{"id":"fe4bbe16-1759-416e-8423-5fa7bc58aa22","arxiv_id":"2507.11028","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"In cylindrical Hastings-Levitov(0) aggregation, the expected one-arm domination time is of order N^2/λ^3, with an exponential tail, and the expected number of trees is asymptotically π N/λ.","lead":"This paper proves that in a cylindrical version of diffusion-limited aggregation, the time until one growing arm dominates all others is proportional to the square of the cylinder width divided by the cube of the particle size, and that this time has an exponential tail. The result gives the first rigorous bounds for a central observable in this model and also determines the expected number of arms.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Upper-tail proof of Theorem 1.4 has an invalid bound at (4.6): the term 200δ²x is not exponentially small for x≍δ^{-2}, so the claimed C9e^{-Kx/2} does not follow as written.","rationale":"I examined the proof of the central expectation result, Theorem 1.3, in detail. The interval-length martingale, the second-moment estimates of Lemma 2.3, the exit-time bounds, and the marked-configuration coupling of Proposition 3.4 all check out. The lower bound's partition argument is valid because each tree has a root in exactly one of the two semicircles, and the upper bound's restart procedure has i.i.d. sessions by the strong Markov property, with the stated cδ success probability producing the δ^{-3} particle count. No fatal flaw is apparent in the N²/λ³ estimate. The reader's flagged weakness about Proposition 2.1 is legitimate: that proposition alone does not construct the unique infinite tree, and the manuscript's claim of a self-contained proof there is an overstatement; however, [NT12] supplies the missing existence/uniqueness, and the upper-bound construction itself produces a.s. an eventually dominant color. My concrete objection is to the upper-tail proof of Theorem 1.4: the passage from (4.6) to the claimed exponential bound uses a false suppression of the 200δ²x term. This is a proof gap in an advertised main result, not a contradiction of the expectation theorem, and it is repairable by replacing the crude NUM tail with the geometric tail. Therefore the appropriate disposition remains conditional, matching the reader's verdict.","tokens_in":27992,"tokens_out":46917,"duration_ms":589590,"concrete_test":"Take δ=10^{-2}, x=δ^{-2}=10^4 and recompute the final bound of (4.6) using the displayed intermediate expressions: the crude term 200δ²x equals 2, whereas C9 e^{-Kx/2}=C9 e^{-50K}; no constant C9 independent of δ can dominate 2 by an exponentially small quantity. Then replace P(NUM>k) by the geometric tail (4δ²)^{k+1}/(1-4δ²) and verify that P(Σ_{j=0}^{NUM} bT_j > δ^{-3}x) decays as e^{-c x} for this δ,x. If it does, the exponential upper tail is recoverable by a local patch.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equations (4.6)–(4.8) are the only place where the exponential upper tail of υ is extracted. At (4.6), the authors bound P(NUM>[x]+1) by the crude term 200δ²x and then assert the whole expression is ≤ C9 e^{-Kx/2}, citing that δx decays faster than e^{-x}. For x ≍ δ^{-2}, the additive term is of order one, while the asserted right-hand side is exp(-const/δ²); no choice of C9 independent of δ makes the displayed inequality true. Moreover δ²x is not small in that regime. The fix is to use the honest tail P(NUM>k) ≤ (4δ²)^{k+1}/(1-4δ²), which is exponentially small in x log(1/δ) and does rescue the estimate. The expectation result Theorem 1.3 is unaffected, and the gap is repairable, but Theorem 1.4's upper tail is not proved by the text as written. Separately, the statement in Definition 1.2 that Proposition 2.1 gives a self-contained proof of the unique infinite tree is an overstatement: Proposition 2.1 establishes only that every interval length converges to 0 or 2π, and existence/uniqueness of a persistent tree is supplied by [NT12] (and later by the success event of the upper-bound construction).