derives an exact Hausdorff dimension distortion formula on a full-measure subset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Invalid squeeze in the converse of Theorem 2.1 is a real but repairable proof gap; the conditional verdict stands.","rationale":"The central claim is the iff criterion in Theorem 2.1. Its sufficiency is solid: the correction from z is at most (R_n+2)(-log p_min), which is o(B_n) under R_n/n->0. The converse has two parts; the first, deriving A_n/B_n->alpha from the assumed local limit, is the load-bearing step because the endpoint-run sublinearity argument assumes it. The proof's displayed squeeze is wrong: the lower bound is too weak and is not implied by z>=p_min and w>=delta. This is a genuine gap in the written proof. However, the ratio structure makes the repair immediate: numerator and denominator are B_n(1+o(1)) times A_n/B_n and 1, respectively, so Q->alpha. Thus I see no counterexample to the theorem; the concern is a proof gap, not a false claim. The reader's rationale also names this false inequality, though their weakest_assumption field identifies strong separation instead. Strong separation is an explicit standing hypothesis, not a hidden weakness. Hence the conditional verdict remains unchanged.","tokens_in":11788,"tokens_out":24959,"duration_ms":220742,"concrete_test":"Re-derive the converse estimate in Theorem 2.1 from the defining ratio Q=(A_n - log z)/(B_n - log w), using z>=p_min and w>=delta. Verify the corrected squeeze A_n/(B_n - log delta) <= Q <= (A_n - log p_min)/B_n and check that both bounds converge to alpha as n->infty. If the corrected squeeze holds, the necessity of (7) follows; if not, Theorem 2.1 requires a new argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the converse direction of Theorem 2.1 (Section 3, after constructing y_n), the proof needs lim Q(x,y_n)=alpha to conclude A_n/B_n -> alpha. The displayed inequality A_n/B_n - log delta <= Q(x,y_n) <= (A_n - log p_min)/B_n does not follow from the preceding bounds and, even if true, would not squeeze to alpha: the lower bound tends only to alpha - log delta. The correct estimates follow directly from Q=(A_n - log z)/(B_n - log w) with 0<p_min<=z<=1 and 0<delta<=w<=1, giving A_n/(B_n - log delta) <= Q <= (A_n - log p_min)/B_n, and both bounds tend to alpha. Thus the converse is incomplete as written, but the gap is local and repairable; the endpoint-run contradiction and the sufficiency direction are not affected.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the pointwise local scaling of distribution functions of self-similar Cantor measures on the real line, under a strong separation condition on the underlying IFS. The main theorem (Theorem 2.1) characterizes, for a point x=π(ω) in the attractor K, the existence of the limit of log|F(y)-F(x)|/log|y-x| as y→x in K in terms of the convergence of A_n(ω)/B_n(ω) to a prescribed α>0 and the sublinear growth of the endpoint runs R_n(ω)/n. A second theorem (Theorem 2.2) constructs a full H^s-measure subset M of K on which the typical exponent is the cross-entropy-to-Lyapunov ratio h(q,p)/χ(q), and derives the exact dimension-distortion formula dim_H F(A) = (χ(q)/h(q,p)) dim_H A for all A⊂M. A three-branch example is used to show that the endpoint-run condition is necessary. The derivation is mostly symbolic and uses standard tools such as the strong law of large numbers, Borel-Cantelli, and Hutchinson's theorem.","tokens_in":11909,"tokens_out":36105,"duration_ms":317127,"significance":"If the proof gap discussed below is fixed, the paper would be a valuable contribution: it gives a genuinely pointwise, if-and-only-if characterization of local scaling for generalized Cantor functions with arbitrary contraction ratios and probability weights, and it connects the pointwise exponent to an exact Hausdorff-dimension distortion formula on a full-measure subset. The paper ships no fitted parameters, the main assertions are falsifiable through the explicit symbolic criterion, and the three-branch example is a useful concrete test of the necessity of the endpoint-run condition. The proofs are largely self-contained, with the external