{"id":"5b6fc965-4f82-4cef-97b8-0dc95ee827df","arxiv_id":"1908.00730","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For Kac polynomials and related random polynomials, the limiting zero distribution of the N_n-th derivative depends on the limit of N_n/n, with a phase transition and rescaling when N_n/n tends to 1.","lead":"This paper studies what happens to the zeros of a random polynomial after you differentiate it many times, when the number of derivatives grows with the polynomial's degree. It shows that the behavior changes in a sharp way depending on how fast the derivative count grows.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's lower-bound estimate (71) is asserted, not proved, and every main phase-transition result depends on it; a complete derivation under A2 is needed before the claims are fully supported.","rationale":"The reader's weakest-assumption identification matches my reading. The Kac-specific Stirling estimates in Section 2 are detailed and internally consistent, and the phase-transition picture is not contradicted by any demonstrated counterexample. The real soft spot is the proof of Theorem 2: the lower bound (71) is load-bearing, and the appendix does not supply the necessary argument for A2, instead invoking [9] in a way that is not a direct black box. Because this is a genuine gap in rigor rather than a known falsehood, the appropriate verdict is CONDITIONAL, which is what the reader assigned. I see no reason to move the verdict in either direction.","tokens_in":19244,"tokens_out":24755,"duration_ms":259137,"concrete_test":"Write out a complete proof of (71) under A2: for fixed z, construct J = {t in [0,T0] : st + log p(t) > I(s) - epsilon}, prove that |{k <= (T0-delta_n)L_n : k/L_n in J}| >= c L_n for all large n, including the case where the supremum is attained at T0 and delta_n > 0, and verify the KZ small-ball estimate with L_n in place of n. If this derivation fails for some A2-compliant sequence, Theorem 2 must be weakened or A2 strengthened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central conclusions for Kac, general, and elliptic polynomials (Theorems 3, 5, and 6) all route through Theorem 2. Appendix A explicitly only sketches the proof, and the decisive lower-bound estimate is (71): P(|F_n(z)| < e^{L_n(I(log|z|)-4epsilon)}) = O(1/sqrt(L_n)). The text says that the set J from the proof in [9] has at least |J|/2 L_n points and that the rest follows by replacing n with L_n. This is not a routine substitution. The Kabluchko-Zaporozhets argument is designed for assumptions A1, which include the R0 lower-bound condition A1.4; A2 has no analogous condition, and the index range itself depends on delta_n. The good set J must be intersected with [0, (T0-delta_n)L_n], and the delta_n^- (log|z|)^+ term in the upper bound (70) must be shown to be harmless. No verification is supplied that the intersection preserves linear cardinality when the supremum of I is attained at T0 and delta_n > 0, nor that the small-ball/Paley-Zygmund step of [9] survives with L_n in place of n. Since the limiting measures in Theorems 3(2)-(3), 5, and 6(2) are inherited from Theorem 2, an unproved (71) leaves the phase-transition conclusions unsupported. A secondary issue is that the printed Legendre-Fenchel transform in Section 2.2 is incorrect as written; differentiating it gives a/(1-r+ar), not ar/(1-r), although the measure claimed afterwards is the derivative of the correct transform.