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the cylindrical Hastings-Levitov(0) process on a cylinder of width N with particle size λ. It defines the one-arm domination time υ_{N,λ} as the first time after which only the unique infinite tree receives particles, and proves that E[υ_{N,λ}] is of order N²/λ³ (Theorem 1.3), that υ_{N,λ} possesses exponential upper and lower tails (Theorem 1.4), that the tree completion time satisfies E[ω_{N,λ}] ≤ (1+ε) log N/(2λ) (Theorem 1.5), and that the expected number of trees tends to πN/λ (Theorem 1.7). The proofs are based on a pull-back interval process on the unit circle, on marked configurations coding the intervals of all trees, and on a backward-process recursion.","tokens_in":28280,"tokens_out":15629,"duration_ms":198838,"significance":"If the results are correct, this is a substantive contribution to the rigorous theory of off-lattice aggregation: it gives the first sharp quantitative control of the time to single-arm domination in a continuum DLA-type model, with universal constants and no fitted parameters. The interval characterization and the marked-configuration coupling are natural and potentially reusable tools. I emphasize that the main expectation theorem is supported by a martingale structure and explicit second-moment estimates that are largely self-contained; the exponential tail currently has a gap in the written proof (see major comment 1) but appears repairable with the authors' own estimates. The paper also builds on [NT12] for the existence of the unique infinite tree, although the text overstates the degree to which that fact is reproved.","major_comments":[{"comment":"The inequality '≤ C9 e^{-Kx/2}' does not follow from the preceding display. The additive term 200δ²x is of order one when x ≍ δ^{-2}, while the asserted right-hand side is exponentially small in δ^{-2}; for any C9 independent of δ the displayed bound fails in that regime. This step is the only place where the exponential upper tail of the sum of the bT_j variables is extracted, so the upper bound in Theorem 1.4 is not proved as written. A fix is available: replace the crude bound P(NUM>[x]+1)≤200δ²x by the honest tail P(NUM>k)≤(4δ²)^{k+1}/(1-4δ²), which is exponentially small in k log(1/δ) and can be absorbed into the desired e^{-Kx} form after choosing K appropriately. The expectation result Theorem 1.3 is unaffected.","section":"Section 4, Eq. (4.6)"},{"comment":"The parenthetical claim that Proposition 2.1 supplies a self-contained proof of the uniqueness of the infinite tree is an overstatement. Proposition 2.1(3) shows that the length of a tracked interval converges almost surely to 0 or 2π; it does not construct a persistent tree nor prove that exactly one tree survives forever. The existence and uniqueness of the infinite tree is imported from [NT12] and is also used implicitly in the success event of the upper-bound construction. This should be stated accurately; if [NT12] is assumed, the definition of υ_{N,λ} is legitimate, but the text should not claim that Proposition 2.1 reproves it.","section":"Definition 1.2 and Proposition 2.1"},{"comment":"The recursion X_{k+1}=1+Σ_{j=1}^{X_k} 1_{B_j} and the subsequent conditional expectation E[X_{k+1}|G_k]=1+(1-2/π arctan(δ/√(1-δ²)))X_k presuppose that each tree in the backward process has a unique base interval whose covering set under a new backward slit has length exactly 4 arctan(δ/√(1-δ²)), independently of the tree. This uniqueness is not proved anywhere in Section 6; without it, a tree could intersect the boundary in several intervals, and the probability that a new backward slit covers it would not be linear in the tree count. The authors should add a lemma establishing the unique-base-interval property for the backward construction, or justify it from the marked-configuration coupling, since Theorem 1.7 and Corollary 1.8 rely on this linearity.","section":"Section 6, recursion for X_k"}],"minor_comments":[{"comment":"The displayed identity |Sinv_x(I)|-|I| = ∫_{I^c}(1-Dθ)dθ is not correct as written for x∈I; the increment is nonnegative and the formula should include the endpoint contribution from Sinv_x({x}). The proof should be re-derived with the correct expression.","section":"Proof of Proposition 2.1(3)"},{"comment":"The notation PPI appears repeatedly in Section 2 without being defined; it presumably denotes the law of the embedded interval chain and should be defined at first use.","section":"Section 2 notations"},{"comment":"The Chernoff/ Bernstein estimate in the proof of Lemma 2.2(7) is applied with parameters that should be displayed more explicitly; the resulting constants 200 and the range m≥200 seem to rely on δ being small, which is fine, but the reader should be able to verify the numeric domination step.","section":"Eq. (2.14)"},{"comment":"The sentence 'Note that δx decays faster than e^{-x}' is misleading: δ²x is not small in the regime x≍δ^{-2}. This is part of the error in Eq. (4.6) and should be corrected along with the bound.","section":"Section 4, sentence near Eq. (4.6)"}],"recommendation":"major_revision","confidential_remarks":"The