dimension-distortion lemma from [7] used transparently as a tool.","major_comments":[{"comment":"The displayed two-sided estimate for Q(x,y_n) after the construction of y_n contains an incorrect lower bound. From Q = (A_n - log z)/(B_n - log w) together with z ≥ p_min and δ ≤ w ≤ 1, the valid lower bound is Q ≥ A_n/(B_n - log δ), not A_n/B_n - log δ. As printed, the lower bound tends to α - log δ and cannot be used to conclude A_n/B_n → α from Q(x,y_n) → α. The corrected lower bound together with the existing upper bound (A_n - log p_min)/B_n both tend to α, and a simple rearrangement gives A_n/B_n → α, so the conclusion is recoverable. This is a local, repairable gap, but the formula and the subsequent sentence 'Taking limits on both sides' must be corrected.","section":"Section 3, proof of Theorem 2.1 (converse)"}],"minor_comments":[{"comment":"The phrase 'after moving each n_j backward to the beginning of the corresponding endpoint run' is confusing: the construction requires n_j to be placed immediately before the run, so that ω_{n_j} ≠ 1 in the first case and ω_{n_j} ≠ m in the second. Please clarify this index shift, since it is essential for the definition of η^(j).","section":"Section 3, endpoint-run contradiction"},{"comment":"The implication μ((x,y]) = 0 ⇒ (x,y) ∩ K = ∅ is used twice but not justified. Add a sentence explaining that if a point of K lay in (x,y), then a sufficiently small cylinder of the IFS with positive μ-mass would also lie in (x,y), giving μ((x,y]) > 0.","section":"Section 4, proof of Theorem 2.2"},{"comment":"In the computation of z(x_2, \\tilde y_k), the value T(21) = p_1 is used implicitly. This follows from Lemma 3.1 together with T(1) = 0, but T(1) = 0 (where 1 denotes the constant coding 111...) is never stated explicitly; please state it at first use.","section":"Section 5, second subsequence"},{"comment":"The title as extracted contains an evident typo, 'DISTOR TION'; please correct it to 'DISTORTION' in the final version.","section":"Title"},{"comment":"Several displayed formulas mix inline text and fraction notation in a way that is hard to parse in the extracted version (for example, the lower bound discussed in the major comment). Please ensure that all fractions are typeset unambiguously.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The only substantive issue I find is the local proof gap in the converse of Theorem 2.1; it is a repairable error and does not appear to undermine the main theorem's validity. The paper is otherwise well structured and the results are of clear interest to the fractal-geometry community. I recommend major revision, not rejection, with the understanding that the corrected lower bound A_n/(B_n - log δ) restores the squeeze argument. If the authors intended the lower bound to be A_n/(B_n - log δ) all along, the formula in the submitted version is at best misleadingly typeset and should be rewritten."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a genuinely useful paper, not a minor variation on classical Cantor-function results. The pointwise criterion in Theorem 2.1—the local scaling exponent exists and equals α iff the bulk quotient A_n/B_n converges to α and the endpoint runs grow sublinearly—is new for arbitrary contraction ratios and probability weights, and the proof strategy of separating bulk from endpoint corrections is the right way to see the problem. The dimension-distortion formula on a full-measure set (Theorem 2.2) is a clean extension of the dimension-squaring law, and the information-theoretic interpretation of the factor χ(q)/h(q,p) is a nice touch. The three-branch example in Section 5 does real work: it shows that the bulk limit alone cannot be sufficient, and the direct computation of two different subsequential limits is convincing.\n\nThe main soft spot is a printed inequality in the converse direction of Theorem 2.1. After constructing the sequence y_n, the proof claims A_n/B_n − log δ ≤ Q(x,y_n) ≤ (A_n − log p_min)/B_n. The lower bound is not implied by the earlier estimates and, even if true, would not squeeze Q to α because α − log δ is not α. The correct lower bound is A_n/(B_n − log δ), which follows from the same estimates (Q = (A_n − log z)/(B_n − log w) with z ≥ p_min and w ≥ δ), and then both bounds tend to α. So the conclusion lim A_n/B_n = α is still correct, but the proof as printed has a gap that needs fixing. This is a one-line repair, not a structural problem.