{"id":"fe604268-cd84-442d-a7dc-b9b3dc47035b","arxiv_id":"1908.01370","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For the additive integer urn, the empirical label distribution, scaled by the number of balls, converges to an exponential distribution multiplied by a random martingale limit A.","lead":"This paper analyzes an urn process where each new ball is labeled by the sum of two randomly drawn labels, and proves that after rescaling, the urn's empirical distribution converges to an exponential law whose scale is a random variable. It is a new limit theorem for an infinite-type, multi-drawing urn model that had not been analyzed before.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1.1 fails for the all-zero initial urn (A=0, P(A≠0)=0), and Lemma 2.1's nontriviality bound is zero whenever S_tau0=0, so the central claim is overclaimed as stated.","rationale":"The central contribution is a quenched exponential limit for multi-drawing infinite-type integer urns, with the exponential scale A being the martingale limit of the normalized mean. I checked the two places this could fail: the martingale convergence in Lemma 2.1 and the contraction argument in Theorem 3.4. The contraction argument is standard and not the main risk. The load-bearing weakness is the assertion P(A≠0)>0 in Theorem 1.1. For U0={0}, the process is frozen at zero, so the theorem is literally false; no qualifier in the statement excludes this. The only nontriviality proof uses a lower bound proportional to A_{τ0}, which vanishes whenever the initial sum is zero, including the {-1,1} configuration the paper highlights. The theorem is therefore overclaimed as stated and underproved for zero-sum starts. This is compatible with the reader's conditional verdict: the exponential-law mechanism need not collapse, but the statement must be amended (at least exclude all-zero starts) and a variance-based lower bound supplied for zero-sum configurations such as {-1,1}. A concrete check of E[A_3^2] for {-1,1} confirms the intended nontriviality is true there, so a correction is plausible rather than a fundamental flaw.","tokens_in":10411,"tokens_out":9614,"duration_ms":101025,"concrete_test":"Recompute the lower bound in Lemma 2.1 for the paper's featured start U0={-1,1}: because S_{τ0}=0 the bound is 0, whereas the one-step recursion gives E[A_3^2]=2/(12^2)=1/72>0. If this positive second moment propagates to all n (equivalently, if E[A^2]>0 can be shown from the R_n,Q_n recursions with R_2=Q_2=2), then the theorem is repairable by excluding only all-zero starts; if not, the nontriviality claim is more seriously under-supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is the nontriviality assertion of Theorem 1.1. The theorem is stated for 'any initial configuration,' but if U0 consists only of balls labeled 0, then every draw is 0, every added ball is 0, μ_n=δ_0, A=lim S_n/[n(n+1)]=0 a.s., and P(A≠0)>0 is false. This is not excluded anywhere in the statement. The only nontriviality argument, Lemma 2.1, shows E[A_n^2] ≥ ∏_{k=τ0}^{n-1}(1-2/(k^2+4k+4))·A_{τ0}; this lower bound is zero whenever S_{τ0}=0. It therefore fails for the paper's featured initial configuration {-1,1} as well as for all-zero urns. For {-1,1}, E[A_3^2]=1/72>0, so a repair is plausible, but Section 2 does not provide it. Thus Theorem 1.1 is false as stated and its proof does not establish P(A≠0)>0 for zero-sum starts; the exponential-limit mechanism itself is not the weak point.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies an urn process in which, at each step, two balls are drawn independently and with replacement and a new ball labeled by their sum is added. The main result, Theorem 1.1, asserts that for any initial configuration the rescaled empirical mean A_n = S_n/(n(n+1)) converges almost surely to a limit A, that P(A ≠ 0) > 0, and that a random draw Z_n from the urn satisfies the conditional limit law Z_n/n → A·Exp(1). The proof uses an L2-bounded martingale argument for the mean and the contraction method for recursive distributional equations, with a vector-valued generalization in Theorem 1.2.","tokens_in":10662,"tokens_out":6703,"duration_ms":72038,"significance":"If the main theorem is established for non-degenerate initial configurations, the paper provides an elegant rigorous analysis of a genuinely novel urn model that combines multiple draws with an infinite type space. The martingale identification of the random scale A and the reduction to a fixed-point equation are clean and convincing, and the use of known contraction-method results is appropriate. The vector-valued extension is a useful addition. The main obstruction is that Theorem 1.1 is false as stated for an all-zero initial urn, and the proof of the nontriviality assertion has a gap for zero-sum initial configurations, including the paper's featured {-1,1} example.","major_comments":[{"comment":"The theorem is