{"id":"e26e33db-267f-428b-bb3e-0331125d672f","arxiv_id":"2507.21605","paper_version":3,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For random Bernoulli convolutions, the Fourier transform is in L^1 almost surely whenever λ_g > 2/π, giving absolute continuity and non-empty interior; polynomial Fourier decay holds for every λ_g.","lead":"This paper proves that for random Bernoulli convolutions, when the geometric mean of the contraction ratios exceeds 2/π, the Fourier transform is integrable almost surely. This yields absolute continuity with a continuous density and a support with non-empty interior, improving the known threshold from about 0.824 to about 0.636.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The threshold 2/π rests on a Riemann-sum limit at a singular endpoint; the proof as written does not justify it.","rationale":"I read the proof in good faith and found the overall strategy coherent. The reader's weakest-assumption point about the uniform conditional distribution in Lemma 9 is correct and genuinely load-bearing, but it is an assumption of the model rather than a flaw in the proof. My main concern is different and more local: in the final step of Theorem 1, the paper passes from a discrete sum to an integral whose value 1 determines the constant 2/π. The function f has an unbounded derivative at 0, so the claimed Riemann-sum limit is not immediate and, as written, involves an infinite term. This does not mean the theorem is wrong, but it means the central numerical threshold is not fully established by the text. I would therefore recommend accepting only after a short rigorous justification of the limit is inserted. The paper otherwise appears mathematically sound: Lemma 4 is standard, Lemmas 5-8 correctly reduce to the conditional probability estimates, Lemma 9's geometric estimate is plausible and detailed, and the Borel-Cantelli arguments in Sections 2 and 3 are coherent. The flaws I noticed are typographical (e.g. Lemma 8's product starting index and the final λ_g^i ≤ ξ^{-1}λ_g inequality in Theorem 3, which should likely be ξ^{-1}λ_g^{-1}) rather than substantive. No ad hominem is intended; the concern is entirely about a missing proof step.","tokens_in":19695,"tokens_out":40383,"duration_ms":476968,"concrete_test":"Provide a rigorous derivation of the limit: split the sum at k0=⌊Mδ⌋, control the k=0 term by f(1/M)→0, use monotonicity or MVT on k≥1, and apply dominated convergence on [δ,∞), then let δ↓0. Confirm the limit equals ∫_0^1 u/√(1-u^2)du = 1. As a numerical cross-check, evaluate S_M = ∑_{k=0}^∞ 0.65^{k/M}(arccos(0.65^{(k+1)/M})-arccos(0.65^{k/M})) for M=10^3,10^4,10^5; if S_M does not approach 1, the threshold is not 2/π.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In the proof of Theorem 1, after Eq. (2.12), the authors set f(x)=arccos(λ_g^x) and claim that lim_{M→∞} ∑_{k=0}^∞ λ_g^{k/M}(f((k+1)/M)-f(k/M)) equals ∫_0^1 u/√(1-u^2)du = 1. This limit is exactly what turns the geometric rate λ_g^{i/(E_i-n)} into the threshold λ_g > 2/π. The difficulty is that f'(x) = -λ_g^x lnλ_g / √(1-λ_g^{2x}) blows up at x=0, so the displayed Riemann-sum identity is not justified by the standard convergence theorem for continuous integrands. In particular, the paper's intermediate expression involving (1/M)λ_g^{k/M}f'(k/M) has an infinite k=0 term, while the original k=0 term f(1/M)-f(0) tends to 0. The identity is very likely correct and can be repaired by splitting off a neighbourhood of 0, but the paper does not supply that argument. Because the central claim's exact constant depends on this limit, this is a load-bearing gap in the proof as written, even though I do not believe the theorem is false.