{"id":"cd286020-a551-4d9a-aaf0-422c52a3c119","arxiv_id":"1908.09020","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Sharp bounds of O(log n/(δσ)) and O(1/(δσ)) for distance to normality when roots avoid a disk or sector, resolving Pemantle's conjecture and the Ghosh-Liggett-Pemantle multivariate CLT question.","lead":"The paper proves sharp quantitative central limit theorems for random variables whose probability generating functions have zeros far from 1, resolving Pemantle's conjecture and a multivariate question for strong Rayleigh distributions. A reader interested in probability, combinatorics, or statistical physics will see exactly how root locations control Gaussian approximation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 4.6 defines the reflected Brownian path B° by an invalid concatenation: after τ2 it switches to B_t, but when B exits through the real axis the two pieces do not meet, so B° is not a Brownian motion and Lemma 4.1 is unproved as written.","rationale":"The reader's weakest-assumption analysis correctly identifies Lemma 4.1 as the pivot of the paper, but it focuses on the Brownian exit-probability estimate and the growth control on the ends of the sector. My check of the proof locates a more specific defect inside Lemma 4.6: the asserted Brownian motion B° is not continuous at the switching time τ2 when the original path exits through the real axis. Since Theorem 4.2 cannot be applied to a discontinuous path, Lemma 4.1 is not established exactly as written. The defect is easily repaired by defining B° as the reflected path α\\overline{B_t} for all times, which is a genuine Brownian motion and makes the three-event decomposition work. Because this repair is local and does not change the architecture or the resulting constants, the central claim remains credible: the issue is a proof gap to be corrected, not a counterexample or a false theorem. The reader's CONDITIONAL verdict is therefore appropriate; I do not see a reason to move it. My agreement is partial because the reader pointed at Lemma 4.1 and its quantitative ingredients but did not flag the specific invalid concatenation in the Brownian coupling.","tokens_in":37768,"tokens_out":19244,"duration_ms":192568,"concrete_test":"Rewrite §4.2 with B°_t := α\\overline{B_t} for all t ≥ 0 and no switching at τ2. Verify directly that (i) B° is a Brownian motion started at z°, (ii) B and B° have the same hitting structure of S_R(0,δ) with respect to the stopping time τ, and (iii) the three-event decomposition in Lemma 4.6 yields (21) with the same probability bound. If this corrected Lemma 4.6 holds, Theorem 1.2 and Theorem 1.4 go through with a minor revision; if it does not, then Lemma 4.1 is unsupported and the main theorems are unproved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 4.1 is the sole mechanism converting weak positivity and harmonicity into the b-decreasing property used by Lemmas 5.1, 6.1, 8.1, and Corollary 9.3. Its proof rests on Lemma 4.6. In the proof of Lemma 4.6 the path is defined by B°_t := α\\overline{B_t} for t ≤ τ2 and B°_t := B_t for t ≥ τ2. This is discontinuous: if the original Brownian motion first hits ∂S_R(0,δ/2) on the ray arg = 0 (event E2), then α\\overline{B_{τ2}} = αB_{τ2} has argument δ, not 0, so the two pieces do not join at τ2. Consequently B° is not a continuous path and cannot be a Brownian motion, so the application of Theorem 4.2 to z° and the equality of hitting times τ are unjustified. The subsequent use of B°_τ = αB_τ in event E2 indicates the intended definition is the reflected path α\\overline{B_t} for all t, which is a Brownian motion and makes the argument work: on E1 the reflected path coincides with B, on E2 it exits through the ray of argument δ where weak positivity applies, and on E3 the end-boundary crude bound applies. Thus the flaw is a localized gap rather than a refutation, but because Lemma 4.1 is load-bearing for both Theorem 1.2 and Theorem 1.4, it must be fixed and checked before the central claim is considered established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proves sharp quantitative central limit theorems for integer-valued random variables from zero-free regions of their probability generating functions. Theorem 1.2 bounds the Kolmogorov distance by O(log n/(δσ)) when no root of f_X lies within distance δ of 1; Theorem 1.4 obtains O(1/(δσ)) when no root lies in a sector of angular width δ around the positive real axis. A multivariate CLT for strong Rayleigh