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the empirical measure of complex zeros of the N_n-th derivative of random polynomials, primarily Kac polynomials with i.i.d. coefficients. The main result for Kac polynomials (Theorem 3) is that the limiting global distribution depends on a = lim N_n/n: for a=0 the zeros uniformly cluster on the unit circle; for a in (0,1) the limit is an explicit rotationally invariant measure (Eq. (15)); for a=1 the zeros collapse to the origin, and after rescaling by n/D_n one obtains a measure with density 1/(2π|z|) on the unit disk (Eq. (17)-(18)). The fixed-degree case is treated by Rouché's theorem (Theorem 4), yielding a random limit related to f_m(z)=Σ ξ_k z^k/k!. The results are extended to general random polynomials satisfying the Kabluchko-Zaporozhets assumptions (Theorem 5), and for elliptic polynomials a different rescaling rate √(n/D_n) with an explicit limiting measure is computed (Theorem 6). The proofs route through a modified Kabluchko-Zaporozhets theorem (Theorem 2), whose proof is sketched in Appendix A.","tokens_in":19574,"tokens_out":13083,"duration_ms":121753,"significance":"If the results are fully established, this is a substantial contribution to the theory of random polynomials. It answers a natural question about the effect of dependent-on-degree differentiation and identifies a sharp phase transition controlled by the ratio N_n/n. The limiting measures are explicit and parameter-free, and the elliptic example demonstrates that the general A1 assumptions are insufficient to determine the rescaling, which is an honest and valuable observation. The coefficient asymptotics in Sections 2 and 4 are computed carefully and in detail, and the two-step heuristic connecting the fixed-degree and growing-degree regimes is illuminating. The main weakness is that the paper's central tool, Theorem 2, is only sketched: the critical lower-bound estimate (71) is asserted to follow from the Kabluchko-Zaporozhets proof by a substitution, but no complete argument is supplied. Since Theorems 3, 5, and 6 all inherit their limits from Theorem 2, the main conclusions are conditional on a rigorous proof of that theorem.","major_comments":[{"comment":"The proof of Theorem 2 is the load-bearing step for all main limit theorems, but the lower-bound estimate (71) is not proved. The text states that the set J from the proof in [9] has at least |J|/2 L_n points after intersection with [0,(T0-δ_n)L_n] and that the rest follows by replacing n with L_n. This is not a routine substitution: the assumptions A2 do not include an analogue of condition A1.4 (the R0 lower-bound condition), and the index range itself depends on δ_n. No argument is given that the intersection preserves linear cardinality when the supremum of I is attained at T0 and δ_n > 0, nor that the small-ball/Paley-Zygmund step of [9] works under A2 with L_n in place of n. Since the limiting measures in Theorem 3(2)-(3), Theorem 5, and Theorem 6(2) are inherited from Theorem 2, an unproved (71) leaves the phase-transition conclusions unsupported. A complete proof of Theorem 2 under A2 is needed before the main results are fully verified.","section":"Appendix A, Eq. (71)"},{"comment":"The displayed Legendre-Fenchel transform I(s) is incorrect as written. For the function log f1(t) actually used, with the corrected definition log f1(t) = a log(t+a) + t log(1+a/t) + (1-a)log(1-a), the transform for s < log(1-a) should contain a term -a s and the denominator e^{-s}-1, not e^{-s}-1+a. Differentiating the printed formula gives a/(1-r+ar) (up to sign) rather than ar/(1-r); the measure claimed in (15) is the derivative of the correct transform. The final formula (15) appears to be correct, but the printed I(s) is internally inconsistent and must be corrected.","section":"§2.2, after Eq. (43)"}],"minor_comments":[{"comment":"In the line preceding Eq. (45), the expression \"1/n f_{k,n}\" should read \"(1/n) log f_{k,n}\"; the logarithm is missing from the displayed formula.","section":"§2.2, Eq. (45)"},{"comment":"The title and abstract contain spacing/OCR artifacts such as \"REPEA TED\" and \"POL YNOMIALS\"; these should be cleaned up in the final version.","section":"Title and abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper has a clear and interesting message, and the Kac and elliptic computations appear correct. However, the central Theorem 2 is only sketched, and the lower-bound estimate (71) is the key unproved ingredient. If the authors can supply a complete proof of Theorem 2 under A2, the paper is likely publishable. The