\n\nThe example's claim that R_n(ω(1)) = O(√n) for the first coding is correct (blocks of 1's of length about 2k between squares at positions k^2). The Borel–Cantelli argument in Proposition 4.1 for R_n/n → 0 is standard but correct. The reliance on [7, Lemma 4.3] for dimension distortion is appropriate; the authors are not hiding the conclusion there.\n\nWho is this for? Fractal geometers and people working on devil's staircases, multifractal analysis, and dimension distortion under singular functions. A serious referee should send this to review; the paper is a solid contribution that needs a small but genuine correction before publication. I would accept it after minor revision.","headline":"A genuinely new pointwise scaling criterion for generalized Cantor functions, with a real but easily repairable proof gap in the converse of Theorem 2.1; worth refereeing.","tokens_in":12450,"tokens_out":8654,"would_cite":true,"duration_ms":74183,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"pith_extraction":{"msc":["28A80","26A30","28A78"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves a complete symbolic criterion for the local scaling exponent of any generalized Cantor function, and derives an exact Hausdorff-dimension distortion law from it.","keywords":["generalized Cantor function","self-similar measure","local scaling exponent","Hausdorff dimension","dimension distortion","symbolic coding","endpoint runs","cross-entropy"],"falsifier":"Construct, or search for, a coding $\\omega$ with $A_n(\\omega)/B_n(\\omega)\\to\\alpha$ and $R_n(\\omega)/n\\to 0$ for which the quotient $\\log|F(y)-F(x)|/\\log|y-x|$ nevertheless has two distinct subsequential limits as $y\\to x$; this would refute the sufficiency direction of Theorem 2.1. Conversely, a point where the scaling limit exists and equals $\\alpha$ but $A_n/B_n$ does not converge to $\\alpha$ would refute the necessity direction. The paper's Section 5 example supplies the template for the first search: coding $\\omega^{(2)}$ has bulk convergence but $R_{n_k}/n_k\\to 1$, and two approaching sequences give limits $\\alpha_0$ and $2\\alpha_0$, so the theorem predicts, and the calculation confirms, no scaling exponent.","tokens_in":11548,"feed_emoji":"📐","tokens_out":11594,"duration_ms":110842,"temperature":0.7,"pith_summary":"This paper establishes a complete pointwise criterion for when a generalized Cantor function, the distribution function of a self-similar Cantor measure, has a power-law local scaling at a point of its Cantor set. The criterion is symbolic: at $x=\\pi(\\omega)$, the scaling exponent exists and equals $\\alpha>0$ exactly when the ratio $A_n(\\omega)/B_n(\\omega)$ of accumulated logarithmic mass to accumulated logarithmic contraction converges to $\\alpha$ and the endpoint runs $R_n(\\omega)$ grow sublinearly. This separates the bulk mass-geometry balance from the corrections produced by long runs along extreme branches, and it works for arbitrary contraction ratios and probability weights under the sole assumption that the defining intervals are pairwise disjoint. As an application, the paper constructs a subset $M\\subset K$ of full Hausdorff measure on which the exponent is $h(\\mathbf q,\\mathbf p)/\\chi(\\mathbf q)$ and proves the exact dimension-distortion identity $\\dim_{\\mathrm H} F(A) = \\frac{\\chi(\\mathbf q)}{h(\\mathbf q,\\mathbf p)} \\dim_{\\mathrm H} A$ for every $A\\subset M$. A non-uniform three-branch example shows that convergence of the bulk ratio alone is not enough, because long endpoint runs can force two different limiting exponents along two sequences approaching the same point.","feed_headline":"One ratio test decides where Cantor functions scale","feed_subtitle":"Mass-to-geometry ratio plus short endpoint runs yields the exponent and an exact dimension-distortion law.","key_machinery":"The load-bearing object is the symbolic recursion $T(\\omega)=c_{\\omega_1}+p_{\\omega_1}T(\\sigma\\omega)$ for $T=F\\circ\\pi$, where $c_i$ is the cumulative mass of the cylinders to the left of branch $i$. This identity lets the proof express the value difference $|F(x)-F(y)|$ as a product of cylinder probabilities times a residual difference, and Lemma 3.2 uses it to bound that residual below by $p_{\\min}^{R_n(\\omega)+2}$. Together with the separation constant $\\delta=\\min_{i<j} \\operatorname{dist}(S_i(K),S_j(K))>0$, which keeps the geometric factor $\\log w(x,y)$ bounded, this reduces the scaling quotient $Q(x,y)$ to $A_n(\\omega)/B_n(\\omega)$ plus an error controlled by $R_n(\\omega)/n$. The sufficiency and necessity of Theorem 2.1 are both obtained from these two estimates.","core_discovery":"On the paper's own terms, the central claim is Theorem 2.1: for $x=\\pi(\\omega)$ in the self-similar set $K$ with strong separation, the logarithmic local scaling exponent $\\lim_{y\\to x,\\,y\\in K} \\frac{\\log|F(y)-F(x)|}{\\log|y-x|}$ equals $\\alpha>0$ if and only if $\\lim_n A_n(\\omega)/B_n(\\omega)=\\alpha$ and $\\lim_n R_n(\\omega)/n=0$. The proof splits the scaling quotient into the bulk ratio $A_{n-1}/B_{n-1}$ plus a correction term bounded by $R_n(\\omega)/n$; the endpoint-run condition forces that correction to vanish, and the contradiction argument shows that any linear endpoint run produces a second sequence with a strictly larger limit. Theorem 2.2 then restricts to $\\mathbf q$-typical codings, where the strong law gives $A_n/B_n \\to h(\\mathbf q,\\mathbf p)/\\chi(\\mathbf q)$ and $R_n/n\\to 0$ almost surely, so the exponent is constant on a full-measure set $M\\subset K^*$, and the distortion lemma converts this into the exact dimension formula $\\dim_{\\mathrm H} F(A) = \\frac{\\chi(\\mathbf q)}{h(\\mathbf q,\\mathbf p)} \\dim_{\\mathrm H} A$ for every $A\\subset M$.","pith_inferences":["Because the criterion is purely symbolic, it gives an algorithm: for any concrete coding, decide existence of the scaling exponent by checking two limits in digit frequencies and run lengths. The same reduction should apply to graph-directed self-similar functions, where endpoint runs are replaced by transitions along the graph.","The paper proves the dimension-distortion formula on a full $\\mathcal H^s$-measure subset, not on all of $K^*$; a natural next question, not addressed here, is whether the identity extends to all of $K^*$ or whether exceptional codings with linear endpoint runs genuinely break it.","The endpoint-run condition suggests a classification of points by $\\limsup R_n/n$: one expects mixed exponents governed by $\\alpha + c\\cdot \\limsup R_n/n$ when approaching from the side of a long run. The two-sequence example in Section 5 is exactly the first instance of such a mixed-exponent phenomenon.","The role of the separation constant $\\delta$ in the proof suggests that overlapping systems need a third term involving iterated overlaps; testing the criterion on a self-similar system with exact overlaps would reveal whether sublinear endpoint runs still suffice."],"forward_implications":["At every point whose coding has $A_n(\\omega)/B_n(\\omega)\\to\\alpha$ and $R_n(\\omega)/n\\to 0$, the local scaling exponent exists and equals $\\alpha$; in particular, the set of $\\alpha$-scaling points is described purely by digit statistics of the coding.","For $\\mathbf q$-typical points, the exponent is $h(\\mathbf q,\\mathbf p)/\\chi(\\mathbf q)$, so when $\\mathbf p=\\mathbf q$ the typical exponent is the similarity dimension $s$ and the dimension-distortion factor is $1/s$.","The dimension-distortion factor $\\chi(\\mathbf q)/h(\\mathbf q,\\mathbf p)$ equals $1/(s+D_{KL}(\\mathbf q\\|\\mathbf p)/\\chi(\\mathbf q))$, so any deviation of $\\mathbf p$ from the natural weights strictly lowers the factor below $1/s$; the classical ternary Cantor function is recovered by $\\mathbf p=\\mathbf q=(1/2,1/2)$.","If the local scaling exponent exists at every point of a set $A\\subset M$, then $F$ multiplies the Hausdorff dimension of every subset by the reciprocal of that exponent; hence the pointwise criterion and the dimension formula are two sides