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies random Bernoulli convolutions μ_ω formed as infinite convolutions with i.i.d. contraction ratios λ_k uniformly distributed on a fixed interval W ⊂ (0,1). Its main result, Theorem 1, states that if λ_g = exp(E log λ_1) > 2/π, then the Fourier transform of μ_ω is in L^1(R) almost surely. Theorem 2 then deduces that μ_ω is absolutely continuous with a continuous density and that its supporting random self-similar set has non-empty interior almost surely. Theorem 3 states that, without any lower bound on λ_g, the Fourier transform decays polynomially almost surely. The proof of Theorem 1 uses the infinite-product representation of |μ̂_ω|, a decomposition of frequencies into dyadic-like intervals I_i, a conditioning estimate for the 'good contraction' events (Lemma 9), and a limiting Riemann-sum computation that yields the constant 2/π. The paper thereby improves the earlier threshold e^{1/2}/2 ≈ 0.824 of Peres, Simon and Solomyak to 2/π ≈ 0.636.","tokens_in":19862,"tokens_out":12333,"duration_ms":141945,"significance":"If the proof is correct, this is a clean and substantial improvement over the previous threshold, and the exact constant 2/π is derived from an explicit integral rather than from any fitted parameter. The paper is self-contained, its structural reductions (Lemmas 4–8) are transparent, and Theorem 3 gives a useful almost-sure polynomial decay result that is new in this random setting. The uniform nature of the contraction-ratio law is genuinely used and is explicitly stated, which is appropriate. The main concern is one local but load-bearing limiting step in the proof of Theorem 1; once repaired, the paper would be a solid contribution to the Fourier analysis of random self-similar measures.","major_comments":[{"comment":"The limiting identity that produces the constant 1, and hence the threshold 2/π, is not justified as written. The paper replaces the sum ∑_{k=0}^∞ λ_g^{k/M} (arccos λ_g^{(k+1)/M} − arccos λ_g^{k/M}) by the Riemann sum (1/M)∑_{k=0}^∞ λ_g^{k/M} f'(k/M), with f(x) = arccos(λ_g^x), and then by the integral ∫_0^∞ λ_g^x f'(x) dx. The difficulty is that f'(x) = −λ_g^x ln λ_g / √(1 − λ_g^{2x}) has a square-root singularity at x = 0. Consequently the term k = 0 in the displayed Riemann sum is infinite for every finite M, while the corresponding term f(1/M) − f(0) in the original sum tends to 0 as M → ∞. Thus the displayed equality between the original sum and that Riemann sum is not valid, and the sentence 'it immediately follows' does not constitute a proof. The limiting identity itself appears to be correct: it can be established by splitting off a neighbourhood of 0, estimating the first interval separately with f(1/M) − f(0) ≈ C/√M, and applying dominated convergence to the tail. Because this limit is what converts the exponent i/(E_i−n) into the numerical condition λ_g > 2/π, this gap is load-bearing and should be repaired in a revision.","section":"§2, proof of Theorem 1, after Eq. (2.12)"},{"comment":"The paper's main theorem depends essentially on the conditional law of each λ_k given the past being absolutely continuous with a bounded density (here, uniform on W). This is exactly what makes the hitting probability of a small target interval proportional to its length in Lemma 9. The authors do state the uniform assumption, but Remark 1 generalizes only to 'another probability absolutely continuous with respect to Lebesgue'. I recommend making explicit that the argument does not extend to deterministic or atomic choices of λ_k, and that the threshold 2/π is tied to the non-atomic nature of the law; otherwise readers may over-read the result as depending only on λ_g. This is a clarification of the scope of the theorem rather than a correction of the proof.","section":"§2, Lemma 9 and Remark 1"}],"minor_comments":[{"comment":"The notation 'ε2i' appears in several places (e.g., in the statement of Lemma 9 and in the displayed estimates) and should be written as 'ε^2 i' to avoid ambiguity with 2ε i.","section":"§2, proof of Theorem 1, after Eq. (2.12)"},{"comment":"The sentence 'the second derivative of f'' is negative' should read 'f'' is negative'; also, the sign of f' is positive only because ln λ_g < 0, which could be stated explicitly.","section":"§2, proof of Theorem 1, after Eq. (2.12)"},{"comment":"The infinite product in the definition of C_n is written with index i in