distributions (Theorem 1.6) is derived from Theorem 1.4 via a Cramér–Wold type argument, and Section 11 gives explicit constructions showing the rates are sharp. The proof develops a harmonic-analysis framework based on weak positivity and symmetry of u = log|f_X|, a Brownian-motion lemma converting weak positivity plus harmonicity into a 'b-decreasing' property, and a sequence of comparison and cumulant-tail lemmas culminating in a characteristic-function approximation and Fourier inversion.","tokens_in":38098,"tokens_out":25649,"duration_ms":258259,"significance":"If correct, these results resolve Pemantle's conjecture in a strong quantitative form, improve the Lebowitz–Pittel–Ruelle–Speer theorem, and answer the Ghosh–Liggett–Pemantle question on multivariate CLTs for strong Rayleigh variables. The paper is largely self-contained, gives explicit absolute constants, and provides matching lower bounds via explicit constructions. Because the main theorems are quantitative with no fitted parameters, and because the sharpness examples are explicit, the contribution would be a definitive and useful addition to probability and analytic combinatorics. The proof strategy, based on b-decreasing harmonic functions, appears flexible and is already applied to Hurwitz-stable and half-plane-stable generalizations.","major_comments":[{"comment":"The path B° is not a Brownian motion as defined. The definition sets B°_t := α\\overline{B_t} for t ≤ τ2 and B°_t := B_t for t ≥ τ2. These two pieces do not agree at τ2 unless B_{τ2} lies on the ray arg = δ/2; in particular, on event E2 (when B hits the real axis at τ2), the value specified for t ≥ τ2 has argument 0, while the value before τ2 has argument δ. Hence B° is discontinuous, is not a continuous path, and Theorem 4.2 cannot be applied to z°. The intended path is presumably B°_t := α\\overline{B_t} for all t ≥ 0, which is a Brownian motion and makes the E1/E2/E3 analysis meaningful; with the written definition, equations (22)–(25) do not follow. Since Lemma 4.1 is used in Theorems 1.2, 1.4, and 12.2, this gap must be repaired before the main claims are established.","section":"Section 4.2, proof of Lemma 4.6"},{"comment":"The step bounding P(Bτ ∈ S*_R(0,φ/2)) by (4/3)(r/R)^{4c/δ} is not justified by the cited application of Lemma 4.3. Lemma 4.3 is stated for symmetric sectors S_R(θ) with |arg| ≤ θ, whereas the event S*_R(0,φ/2) is a one-sided sector of angles [0,φ/2]. This event is contained in S_R(δ/4) only when φ ≤ δ/2, but the proof must handle all θ1,θ2 ∈ (0,δ/2), for which φ = θ1 + θ2 can be arbitrarily close to δ. Thus the displayed exponent 4c/δ is not established by the argument as written; a sharper exit-probability estimate or a genuinely different embedding is needed. This is load-bearing because it is exactly the step that produces the factor log n in Theorem 1.2.","section":"Section 4, proof of Lemma 4.1"},{"comment":"The statement of Theorem 11.4 is too broad as written. The proof sets k := ⌊log n/(100δ)⌋ and defines X := kY; when δ > log n/100, k = 0 and the construction is undefined, since k⌊n/k⌋ involves division by zero. The hypotheses allow such δ (for example, n = 10, δ = 10^9, σ = 1 satisfy log n/(δσ) ≤ 1), so the theorem is not proved for all δ > 0. The statement should either impose a restriction such as δ ≤ c log n or provide a separate construction covering the case k = 0.","section":"Section 11, Theorem 11.4"}],"minor_comments":[{"comment":"The quantity a2 is stated as −σ/2 in both places; it should be σ²/2. With the stated sign, the displayed identity ψ_{X*}(ξ) = exp(−ξ²/2 + R(ξ)) does not follow from the preceding expansion.","section":"Section 2, Eq. (3) and proof of Lemma 8.1"},{"comment":"The proof cites Lemma 5.4 for the inequality U''(t0) ≥ 0; the correct reference is Lemma 5.2.","section":"Proof of Lemma 5.4"},{"comment":"The event E3 is written as {B_{τ2} ∈ S*_R(δ/2)}, while the statement of Lemma 4.6 and the final inequality use the one-sided set S*_R(0,δ/2). These are different sets (two-sided versus one-sided ends), and the notation should be made consistent.","section":"Proof of Lemma 4.6"}],"recommendation":"major_revision","confidential_remarks":"The concatenation flaw in Lemma 4.6 is localized and almost certainly repairable by defining B° as the reflected path for all times. The sector-width issue in Lemma 4.1 also appears repairable with a standard sharp harmonic-measure estimate, but it is a genuine gap in the written proof. The overclaim in Theorem 11.4 should also be fixed. I see no concerns about novelty or attribution; the paper cites earlier work appropriately. My recommendation of major revision reflects the need to repair these load-bearing technical points and to correct the statement of Theorem 11.4."