Legendre-Fenchel transform error in §2.2 is local and should be straightforward to fix. I recommend major revision rather than rejection because the identified gap is fillable within the manuscript's scope and there is no evidence of a fundamental flaw in the main results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The paper asks the right question—what happens to zeros of random polynomials when the number of derivatives grows with the degree—and gives a clean answer for Kac polynomials. Previous work only handled fixed-order derivatives. The phase diagram depending on lim N_n/n, with explicit limiting measures and rescaling limits, is genuinely new. The Kac coefficient estimates via Stirling are careful, the counterexample showing A1 is insufficient in the N_n/n→1 regime is honest, and the elliptic polynomial computation is concrete. This is a solid paper in intention and execution, modulo one gap.\n\nThe gap is Theorem 2. The appendix explicitly labels the proof a sketch. The upper bound (70) is derived, but the lower bound (71) is asserted: the set J from [9] is said to have at least |J|/2 L_n points after intersection with [0,(T0−δ_n)L_n], and the rest follows by replacing n with L_n. That is not a routine substitution. A2 lacks the R0 lower-bound condition from [9]; the index cutoff depends on δ_n; and the small-ball step needs rechecking with L_n in place of n. No details are supplied. Since Theorems 3, 5, and 6(2) all route through Theorem 2, the phase-transition conclusions are not fully supported as written. I suspect the gap is fixable—the heuristics look right—but it is a real gap.\n\nMinor issue: the Legendre-Fenchel transform printed in Section 2.2 is wrong as written—it does not differentiate to the measure claimed just below. The measure is correct, so this is a typo, but it should be fixed.\n\nBottom line: this deserves a serious referee. The main results are important and likely true; the proof of Theorem 2 needs to be completed. I'd send it to peer review, with a request that the referee ask for a full proof of (71). Anyone working on zero distributions of random polynomials will want to read this, even in its current conditional form.","headline":"The Kac phase diagram for repeated derivatives is new and likely correct, but the main results all rest on a general theorem whose proof is a sketch with an unproved lower bound.","tokens_in":20096,"tokens_out":5948,"would_cite":true,"duration_ms":54207,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["30C15","60F10","60G57"],"pacs":[],"model":"deepseek-v4-flash","headline":"Zeros of Kac polynomials that have been differentiated $N_n$ times have a limiting global distribution that depends only on the limiting ratio $N_n/n$, with a collapse-and-rescaling regime when the ratio tends to 1.","keywords":["random polynomials","Kac polynomials","zeros of random polynomials","repeated differentiation","phase transition","rescaling limits","Legendre-Fenchel transform","elliptic polynomials"],"falsifier":"Simulate zeros of $K_n^{(N_n)}$ for $N_n=\\lfloor n/2\\rfloor$ with, say, i.i.d. standard complex Gaussian coefficients: the theorem predicts $(1/D_n)\\mu^K_{D_n}(\\mathbb{D}_r)\\to ar/((1-a)(1-r))$ for $r<1-a$ and $1$ for $r\\ge1-a$; a stable deviation from this radial curve at large $n$ would refute the central claim.","tokens_in":18999,"feed_emoji":"🎲","tokens_out":11948,"duration_ms":115769,"temperature":0.7,"pith_summary":"Repeatedly differentiating a random polynomial changes where its zeros live, and this paper shows that for the classical Kac polynomials the change is controlled entirely by the limiting ratio $N_n/n$ of the number of derivatives to the degree. If the ratio tends to $0$, the zeros still converge to the uniform measure on the unit circle. If it tends to $a\\in(0,1)$, the limiting zero measure becomes a rotationally invariant measure with $\\mu^K_a(\\mathbb{D}_r)=ar/((1-a)(1-r))$ for $r<1-a$ and $1$ beyond that radius. If the ratio tends to $1$, the zeros collapse to the