one place and j in another; the index should be consistent.","section":"§2, Eq. (2.12)"},{"comment":"The phrase 'for all large i large enough' should be 'for all sufficiently large i'.","section":"§3, Lemma 13"},{"comment":"In the passage following Eq. (3.16), the statement that the right-hand side tends to infinity as p → 1 is correct, but the coefficient (ε − 1 + p/2) is negative for ε < 1/2; a brief explanation of why its product with log(1 − p) tends to +∞ would help the reader.","section":"§3, proof of Lemma 13"}],"recommendation":"major_revision","confidential_remarks":"The Riemann-sum gap highlighted in the major comments is real but local; I believe the theorem is true and the gap is repairable within the scope of this paper. This is not a rejection. The paper improves a known threshold and is well suited to the journal; I would encourage the editor to seek a revision with the missing limiting argument supplied."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Good paper, worth refereeing. The main new content is real: Theorem 1 drops the threshold for an almost-sure L^1 Fourier transform (hence absolute continuity with continuous density and non-empty interior) from e^{1/2}/2 ≈ 0.824 to 2/π ≈ 0.636, and Theorem 3 gives polynomial Fourier decay for every λ_g without any threshold assumption. The method is also genuinely different from PSS06: the paper estimates the L^1 norm directly, via a hitting-probability estimate for products of random contractions, rather than going through Sobolev dimension. Lemma 9 is the core probabilistic estimate and it looks sound; the conditioning on the past is handled carefully, and the randomness of the λ_k is really doing the work, not just as a technical convenience.\n\nI checked the stress-test on the Riemann-sum identity and it lands. In the proof of Theorem 1, after (2.12), the authors write that the series is immediately an integral. The problem is that the integrand g(x) = λ^x f'(x) is singular at x = 0 (it behaves like c x^{-1/2}), so the standard convergence theorem for Riemann sums does not apply. The k = 0 term in their original sum is f(1/M) - f(0), which tends to 0, not the left-endpoint value g(0)/M. The limit itself is almost certainly correct: split off a neighbourhood of 0, use integrability of g and the mean-value estimates already in the paper, and the identity follows. But as written it is a genuine gap at a load-bearing point: the constant 1 in that limit is exactly what converts the geometric rate into the threshold λ_g > 2/π. I do not think the theorem is false, and a competent referee should ask for this repair, not reject over it.\n\nMinor issues: a couple of index typos in display estimates, e.g. the product in (2.12) has an 'i' where a 'j' is meant, and the notation F_n(h_1, ..., h_{k-1}) is a bit abusive since the subscript n is fixed while k changes. None of this threatens the argument.\n\nBottom line: this is a solid, honest contribution for people working on random self-similar sets and Fourier decay. The citation pattern is normal, there are no fitted constants, and no circularity. I would send it to a serious referee and expect acceptance after a small but necessary revision.","headline":"A real improvement on random Bernoulli convolutions; the 2/π threshold is new and mostly well proved, but one Riemann-sum step at a singular endpoint needs a fix.","tokens_in":20487,"tokens_out":4460,"would_cite":true,"duration_ms":51860,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["28A80","42A38","60G57"],"pacs":[],"model":"deepseek-v4-flash","headline":"For random Bernoulli convolutions, the geometric mean contraction exceeding 2/π puts the Fourier transform in L^1 almost surely, forcing a continuous density and interior support.","keywords":["random Bernoulli convolutions","Fourier transform","L1 integrability","absolute continuity","self-similar sets","Rajchman measures","polynomial Fourier decay","random iterated function systems"],"falsifier":"Simulate many realizations with $W$ chosen so that $\\lambda_g$ is just above $2/\\pi$, and for each realization numerically approximate the integral of the truncated product $\\prod_{j=1}^N |\\cos(\\pi\\xi\\prod_{k=1}^j \\lambda_k)|$ over growing frequency windows; if these integrals diverge as $N$ grows for a positive-measure set of $\\omega$, Theorem 1 is false, whereas bounded growth on typical draws is the theorem's prediction.","tokens_in":19447,"feed_emoji":"📉","tokens_out":15100,"duration_ms":165309,"temperature":0.7,"pith_summary":"The paper studies random Bernoulli convolutions, the distributions of infinite sums $\\sum a_j \\prod_{k=1}^j \\lambda_k$ with $a_j\\in\\{0,1\\}$ fair coins and with contraction ratios $\\lambda_k$ i.i.d. uniform on a fixed interval $W\\subset(0,1)$. Its central claim is that when the geometric mean contraction $\\lambda_g=\\exp\\mathbb{E}(\\log\\lambda_1)$ exceeds $2/\\pi$, the Fourier transform of the random measure is in $L^1(\\mathbb R)$ almost surely. From there the measure is automatically absolutely continuous with a continuous density, and the random self-similar set supporting it has non-empty interior almost surely. This lowers the previously known threshold for interior points, about $0.824$, to about $0.636$, by estimating the $L^1$ norm of the Fourier transform directly. The paper also proves, without any restriction on $\\lambda_g$, an almost-sure uniform polynomial decay rate for the Fourier transform.","feed_headline":"Random Bernoulli sums gain a density above 2/π","feed_subtitle":"As soon as the geometric mean contraction exceeds 2/π, the random self-similar set has interior points almost surely.","key_machinery":"The engine is the infinite-product identity\n\\[\n|\\widehat\\mu_\\omega(\\xi)|=\\prod_{j=1}^\\infty\\left|\\cos\\left(\\pi\\xi\\prod_{k=1}^j\\lambda_k\\right)\\right|,\n\\]\nwhich converts the $L^1$ question into a quantitative question about how often the randomly scaled products $\\xi\\prod_{k=1}^j\\lambda_k$ land near integer multiples of $\\pi$. The proof partitions frequencies into geometric shells $I_i=[\\lambda_g^{-i},\\lambda_g^{-i-1})$, truncates the product at $E_i=\\lfloor(1-\\varepsilon)i\\rfloor$, and conditions on the past through a martingale filtration. The key estimate (Lemma 9) bounds the conditional probability that the next cosine factor is as small as $\\lambda_g^{h/M}$ by $(1+3\\lambda_g^{\\varepsilon^2 i}/\\Delta)$ times the arccos-interval length divided by $\\pi/2$. Summing these bounds yields a geometric series whose ratio is governed by the integral $\\int_0^1 u/\\sqrt{1-u^2}\\,du=1$, and $\\lambda_g>2/\\pi$ is exactly the condition that makes the ratio less than $1$.","core_discovery":"On the paper's own terms, the discovery is that the randomness of the contraction ratios is strong enough to overcome the algebraic obstructions that can make deterministic Bernoulli convolutions singular. For almost every $\\omega$, whenever $\\lambda_g > 2/\\pi$ the Fourier transform satisfies $\\widehat\\mu_\\omega \\in L^1(\\mathbb R)$, which by the classical criterion means $\\mu_\\omega$ is absolutely continuous with a continuous density and hence $\\Lambda_\\omega$ has non-empty interior. The proof actually gives more: there is a universal exponent $\\rho>0$ such that almost surely $|\\widehat\\mu_\\omega(\\xi)|\\le C_\\omega|\\xi|^{-\\rho}$ for every $\\xi\\neq0$, with no hypothesis on $\\lambda_g$ except the standing assumption $\\lambda_{\\max}<1$.","pith_inferences":["For other atomless distributions on the contraction ratios, the same conditioning scheme should produce a threshold tied to how concentrated the distribution is; the value $2/\\pi$ is a feature of the uniform model, not a universal constant for random self-similar sums.","Because the conditioning estimate uses the non-atomic law of each $\\lambda_k$, the theorem cannot be transplanted to deterministic contraction sequences by continuity; known singular examples with rigid algebraic parameters show the analogous conclusion can fail in the deterministic setting.","A natural numerical probe is to let $\\lambda_g = 2/\\pi$ and watch