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Marcus and Julian have produced the definitive quantitative answer to Pemantle's conjecture and the GLP question. The bounds in Theorems 1.2 and 1.4 are optimal, the sharpness constructions are convincing, and the multivariate theorem falls out cleanly. The method—weak positivity plus harmonicity converted into b-decreasing via Brownian coupling—is genuinely new, and the paper is careful, with explicit constants throughout. This is a major result for probability and combinatorics.\n\nThat said, the reader's report is too generous about soundness. The sign typo in Lemma 8.1 (a2 should be σ^2/2, not −σ/2), the wrong cross-reference in Lemma 5.4, and the overbroad statement in Theorem 11.4 are all real but minor. The bigger problem is Lemma 4.6. The path B° is defined as α\\bar B_t for t ≤ τ2 and B_t for t ≥ τ2. The proof says the two pieces agree at τ2. They don't. If B exits the small sector through the real axis (event E2), then α\\bar B_{τ2} has argument δ while B_{τ2} has argument 0; the trajectory jumps. Same for the circular ends unless the angle is exactly δ/2. So B° is not a continuous Brownian motion, and the appeal to Theorem 4.2 to compute u(z°) = Eu(B°_τ) is unjustified. This is load-bearing: Lemma 4.1, the sole mechanism producing b-decreasing, uses Lemma 4.6. The stress-test note suggests using the reflected path α\\bar B for all t. I don't think that fix is immediate: on the event that B hits the intermediate ray δ/2, the reflected path only touches B at that instant and then diverges, so the E1 contribution is not automatically zero. The authors need to supply a correct coupling or a different argument. This is not a typo; it's a gap in a central lemma. My guess is the result is true and the gap is repairable, but as written the proof of Theorems 1.2 and 1.4 is incomplete.\n\nWho is this for: anyone working in negative dependence, stable polynomials, or CLTs from zero-free regions. It deserves a serious referee, but only on condition that Lemma 4.6 is fixed. I would not cite the proof until then, but I'd certainly read a revised version closely.","headline":"Strong, sharp results on two conjectures, but the proof has a real gap in the Brownian motion coupling (Lemma 4.6) that the reader's report missed; fixable, but the write-up as it stands is incomplete.","tokens_in":38656,"tokens_out":13272,"would_cite":false,"duration_ms":128287,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60F05","30C15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A root-free disk around 1 forces a nearly Gaussian distribution, with optimal error O(log n/(δσ)).","keywords":["central limit theorem","probability generating function","zero-free regions","real-stable polynomials","strong Rayleigh distributions","cumulants","Brownian motion","sharp quantitative bounds"],"falsifier":"Search for a counterexample to the disk theorem: any sequence with σ_n δ_n/log n → ∞, generating polynomial root-free in B(1, δ_n), and sup_t |F_n(t) − Φ(t)| bounded below by a fixed positive constant would refute it; the scaled-Bernoulli-sum family in Section 11 is the natural place to test, since the paper shows it saturates the bound.","tokens_in":37554,"feed_emoji":"📊","tokens_out":12821,"duration_ms":118543,"temperature":0.7,"pith_summary":"The paper proves that the geometry of the roots of a probability generating function controls how close a random variable is to Gaussian. If the roots stay a distance δ away from 1, then the normalized variable is within O(log n/(δσ)) of a standard normal, measured by the maximum discrepancy between cumulative distribution functions; if the roots stay away from a whole sector around the positive axis, the sharper O(1/(δσ)) holds. These rates are optimal, resolving a conjecture on root-free disks and sharpening an earlier central limit theorem from statistical physics. The sector version also yields a sharp multivariate central limit theorem for real-stable, strong-Rayleigh distributions.","feed_headline":"No roots near 1 forces near-Gaussian behavior","feed_subtitle":"Optimal log n/(δσ) error resolves the disk conjecture and a multivariate CLT follows.","key_machinery":"The carrying object is the logarithmic potential u(z) = log|f_X(z)|, which