stated for 'any initial configuration', but if U0 consists only of balls labeled 0, then every drawn ball and every added ball is 0, so μ_n = δ_0 and A = 0 almost surely. The assertion P(A ≠ 0) > 0 is then false. The statement must either exclude all-zero initial configurations or be modified to a form that remains true for them.","section":"Section 1.1, Theorem 1.1"},{"comment":"The lower-bound display in the proof of Lemma 2.1 reads E[A_n^2] ≥ ∏_{k=τ0}^{n-1}(1 - 2/(k^2+4k+4)) · A_{τ0}. As written this is dimensionally inconsistent: the right-hand side is nonpositive when A_{τ0} < 0 and is zero whenever S_{τ0} = 0. Even with the likely intended factor A_{τ0}^2, the argument gives no positive lower bound for zero-sum initial configurations. In particular, for the paper's featured initial configuration {-1,1}, S_{τ0}=0, so Lemma 2.1 does not establish the nontriviality assertion P(A ≠ 0) > 0 that is needed for Theorem 1.1. A separate argument is required for zero-sum starts.","section":"Section 2, Lemma 2.1"},{"comment":"Lemma 3.3 assumes X(i) ≥ 0 almost surely for all coordinates, but in the application the vector A can have negative coordinates (as in the {-1,1} initial configuration, where A can be negative). The characteristic-function computation actually extends to signed A, but the lemma as stated does not cover the case used in the proof of Theorems 1.1 and 1.2. The paper should state and prove the signed version, or explicitly condition on the sign of A.","section":"Section 3, Lemma 3.3"},{"comment":"The proof divides by A in the step 'Since \\tilde{X}_{I_n}/A converges in L2 to X ∼ Gamma(2,1)'. This presupposes A ≠ 0, but P(A ≠ 0) > 0 is exactly one of the assertions being proved. The proof should first establish nontriviality of A, or should handle the event {A = 0} separately, before dividing by A.","section":"Section 3.3, proof of Theorems 1.1 and 1.2"}],"minor_comments":[{"comment":"There is a typo in the remark: 'the the nth labeled added' should read 'the nth label added'.","section":"Section 1.1, Remark"},{"comment":"The notation A_τ0 is used ambiguously: it should be made explicit that the lower bound concerns A_τ0(i)^2 for each coordinate i, and the displayed inequality should be corrected to include the square on A_τ0.","section":"Section 2, Lemma 2.1"},{"comment":"In Theorem 1.1, A is defined as lim E[Z_n|μ_n]/(n+1), while Section 2 defines A_n = S_n/(n(n+1)). Since E[Z_n|μ_n] = S_n/n, these agree, but the paper would be easier to follow if this identity were stated explicitly before the theorem.","section":"Section 1.1 and Section 2"},{"comment":"The phrase 'converges in L2' for \\tilde{X}_{I_n}/A should be 'converges in the Wasserstein L2 metric' or 'converges in distribution with convergence of second moments', since the contraction method gives distributional convergence, not almost-sure convergence.","section":"Section 3.3"},{"comment":"The proof states that the integral equation for the characteristic function has a unique solution, but no justification or reference is given for this uniqueness. A citation or a short argument would be helpful.","section":"Section 3.1, proof of Lemma 3.3"}],"recommendation":"major_revision","confidential_remarks":"The all-zero initial configuration is a simple but genuine counterexample to Theorem 1.1 as stated, and the lower-bound display in Lemma 2.1 appears to have a missing square. Both are fixable within the scope of the paper, and the main contraction-method mechanism seems sound for non-degenerate starts. I would therefore not reject the paper, but the public claims need to be corrected and the zero-sum case needs a proper treatment before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my read. The paper gives the first limiting result for an urn that combines multiple draws per step with an infinite type space: if you draw two integers with replacement and add their sum, the empirical measure, normalized, converges to an exponential with a random scale. That is a real contribution. The martingale argument for the mean is straightforward and correct, the contraction method is standard but carefully applied, and the paper is honest that the distribution of the scale A is not identified. The related-work section is accurate; the model really does seem new, and the signal-game motivation is contextual rather than load-bearing.\n\nThe soft spot is exactly where the reader put it. Theorem 1.1 is stated for “any initial configuration.” The all-zero urn is a counterexample: every ball is 0, A=0 a.s., so P(A≠0)>0 fails. That is not an exotic corner case; it is the most degenerate possible start. Lemma 2.1, the only place nontriviality is proved, gives a lower bound for E[A_n^2] proportional to A_{tau0}, the initial value of the martingale. When S_{tau0}=0, including the paper's own {-1,1} configuration, that bound is zero and proves nothing. The bound should involve A_{tau0}^2, or the proof needs a different argument. For {-1,1} the claim is likely true, and the simulations suggest it, but the paper as written does not prove it. This is a fixable gap, not a collapse of the main idea.