invalid inequality at (4.6) is the main technical obstacle; the fix suggested by the authors' own bound (3.16) is straightforward but must be written out, because the text as it stands fails to prove the upper tail of Theorem 1.4. I also recommend asking the authors to clarify the role of [NT12] and to prove the unique-base-interval property in Section 6; these are load-bearing but not reasons to reject. If the fixes are made, the paper would be a solid contribution appropriate for the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Main thing to know: the core result E[υ] ≍ N²/λ³ for one-arm domination in CHL(0) looks solid and is genuinely new. The interval-variation and marked-configuration machinery is a real contribution; it turns tree dynamics into a tractable Markov chain on boundary intervals, and the tree-density computation leading to E[N∞] = π/(2 arctan(...)) is clean and useful.\n\nWhat is good: Theorem 1.3 is well supported by the martingale exit estimates and optional stopping. The upper-bound proof via marked intervals is involved but coherent. The lower tail of Theorem 1.4 also appears fine, using the uniform interior-range event and Poisson time estimates.\n\nSoft spots, in order of severity. First, the upper-tail part of Theorem 1.4 has a real gap at (4.6). The text bounds P(NUM > [x]+1) by the crude term 200δ²x and then asserts the whole expression is ≤ C9 e^{-Kx/2}, citing that δx decays faster than e^{-x}. For x ≍ δ^{-2}, the additive term is order one while the claimed right-hand side is e^{-c/δ²}; no constant C9 independent of δ makes that inequality true. The stress-test is correct. The fix is also correct: use the honest tail P(NUM > k) ≤ (4δ²)^{k+1}/(1−4δ²), which gives e^{-c x log(1/δ)} and rescues the estimate. So the theorem is probably true, but the text as written does not prove the upper tail. This is a load-bearing defect in that one subsection, not in the main expectation result.\n\nSecond, Definition 1.2 claims a self-contained proof of the unique infinite tree in Proposition 2.1. That is an overstatement: Proposition 2.1 shows any tracked interval length converges to 0 or 2π, but existence and uniqueness of the persistent tree is imported from [NT12] and used critically in the definition of υ. Not fatal, but the text should not claim self-containment.\n\nThird, Section 6's backward-process tree count assumes each tree has a unique base point interval on T¹; this is intuitive and probably true, but not proved, and the recursion (6.1) relies on covering lengths being 4 arctan(δ/√(1−δ²)) independent of tree shape. The additivity argument is fine; just flag the missing justification. There is also a misprinted identity in the proof of Proposition 2.1(3), harmless but should be cleaned.\n\nVerdict: this paper deserves a serious referee. The main scaling law and the technique are worth publishing, and the tail gap is repairable. I would accept a conditional review and ask the authors to fix (4.6) and soften the self-containment claim.","headline":"Genuinely new scaling law with a repairable gap in the upper tail: the main theorem stands, but (4.6) needs the honest binomial tail.","tokens_in":28801,"tokens_out":1927,"would_cite":true,"duration_ms":21983,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","60D05","60J05"],"pacs":[],"model":"deepseek-v4-flash","headline":"In the cylindrical Hastings–Levitov process, the expected time until one tree permanently dominates is of exact order N²/λ³.","keywords":["one-arm domination time","Cylindrical Hastings-Levitov(0)","diffusion-limited aggregation","conformal slit maps","harmonic measure","interval Markov chain","random growth processes","tree coalescence"],"falsifier":"Run the rescaled CHL(1, λ/N) process in simulation, record the first particle index after which only one tree is ever hit again, and compare mean takeover times for several widths $N$ at fixed $\\lambda$: Theorem 1.3 requires $\\mathbb{E}[\\upsilon]$ to lie between two universal constants times $N^2/\\lambda^3$ with no residual $N$-dependence. A cheaper and sharper check targets the microscopic estimate: measure the one-step variance of the length of a $\\delta$-length interval in the embedded chain, $\\delta \\approx \\lambda/(2N)$, and verify it equals $\\frac{32}{3\\pi}\\delta^3 - 2\\delta^4 + O(\\delta^5)$; since every exit-time bound in the paper is built on this $\\delta^3$ scale, a failure of the cubic law there would falsify the whole hierarchy.","tokens_in":27803,"feed_emoji":"🌳","tokens_out":17741,"duration_ms":171320,"temperature":0.7,"pith_summary":"This