origin, and after rescaling by $n/D_n$ they have the density $1/(2\\pi|z|)$ on the unit disk. This matters because it identifies the critical growth order at which unit-circle clustering fails and shows that the limiting shape changes abruptly rather than gradually.","feed_headline":"Random polynomial zeros leave the unit circle as derivatives pile up","feed_subtitle":"For Kac polynomials the limit law is set by N_n/n; clustering survives only when N_n/n→0.","key_machinery":"The load-bearing object is the modified coefficient-asymptotics theorem, stated as Theorem 2: for random polynomials whose coefficient magnitudes are approximated by a profile $p(k/L_n\\wedge T_0)$, the normalized zero measure converges to a rotationally invariant measure determined by the Legendre-Fenchel transform of $-\\log p$, namely $\\mu(\\mathbb{D}_r)=I'(\\log r)$. The proof is sketched in Appendix A and adapts the upper- and lower-bound estimates of the cited theorem, with the lower-bound estimate (71) asserted to follow by replacing $n$ with $L_n$. For each case, the paper chooses $L_n$, a profile $p$, and a rescaling factor, then uses standard factorial asymptotics to show that the derivative coefficients $f_{k,n}=(k+N_n)!D_n!/(k!n!)$ fit the profile, so Theorem 2 delivers the limiting measure. The rescaling cases work by multiplying the coefficients by $R_n^{D_n-k}$ or $R_n^{-(D_n-k)/2}$ so that the dominant coefficient terms obey a new profile.","core_discovery":"The paper establishes that for Kac polynomials $K_n(z)=\\sum_{k=0}^n \\xi_k z^k$ with i.i.d. nondegenerate coefficients satisfying $\\mathbb{E}\\log(1+|\\xi_0|)<\\infty$, the limiting global zero distribution of the $N_n$-th derivative is governed by $a=\\lim N_n/n$. If $a=0$, then $(1/D_n)\\mu^K_{D_n}$ converges in probability to the uniform measure on the unit circle. If $a\\in(0,1)$, it converges to the rotationally invariant measure $\\mu^K_a$ with $\\mu^K_a(\\mathbb{D}_r)=\\frac{ar}{(1-a)(1-r)}$ for $0<r<1-a$ and $1$ otherwise. If $a=1$ and $D_n=n-N_n\\to\\infty$, the zeros collapse to the origin globally, while the rescaled polynomials $K_n^{(N_n)}(z/R_n)$ with $R_n=n/D_n$ have limiting zero density $\\frac{1}{2\\pi|z|}$ on the unit disk. If the residual degree $D_n=m$ is fixed, scaling by $n$ gives convergence in distribution to the random zeros of $\\sum_{k=0}^m \\xi_k z^k/k!$. For general random polynomials whose coefficients satisfy a profile assumption, the $a=0$ and $a\\in(0,1)$ results carry over explicitly, while for $a=1$ the random elliptic polynomials give a different rescaling rate, $\\sqrt{D_n/n}$, with density $r(\\sqrt{4+r^2}-r)/2$, showing that the profile assumption alone does not determine the rescaling.","pith_inferences":["Beyond the paper, the two rates $D_n/n$ and $\\sqrt{D_n/n}$ suggest a general principle: the rescaling rate is set by the slowest coefficient growth near the top of the degree, so the quantities $\\eta_n$ and $b_n$ introduced in the paper could predict when zeros collapse and at what speed.","One could conjecture a one-parameter family of limiting densities that interpolates between the $a\\in(0,1)$ measure and the $1/(2\\pi|z|)$ disk density as $N_n/n\\to1$, with the two-step zooming heuristic in the paper indicating how the interpolation should look.","The fixed-degree limits for Kac and elliptic polynomials provide concrete test beds: perturbing the coefficient profile near the top of the degree while keeping the profile assumption intact should change the random limit in a way that tracks these baseline polynomials."],"forward_implications":["For Kac polynomials, the classical clustering of zeros near the unit circle survives repeated differentiation exactly when $N_n/n\\to0$; any positive limiting ratio destroys it.","When $N_n/n\\to1$ with $D_n\\to\\infty$, typical zeros of $K_n^{(N_n)}$ lie at distance of order $D_n/n$ from the origin, and after that rescaling the limiting density is $1/(2\\pi|z|)$ on the unit disk.","For fixed residual degree $m$, the scaled zeros have a random limit described by the zeros of $\\sum_{k=0}^m \\xi_k z^k/k!