\n\nI also find Lemma 2.1's statement dimensionally odd: it claims an infimum over expectations greater than zero using a product times A_{tau0}, which is a random variable, not a positive constant. Even when S_{tau0}>0, the displayed inequality is at least poorly phrased. A referee should ask for a clean statement and a repair for the zero-sum case.\n\nBottom line: the exponential limit mechanism is sound, the paper deserves serious refereeing, and with a small amendment to the theorem and a corrected nontriviality proof it becomes a solid contribution. I would send it out.","headline":"A novel infinite-type, multi-draw urn limit with a clean exponential law, but Theorem 1.1 overclaims for zero-sum initial configurations and the nontriviality proof has a gap.","tokens_in":11163,"tokens_out":2633,"would_cite":true,"duration_ms":25855,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60F05","60G42","60C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"An urn that adds the sum of two random draws converges, after rescaling, to an exponential whose random scale is the martingale limit of the mean.","keywords":["urn models","random additions","integer labels","exponential limit law","contraction method","recursive distributional equations","martingale convergence","multi-draw urns"],"falsifier":"The all-zero urn is a direct counterexample to the theorem as stated: with initial labels $\\{0,0\\}$, every added ball has label $0$, so $A=0$ almost surely and $P(A\\neq 0)=0$. For the $\\{-1,1\\}$ configuration, the proof's lower bound for $\\mathbb{E}[A_n^2]$ is proportional to the initial mean, which is zero, so the argument that $A$ is nontrivial does not apply; simulating many runs and checking whether the empirical distribution of $A$ has positive mass away from $0$ would settle whether the theorem's conclusion still holds there.","tokens_in":10208,"feed_emoji":"🎲","tokens_out":14259,"duration_ms":130150,"temperature":0.7,"pith_summary":"This paper analyzes the $\\mathbb{Z}$-urn: balls carry integer labels, and at each step two balls are drawn independently with replacement and a new ball labeled by their sum is added. The main result, Theorem 1.1, asserts that for any starting configuration, if $Z_n$ is a uniformly random draw from the urn after $n$ balls, then $A=\\lim_{n\\to\\infty}\\mathbb{E}[Z_n\\mid \\mu_n]/(n+1)$ exists almost surely, $P(A\\neq 0)>0$, and the conditional law of $Z_n/n$ converges to that of $A$ times an $\\mathrm{Exp}(1)$ variable. The proof splits the claim in two: a martingale argument shows the rescaled mean converges, and a contraction-method argument shows the rescaled added ball converges to a Gamma distribution, which a uniform index then turns into an exponential. This matters because the model combines drawing multiple balls and having infinitely many ball types, and the limit law is simple enough to use.","feed_headline":"Sum-adding integer urns converge to a scaled exponential","feed_subtitle":"The random scale is set by the starting labels, so independent runs end up at different exponential shapes.","key_machinery":"The argument is carried by two mechanisms. First, the rescaled total $A_n=S_n/(n(n+1))$ is a martingale: adding a ball whose conditional mean is $2S_n/n$ leaves $A_n$ unchanged in expectation, and its second moment is bounded, so $A_n\\to A$ almost surely. Second, the scaled added ball $\\tilde{X}_n=X_n/n$ obeys the recursive distributional equation $\\tilde{X}_n\\stackrel{d}{=}(I_n^1/n)\\tilde{X}_{I_n^1}+(I_n^2/n)\\tilde{X}_{I_n^2}$, whose limit must satisfy $X\\stackrel{d}{=}U_1X_1+U_2X_2$ with $U_i$ uniform. The unique solution with mean $\\mu$ is $G\\cdot\\mu$ with $G\\sim\\Gamma(2,1)$ (Lemma 3.3); a uniformly chosen past index introduces an extra independent uniform factor $U$, and $U\\cdot G$ is exponential (Lemma 3.2). The contraction method in the Wasserstein $L^2$ metric (a notion of distance between probability laws that also tracks second moments) turns the recursion into a contraction and forces the convergence.","core_discovery":"On the paper's own terms, the discovery is the exponential limit law of Theorem 1.1. For the integer urn with any initial configuration, the martingale $A_n = S_n/(n(n+1))$ converges almost surely to a random variable $A$, and the conditional law of $Z_n/n$ given the urn history converges to $\\mathcal{L}(A\\cdot \\mathrm{Exp}(1))$. In particular, for the starting configuration $\\{-1,1\\}$, the limiting urn is supported on one side of zero, positive or negative according to the sign of $A$, which explains the one-sided exponential histograms the simulations show. The vector version, Theorem 1.2, says that the $n$-th added ball $X_n$, rescaled by $n$, converges