is harmonic wherever f_X has no roots, symmetric because the coefficients are real, and weakly positive (u(|z|) ≥ u(z)) because the coefficients are non-negative probabilities. The main technical lemma, Lemma 4.1, shows that weak positivity plus harmonicity forces u to be b-decreasing — the value of u along a ray cannot drop by more than b as the angle increases — using a Brownian-motion estimate for the probability of exiting a truncated sector through its ends, which carries the exponential dependence on 1/δ. Once b-decreasing is established, the paper controls the tail of the normalized cumulant sequence, uses weak positivity to dominate all higher cumulants by the variance, and packages the result as a factorized characteristic function exp(−ξ²/2 + R(ξ)) with |R(ξ)| ≤ C|ξ|³/(εσ). A Fourier-inversion lemma converts this into the stated uniform distributional bounds. For the sector version, the same proof runs over an unbounded sector, so the Brownian exit probability decays and no logarithmic factor appears.","core_discovery":"The central discovery, stated as Theorems 1.2 and 1.4, is a quantitative central limit theorem in which the only information about the random variable X is its variance σ and the location of the roots of its probability generating function f_X(z) = E[z^X]. When the roots satisfy |ζ − 1| ≥ δ, the normalized variable X* = (X − μ)/σ satisfies sup_t |P(X* ≤ t) − P(Z ≤ t)| = O(log n/(δσ)); when the roots satisfy |arg ζ| ≥ δ, the same discrepancy is O(1/(δσ)). The proof represents the characteristic function of X* as exp(−ξ²/2 + R(ξ)) with |R(ξ)| controlled by ξ³/(δσ), and the logarithmic factor in the first case is shown to be unavoidable. In the multivariate direction, Theorem 1.6 proves that if the generating functions are real-stable and the maximum variance tends to infinity, the normalized random vectors converge to a multivariate normal under no further conditions.","pith_inferences":["The logarithmic factor appears tied to the disk geometry: the Brownian-exit estimate has exponential dependence on 1/δ, and an intermediate zero-free region, such as a curve tangent to 1, might produce rates interpolating between 1/(δσ) and log n/(δσ).","The multivariate theorem is stated as a limit; tracking explicit constants and the dependence on the dimension and on the conditioning of the covariance matrix would yield finite-sample bounds for random spanning trees, matchings, and determinantal measures.","The method is not restricted to real-stable families; any class of polynomials with a known sector zero-free region is a candidate for the same b-decreasing route, so one could test it on independence polynomials or matching polynomials away from the positive axis."],"forward_implications":["For a sequence with σ_n δ_n/log n → ∞ and roots avoiding the disk B(1, δ_n), the normalized variables converge in distribution to a standard normal; the condition is best possible.","For a sequence with σ_n δ_n → ∞ and roots avoiding the sector |arg ζ| < δ_n, normality holds with no logarithmic loss.","Every real-stable strong-Rayleigh sequence whose maximum variance tends to infinity and whose normalized covariance matrices converge converges in distribution to the corresponding multivariate normal.","The same machinery applies to power series and general analytic generating functions satisfying a mild growth condition, and to other classes with a sector zero-free property.","The sharpness examples show the bound cannot be improved: within the stated root-free classes, O(log n/(δσ)) and O(1/(δσ)) are the correct orders."],"supporting_citations":[{"why":"Formulates the conjecture that a root-free disk around 1 plus growing variance forces asymptotic normality; Theorem 1.2 resolves it in sharp quantitative form.","marker":"[55]"},{"why":"Establishes an earlier CLT under a root-free neighborhood of 1 with the stronger growth condition σ_n n^{-1/3} → ∞; Theorem 1.2 removes the extra growth.","marker":"[40]"},{"why":"Proves a multivariate CLT for real-stable laws under σ_n n^{-1/3} → ∞ and raises the question whether σ_n → ∞ suffices; Theorem 1.6 answers it, and Lemma 10.1 supplies the sector-root-free projection fact.","marker":"[27]"},{"why":"Prior counterexamples show the variance threshold cannot be reduced to a constant multiple of log n, setting the sharp context for Theorem 1.2.","marker":"[50]"},{"why":"Provides the Brownian-motion and harmonic-function theorems used to prove Lemma 4.1's b-decreasing property.","marker":"[53]"},{"why":"Introduces