of one mechanism."],"supporting_citations":[{"why":"Supplies existence and uniqueness of the self-similar measure and the identity identifying the natural measure with normalized Hausdorff measure, used in Lemma 4.2.","marker":"[12]"},{"why":"Supplies the distortion lemma used in Theorem 2.2: a homeomorphism with constant local scaling exponent $\\alpha$ scales Hausdorff dimension by $1/\\alpha$.","marker":"[7]"},{"why":"Provides the classical ternary dimension-squaring baseline that Theorem 2.2 generalizes to arbitrary contraction ratios and weights.","marker":"[2]"},{"why":"Recovers dimension-squaring for natural self-similar measures with finitely many branches; the present result extends that law to arbitrary probability weights.","marker":"[4]"},{"why":"Provides the one-sided multifractal viewpoint that one-sided increments of distribution functions are the natural objects for local scaling analysis.","marker":"[9]"}],"fun_headline_variants":["Ratio test plus short runs pin Cantor scaling exponent","One ratio reveals Cantor scaling exponent, short runs required","Cantor scaling exponent fixed by mass-geometry ratio and short runs","Exact dimension distortion from Cantor scaling ratio test","Almost everywhere Cantor local scaling exponent is exact"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The single premise the argument cannot do without is that the intervals $S_i([0,1])$ are pairwise disjoint: this positive gap is what makes different codings geometrically separated and is used in both directions of the proof of Theorem 2.1.","fun_headline_variants_meta":{"raw":{"variants":["Ratio test plus short runs pin Cantor scaling exponent","One ratio reveals Cantor scaling exponent, short runs required","Cantor scaling exponent fixed by mass-geometry ratio and short runs","Exact dimension distortion from Cantor scaling ratio test","Almost everywhere Cantor local scaling exponent is exact"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000786,"raw_usage":{"total_tokens":3563,"prompt_tokens":1136,"completion_tokens":2427,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":752,"completion_tokens_details":{"reasoning_tokens":2348}},"tokens_in":752,"tokens_out":2427,"duration_ms":18623,"temperature":1.0,"reasoning_tokens":2348,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-27T18:33:46.992022+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct, or search for, a coding $\\omega$ with $A_n(\\omega)/B_n(\\omega)\\to\\alpha$ and $R_n(\\omega)/n\\to 0$ for which the quotient $\\log|F(y)-F(x)|/\\log|y-x|$ nevertheless has two distinct subsequential limits as $y\\to x$; this would refute the sufficiency direction of Theorem 2.1. Conversely, a point where the scaling limit exists and equals $\\alpha$ but $A_n/B_n$ does not converge to $\\alpha$ would refute the necessity direction. The paper's Section 5 example supplies the template for the first search: coding $\\omega^{(2)}$ has bulk convergence but $R_{n_k}/n_k\\to 1$, and two approaching sequences give limits $\\alpha_0$ and $2\\alpha_0$, so the theorem predicts, and the calculation confirms, no scaling exponent.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies existence and uniqueness of the self-similar measure and the identity identifying the natural measure with normalized Hausdorff measure, used in Lemma 4.2."},{"cited_title":"Dovgoshey, O","cited_arxiv_id":null,"evidence_quote":"Supplies the distortion lemma used in Theorem 2.2: a homeomorphism with constant local scaling exponent $\\alpha$ scales Hausdorff dimension by $1/\\alpha$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the classical ternary dimension-squaring baseline that Theorem 2.2 generalizes to arbitrary contraction ratios and weights."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Recovers dimension-squaring for natural self-similar measures with finitely many branches; the present result extends that law to arbitrary probability weights."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the one-sided multifractal viewpoint that one-sided increments of distribution functions are the natural objects for local scaling analysis."}],"review_version":1}