to $G\\cdot A$ with $G\\sim\\Gamma(2,1)$; since a random draw is a uniform past ball, the factor $G$ is multiplied by an independent uniform variable and becomes exponential.","pith_inferences":["The identity $X\\stackrel{d}{=}U_1X_1+U_2X_2$ encodes the entire dependence on the initial state in the single random vector $A$; a reading the paper leaves implicit is that the two-draw addition rule has exactly one stable shape, a ray of scaled exponentials, and the initial labels only choose the ray and the scale.","For the signaling-game motivation discussed in the paper, the one-sided exponential limit means the long-run distribution of strategies is sign-selected: each realization ends up predominantly positive or predominantly negative, and the random scale $A$ acts as the strength of the convention, so independent runs produce different conventions.","A testable consequence of the vector theorem: with labels in $\\mathbb{Z}^d$, rescaled draws should collapse onto the random ray $\\{tA:t\\ge 0\\}$ (modulated by an exponential), so an empirical scatter plot of rescaled draws should look one-dimensional rather than filling the ambient space.","For $k>2$, the fixed-point equation loses its explicit Gamma solution, so the limiting law becomes a genuinely new family; simulating the $k=3$ and $k=4$ urns and comparing histograms with solutions of $\\sum_{i=1}^k U_iX_i$ would show how quickly the exponential shape deforms."],"forward_implications":["For the $\\{-1,1\\}$ start, the limiting draw is exponential on $(0,\\infty)$ or $(-\\infty,0)$ according to the sign of $A$; the urn's labels become almost all positive or almost all negative.","For vector-valued labels in $\\mathbb{Z}^d$, every coordinate of the rescaled added ball converges to the same $\\Gamma(2,1)$ factor multiplied by the coordinate's martingale limit, so the coordinates are asymptotically perfectly dependent.","The annealed limit of $Z_n/n$ is a mixture of exponentials with random scale $A$, and the mean of $A$ is determined by the initial configuration as $S_{\\tau_0}/(\\tau_0(\\tau_0+1))$.","For the $k$-draw variant, the limit satisfies $X\\stackrel{d}{=}\\sum_{i=1}^k U_iX_i$; a Gamma law no longer solves this identity, so the exponential shape observed at $k=2$ should be replaced by a family of limit laws indexed by $k$."],"supporting_citations":[{"why":"Introduces the group-valued urn model whose integer analogue this paper studies; supplies the original motivation and the uniform-limit baseline.","marker":"[SY05]"},{"why":"Provides the $L^2$ martingale convergence theorem used to prove that $A_n$ converges almost surely.","marker":"[Wil91]"},{"why":"Supplies the Wasserstein metric and optimal-coupling results on which the contraction method relies.","marker":"[BF81]"},{"why":"Contains the contraction-method theorem (Theorem 4.1 and Corollary 4.2) that the paper's Theorem 3.4 specializes and reproves.","marker":"[Nei01]"},{"why":"Gives the characterization of the exponential distribution used to identify the limit of the uniform-index mixture.","marker":"[KS88]"}],"fun_headline_variants":["Integer urn sums: random-scaled exponential limit","Sum-adding urn of integers yields random exponential scale","Exponential law with random scale for integer urns","Initial urn labels set the scale of exponential limit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theorem's nontriviality assertion depends on the initial urn not being the all-zero urn—if every starting label is $0$, the limit $A$ is identically $0$—and the proof's second-moment lower bound does not cover configurations whose initial labels sum to $0$, such as $\\{-1,1\\}$.","fun_headline_variants_meta":{"raw":{"variants":["Integer urn sums: random-scaled exponential limit","Sum-adding urn of integers yields random exponential scale","Exponential law with random scale for integer urns","Initial urn labels set the scale of exponential limit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000678,"raw_usage":{"total_tokens":3052,"prompt_tokens":882,"completion_tokens":2170,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":2108}},"tokens_in":498,"tokens_out":2170,"duration_ms":16544,"temperature":1.0,"reasoning_tokens":2108,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:15:51.390925+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The all-zero urn is a direct counterexample to the theorem as stated: with initial labels $\\{0,0\\}$, every added ball has label $0$, so $A=0$ almost surely and $P(A\\neq 0)=0$. For the $\\{-1,1\\}$ configuration, the proof's lower bound for $\\mathbb{E}[A_n^2]$ is proportional to the initial mean, which is zero, so the argument that $A$ is nontrivial does not apply; simulating many runs and checking whether the empirical distribution of $A$ has positive mass away from $0$ would settle whether the theorem's conclusion still holds there.","supporting_citations":[],"review_version":1}