the truncated $L^1$ integrals used in the proof: divergence there would indicate the constant is sharp, while convergence just below the boundary would suggest the true threshold is lower."],"forward_implications":["If $\\lambda_g > 2/\\pi$, then for almost every contraction sequence the measure $\\mu_\\omega$ has an $L^1$ Fourier transform, hence a continuous density, and the random self-similar set $\\Lambda_\\omega$ has non-empty interior almost surely.","The interior-point threshold drops from the previous $e^{1/2}/2 \\approx 0.824$ to $2/\\pi \\approx 0.636$, widening the parameter range known to produce absolutely continuous random Bernoulli convolutions.","Without any restriction on $\\lambda_g$ (while $\\lambda_{\\max}<1$), there is a single exponent $\\rho>0$ such that almost surely $|\\widehat\\mu_\\omega(\\xi)| \\le C_\\omega |\\xi|^{-\\rho}$ for all $\\xi\\neq0$; in particular, the measures are almost surely Rajchman with a uniform polynomial decay rate.","The paper does not settle whether $2/\\pi$ is the sharp threshold, so the optimal boundary remains an open question inside $(1/2, 2/\\pi]$, since $\\lambda_g \\le 1/2$ gives singular measures almost surely."],"supporting_citations":[{"why":"Establishes the previous absolute-continuity threshold for random Bernoulli convolutions via a Sobolev dimension estimate; the paper's Theorem 1 improves its interior-point threshold from $e^{1/2}/2$ to $2/\\pi$.","marker":"[PSS06]"},{"why":"Supplies the standard infinite-product formula for the Fourier transform, the criterion that an $L^1$ Fourier transform gives absolute continuity with continuous density, and the uniform Lipschitz bound used to extend decay from a countable grid to all frequencies.","marker":"[Mat15]"},{"why":"Constructs singular deterministic Bernoulli convolutions at algebraically rigid parameter values, the obstruction that the random model is designed to bypass.","marker":"[Erd39]"},{"why":"Provides the classical Fourier-decay scheme that Section 3 adapts to prove the polynomial decay statement in Theorem 3.","marker":"[Kah79]"},{"why":"Raises the open question of self-similar sets with positive measure but empty interior, to which Theorem 2 contributes by proving interior points almost surely.","marker":"[PS00]"}],"fun_headline_variants":["Random Bernoulli folds get density above 2/π","Almost sure interior for random self-similar sets","Fourier L^1 for random Bernoulli past 2/π","Random contractions: polynomial decay, a.s.","Random Bernoulli: absolute continuity at 2/π"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each contraction ratio $\\lambda_k$ is chosen randomly with a continuous range of possible values, uniformly on a fixed interval, so that the conditional probability of landing in any small target set is proportional to its length; if the $\\lambda_k$ were fixed or atomic, the analogous conclusion can fail even when $\\lambda_g>2/\\pi$.","fun_headline_variants_meta":{"raw":{"variants":["Random Bernoulli folds get density above 2/π","Almost sure interior for random self-similar sets","Fourier L^1 for random Bernoulli past 2/π","Random contractions: polynomial decay, a.s.","Random Bernoulli: absolute continuity at 2/π"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1414,"prompt_tokens":924,"completion_tokens":490,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":413}},"tokens_in":540,"tokens_out":490,"duration_ms":6205,"temperature":1.0,"reasoning_tokens":413,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:37:57.515484+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate many realizations with $W$ chosen so that $\\lambda_g$ is just above $2/\\pi$, and for each realization numerically approximate the integral of the truncated product $\\prod_{j=1}^N |\\cos(\\pi\\xi\\prod_{k=1}^j \\lambda_k)|$ over growing frequency windows; if these integrals diverge as $N$ grows for a positive-measure set of $\\omega$, Theorem 1 is false, whereas bounded growth on typical draws is the theorem's prediction.","supporting_citations":[],"review_version":1}