paper studies the cylindrical Hastings–Levitov(0) process, an off-lattice model of diffusion-limited aggregation on a cylinder of width $N$ fed by particles of size $\\lambda$. As the process runs, almost surely a single infinite tree survives and eventually every new particle attaches to it; the paper defines the one-arm domination time $\\upsilon_{N,\\lambda}$ as the first time after which only that tree receives particles, and asks how large it is. The main theorem gives the exact order: universal constants $c, C > 0$ such that $cN^2/\\lambda^3 \\le \\mathbb{E}[\\upsilon_{N,\\lambda}] \\le CN^2/\\lambda^3$, together with an exponential tail around this scale. The proof's insight is that the whole coalescence can be read off a one-dimensional object — the length of a boundary interval pulled back by the random slit maps — whose increments have variance of order $\\lambda^3/N^3$; gambling-ruin exit estimates then convert this microscopic scale into the macroscopic $N^2/\\lambda^3$ law. A curious reader should care because this is one of the few off-lattice growth models where a dynamical quantity of this global character is determined to within universal constants.","feed_headline":"N²/λ³ time until one arm permanently captures all new particles","feed_subtitle":"The expected takeover time in cylindrical Hastings–Levitov(0) is exactly of this order, with an exponential tail.","key_machinery":"The argument runs on two coupled mechanisms. First is the interval characterization: after rescaling the width-$N$ cylinder to a unit circle, each slit map is inverted on boundary arcs, and the length $|I_k|$ of any tracked interval evolves as a Markov jump process. That length is a positive martingale converging almost surely to $0$ or $2\\pi$ (Proposition 2.1), and its one-step variance is of order $\\delta^3$, where $\\delta \\asymp \\lambda/N$; explicitly $m_\\delta = \\frac{32}{3\\pi}\\delta^3 - 2\\delta^4 + O(\\delta^5)$ (Lemma 2.3). Lemma 2.4 converts this variance scale into gambler's-ruin exit times of order $\\delta^{-3}$, the quantitative heart of the $N^2/\\lambda^3$ result. Second is the marked-configuration coupling (Definitions 3.2–3.3): each tree is represented by a boundary interval whose length equals that tree's share of the harmonic measure — the chance that the next incoming particle attaches to that tree — so the event 'everything is captured by one tree' becomes 'one marked interval occupies almost the whole circle.' The upper bound launches independent explorations of $\\delta$-length intervals, each needing $\\sim\\delta^{-2}$ steps and succeeding with probability $\\sim\\delta$; the lower bound compares domination against a semicircular interval's first exit from a macroscopic range.","core_discovery":"The paper's central claim is Theorem 1.3: for every fixed particle size $\\lambda$ there are universal constants $c, C > 0$ and a threshold $N_0(\\lambda)$ such that for all cylinder widths $N \\ge N_0(\\lambda)$, the one-arm domination time $\\upsilon_{N,\\lambda}$ — the first time after which the unique infinite tree receives every new particle — satisfies $cN^2/\\lambda^3 \\le \\mathbb{E}[\\upsilon_{N,\\lambda}] \\le CN^2/\\lambda^3$. The same statement comes with an exponential tail (Theorem 1.4): $C_3 e^{-C_4 t} \\le P(\\upsilon_{N,\\lambda} > \\lambda^{-3}N^2 t) \\le C_1 e^{-C_2 t}$. The mechanism producing this is that the pull-back of any boundary interval under the random slit maps has length that is a positive martingale converging almost surely to $0$ or $2\\pi$, with one-step variance of order $\\delta^3$ where $\\delta \\asymp \\lambda/N$; gambler's-ruin exit estimates then give a $\\delta^{-3}$ particle count, and at Poisson rate $2\\pi N$ this is $N^2/\\lambda^3$ time. Two further results follow from the same marked-interval machinery: the last time a new tree ever appears has expectation at most $(1+\\varepsilon)\\log N/(2\\lambda)$, and the total number of trees converges almost surely to $N_\\infty$ with $\\mathbb{E}[N_\\infty] = \\frac{\\pi}{2\\arctan(\\delta/\\sqrt{1-\\delta^2})} \\sim \\frac{\\pi N}{\\lambda}$, implying each particle spawns exactly one child in expectation.","pith_inferences":["The same one-step variance $m_\\delta \\asymp \\delta^3$ should control the fluctuations of every tree's harmonic-measure share on the $N^2/\\lambda^3$ time scale; the paper does not state this consequence, but Lemma 2.3 supplies the quantity needed to attempt it.","Replacing $(1+\\varepsilon)$ by $1+o(1)$ in Theorem 1.5 would prove the conjecture of Remark 