the real-stable/strong-Rayleigh framework and its negative-dependence theory, which underpins the multivariate theorem.","marker":"[13]"},{"why":"A sharp projection theorem that lets convergence of many one-dimensional projections imply convergence of the full multivariate law.","marker":"[19]"},{"why":"Develops the positivity-of-difference method for coefficients of large powers of polynomials, refined here in Lemma 7.5 to dominate higher cumulants by the variance.","marker":"[8]"},{"why":"A precursor of the same positivity technique used to control the cumulant sequence.","marker":"[20]"},{"why":"Supplies the Fourier-inversion inequality that converts characteristic-function estimates into uniform distributional bounds.","marker":"[24]"}],"fun_headline_variants":["Root-free zone forces Gaussian with optimal rate","CLT error O(log n/δσ) when roots flee 1","Sharp CLT from polynomial root geometry","Real-stable roots: multivariate normality follows","Pemantle's conjecture: proved and sharpened"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole proof leans on Lemma 4.1, which turns weak positivity and harmonicity into a b-decreasing property by a Brownian-motion estimate of the chance that a path exits a thin sector through its ends; if that estimate were materially weaker, the quantitative bounds would fail.","fun_headline_variants_meta":{"raw":{"variants":["Root-free zone forces Gaussian with optimal rate","CLT error O(log n/δσ) when roots flee 1","Sharp CLT from polynomial root geometry","Real-stable roots: multivariate normality follows","Pemantle's conjecture: proved and sharpened"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000335,"raw_usage":{"total_tokens":1945,"prompt_tokens":1118,"completion_tokens":827,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":734,"completion_tokens_details":{"reasoning_tokens":752}},"tokens_in":734,"tokens_out":827,"duration_ms":8226,"temperature":1.0,"reasoning_tokens":752,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:24:38.735350+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Search for a counterexample to the disk theorem: any sequence with σ_n δ_n/log n → ∞, generating polynomial root-free in B(1, δ_n), and sup_t |F_n(t) − Φ(t)| bounded below by a fixed positive constant would refute it; the scaled-Bernoulli-sum family in Section 11 is the natural place to test, since the paper shows it saturates the bound.","supporting_citations":[{"cited_title":"Pemantle","cited_arxiv_id":null,"evidence_quote":"Formulates the conjecture that a root-free disk around 1 plus growing variance forces asymptotic normality; Theorem 1.2 resolves it in sharp quantitative form."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes an earlier CLT under a root-free neighborhood of 1 with the stronger growth condition σ_n n^{-1/3} → ∞; Theorem 1.2 removes the extra growth."},{"cited_title":"Ghosh, T","cited_arxiv_id":null,"evidence_quote":"Proves a multivariate CLT for real-stable laws under σ_n n^{-1/3} → ∞ and raises the question whether σ_n → ∞ suffices; Theorem 1.6 answers it, and Lemma 10.1 supplies the sector-root-free projection fact."},{"cited_title":"Michelen and J","cited_arxiv_id":null,"evidence_quote":"Prior counterexamples show the variance threshold cannot be reduced to a constant multiple of log n, setting the sharp context for Theorem 1.2."},{"cited_title":"M¨ orters and Y","cited_arxiv_id":null,"evidence_quote":"Provides the Brownian-motion and harmonic-function theorems used to prove Lemma 4.1's b-decreasing property."},{"cited_title":"Borcea, P","cited_arxiv_id":null,"evidence_quote":"Introduces the real-stable/strong-Rayleigh framework and its negative-dependence theory, which underpins the multivariate theorem."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A sharp projection theorem that lets convergence of many one-dimensional projections imply convergence of the full multivariate law."},{"cited_title":"Bergweiler, A","cited_arxiv_id":null,"evidence_quote":"Develops the positivity-of-difference method for coefficients of large powers of polynomials, refined here in Lemma 7.5 to dominate higher cumulants by the variance."},{"cited_title":"De Angelis","cited_arxiv_id":null,"evidence_quote":"A precursor of the same positivity technique used to control the cumulant sequence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Fourier-inversion inequality that converts characteristic-function estimates into uniform distributional bounds."}],"review_version":1}