$, rather than a deterministic measure.","For general random polynomials satisfying the coefficient-profile assumption, the $N_n/n\\to0$ and $N_n/n\\to a\\in(0,1)$ limits are explicit in terms of the original profile, while the $N_n/n\\to1$ rescaling is not determined by the profile alone, as the Kac and elliptic cases give different rates ($D_n/n$ versus $\\sqrt{D_n/n}$)."],"supporting_citations":[{"why":"Supplies the asymptotic zero-distribution theorem for random analytic functions that Theorem 2 adapts; the upper- and lower-bound estimates in its proof are the template.","marker":"[9]"},{"why":"Establishes the almost-sure convergence of Kac zero empirical measures to the uniform measure on the unit circle, the baseline for the $N_n/n\\to0$ case.","marker":"[5]"},{"why":"Provides the classical complex-zero result for Kac polynomials and serves as a benchmark for the clustering property.","marker":"[11]"},{"why":"Treats roots of random polynomials with logarithmic tails, motivating condition (3) and the discussion of what happens when that condition fails.","marker":"[8]"},{"why":"Gives distributional results for zeros of Kac-type random polynomials and supports the framing around condition (3).","marker":"[6]"}],"fun_headline_variants":["Derivative order ratio caps where Kac zeros cluster","Random polynomial zeros: the limit of N_n/n decides","High-order derivatives: zeros collapse then rescale","Kac zeros: clustering fails if derivatives grow linearly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument rests on the assumption that Theorem 2 is correct even though its proof is only sketched; in particular, the lower-bound estimate (71), asserted to follow from [9] by replacing $n$ with $L_n$, has to hold for every main limit theorem to go through.","fun_headline_variants_meta":{"raw":{"variants":["Derivative order ratio caps where Kac zeros cluster","Random polynomial zeros: the limit of N_n/n decides","High-order derivatives: zeros collapse then rescale","Kac zeros: clustering fails if derivatives grow linearly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"total_tokens":1545,"prompt_tokens":1098,"completion_tokens":447,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":714,"completion_tokens_details":{"reasoning_tokens":384}},"tokens_in":714,"tokens_out":447,"duration_ms":4701,"temperature":1.0,"reasoning_tokens":384,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:33:39.007210+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate zeros of $K_n^{(N_n)}$ for $N_n=\\lfloor n/2\\rfloor$ with, say, i.i.d. standard complex Gaussian coefficients: the theorem predicts $(1/D_n)\\mu^K_{D_n}(\\mathbb{D}_r)\\to ar/((1-a)(1-r))$ for $r<1-a$ and $1$ for $r\\ge1-a$; a stable deviation from this radial curve at large $n$ would refute the central claim.","supporting_citations":[{"cited_title":"Hough , M","cited_arxiv_id":null,"evidence_quote":"Supplies the asymptotic zero-distribution theorem for random analytic functions that Theorem 2 adapts; the upper- and lower-bound estimates in its proof are the template."},{"cited_title":"Farmer and M","cited_arxiv_id":null,"evidence_quote":"Establishes the almost-sure convergence of Kac zero empirical measures to the uniform measure on the unit circle, the baseline for the $N_n/n\\to0$ case."},{"cited_title":"Ibragimov and D.N","cited_arxiv_id":null,"evidence_quote":"Provides the classical complex-zero result for Kac polynomials and serves as a benchmark for the clustering property."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Treats roots of random polynomials with logarithmic tails, motivating condition (3) and the discussion of what happens when that condition fails."},{"cited_title":"Feng , Zeros of derivatives of Gaussian random polynomials on plane domain, preprint","cited_arxiv_id":null,"evidence_quote":"Gives distributional results for zeros of Kac-type random polynomials and supports the framing around condition (3)."}],"review_version":1}