1.6, $\\lim_N \\mathbb{E}[\\omega_{N,\\lambda}]/\\log N = 1/(2\\lambda)$; the Section 5 recursion is a plausible route to that sharper constant.","Numerical simulation of CHL(0) over a range of widths and particle sizes should collapse the takeover time onto a single curve in $\\lambda^3 t/N^2$; the paper contains no such experiments, so this is a direct testable prediction.","The paper is explicit that no rigorous link between CHL and lattice DLA on a cylinder is known; transferring the $N^2/\\lambda^3$ law to the lattice model would require a new coupling, and would make the domination time there tractable for the first time."],"forward_implications":["At fixed particle size $\\lambda$, doubling the cylinder width multiplies the expected domination time by four; the law $N^2/\\lambda^3$ identifies the microscopic slit-map coalescence scale as the single determinant of macroscopic takeover.","The exponential tail makes domination a sharp cut-off phenomenon: with probability approaching 1 the takeover is complete within a bounded factor of $N^2/\\lambda^3$, and the matching lower tail shows no provably earlier universal takeover time exists.","The expected total number of trees is asymptotically $\\pi N/\\lambda$, so the forest has a well-defined linear density in the width, and each particle has exactly one child on average — the genealogy of trees is balanced at criticality.","New trees stop appearing after at most $(1+\\varepsilon)\\log N/(2\\lambda)$ time in expectation, so the active frontier of the process shrinks logarithmically slowly relative to the quadratic domination time."],"supporting_citations":[{"why":"Supplies the result that a single infinite tree grows forever in the process, the premise that defines the one-arm domination time, plus the coalescing Brownian flow limit used as the coalescence-scale guide.","marker":"[NT12]"},{"why":"Defines the cylindrical Hastings–Levitov model and its slit maps, including the growth-rate scaling that the paper rescales to a unit cylinder with particle size λ/N.","marker":"[PZ23]"},{"why":"Supplies the distributional identity that the backward CHL process equals the forward process at fixed time, which the proof of the expected tree count relies on.","marker":"[BPT22]"},{"why":"Introduces the conformal slit-map aggregation construction; the paper's interval machinery repeatedly inverts these slit maps, so the whole argument is built on this construction.","marker":"[HL98]"}],"fun_headline_variants":["One-arm domination time scales as N²/λ³","Exponential tail for one-arm takeover in CHL0","N²/λ³ time until one tree claims all particles","CHL0: unique arm persists after N²/λ³","Takeover time in cylindrical HL(0): N²/λ³"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central quantity presumes the cited result [NT12] that a single infinite tree exists and is unique in this process; the paper's own Proposition 2.1 proves only that a tracked interval's length converges almost surely to $0$ or $2\\pi$, which by itself does not construct the persistent tree that defines $\\upsilon_{N,\\lambda}$.","fun_headline_variants_meta":{"raw":{"variants":["One-arm domination time scales as N²/λ³","Exponential tail for one-arm takeover in CHL0","N²/λ³ time until one tree claims all particles","CHL0: unique arm persists after N²/λ³","Takeover time in cylindrical HL(0): N²/λ³"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000527,"raw_usage":{"total_tokens":2588,"prompt_tokens":1038,"completion_tokens":1550,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":1463}},"tokens_in":654,"tokens_out":1550,"duration_ms":15080,"temperature":1.0,"reasoning_tokens":1463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:21:25.515498+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the rescaled CHL(1, λ/N) process in simulation, record the first particle index after which only one tree is ever hit again, and compare mean takeover times for several widths $N$ at fixed $\\lambda$: Theorem 1.3 requires $\\mathbb{E}[\\upsilon]$ to lie between two universal constants times $N^2/\\lambda^3$ with no residual $N$-dependence. A cheaper and sharper check targets the microscopic estimate: measure the one-step variance of the length of a $\\delta$-length interval in the embedded chain, $\\delta \\approx \\lambda/(2N)$, and verify it equals $\\frac{32}{3\\pi}\\delta^3 - 2\\delta^4 + O(\\delta^5)$; since every exit-time bound in the paper is built on this $\\delta^3$ scale, a failure of the cubic law there would falsify the whole hierarchy.","supporting_citations":[],"review_version":1}