{"id":"4bb90726-f73a-43d4-b48a-519ee56156fd","arxiv_id":"1908.10260","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For a multiple-entry reinforcement process, block sizes grow like t^beta, satisfy a mixed-normal CLT, and a single block persistently leads almost surely.","lead":"This paper introduces a broad class of reinforced random partition processes in which several numbers arrive at each step and join existing blocks with probability proportional to block size. The authors prove the asymptotic block sizes, central limit theorems, and that one block eventually becomes, and stays, the largest.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 6's proof uses a false binomial-tail inequality; the Mori supermartingale and the Lp part of Theorem 1 rest on it, but a corrected O(t^-2) bound appears to repair the gap.","rationale":"The reader's stated weakest assumption (A.2) is a hypothesis rather than a gap: the paper explicitly assumes it and it is used to get φ_t/t^β→b. The real soft spot is the proof of Lemma 6. The chain of inequalities in (13) uses C(r,j)≤2R and P(Δd≥j)≤2R(d/(Rt))^j; the second is plainly false for moderate R and p, and the first is also false. Since Lemma 6 is the only source of uniform moment bounds for the normalized cardinalities, the L^p part of Theorem 1 and the L^p statement in Theorem 4 are unsupported as written. I do not see a fatal consequence: the false bound can be replaced by the exact binomial identity E[C(Δd,j)|F] ≤ C(R,j)(d/(Rt))^j, and the ratio of binomial coefficients is ≤m^j/d^j, so the j≥2 contribution is O_{m,R}(t^{-j}) and sums to O(t^{-2}). The constant R(4m)^R in φ(m) can absorb the correction. Lemma 21's omitted proof is straightforward via I_n→I in probability. Other potential concerns (the citation [12, Lemma 2] in Proposition 7, the density arguments in Section 3.3) did not reveal a separate fatal flaw; the central argument is structurally sound. Hence the verdict remains CONDITIONAL rather than ACCEPT, because the proof gap should be fixed, but no change to the reader's verdict is needed.","tokens_in":56,"tokens_out":34298,"duration_ms":719083,"concrete_test":"Independently re-derive the j≥2 estimate in Lemma 6 using the binomial identity E[C(Δd,j)|F_t] = Σ_u g_{t+1}(u) C(u,j)(d/(Rt))^j and the ratio bound C(d+m-1,m-j)/C(d+m-1,m) ≤ m^j/d^j. If this yields a total j≥2 contribution ≤ C_{m,R}/t^2 for some explicit C_{m,R}, then Lemma 6 and its consequences (Lemma 8, L^p convergence in Theorems 1 and 4) go through with the constant in the definition of φ(m) adjusted. Also run a small enumeration over R≤20, d≤Rt≤1000 to confirm both printed inequalities in (13) fail, verifying that the gap is real.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The concrete load-bearing gap is in Lemma 6 (Section 2.2), not in the regularity assumption (A.2). To prove the Mori supermartingale, the proof bounds the j≥2 terms by C(d+m-1,m)·[m!/(m-j)!·(d-1)!/(d+j-1)!]·2R·P(Δd≥j), then uses P(Δd≥j|F_t)≤2R(d/(Rt))^j. The second inequality is false: for R=10, j=2, p=0.1, P(Bin(10,0.1)≥2)≈0.264>0.2=2R p^2; the bound C(r,j)≤2R is also false. Lemma 6 underpins Lemma 8 (uniform moment bounds for d_t/φ_t), hence the L^p part of Theorem 1 and the L^p convergence in Theorem 4. This is a genuine gap in the written proof. It is, however, repairable: using E[C(Δd,j)|F_t]≤C(R,j)(d/(Rt))^j and C(d+m-1,m-j)/C(d+m-1,m)≤m^j/d^j gives a j≥2 contribution O_{m,R}(t^{-j})≤O_{m,R}(t^{-2}), so the supermartingale holds with a larger t^{-2} constant in φ(m). The omitted proof of Lemma 21 is minor and does not affect the central argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces a class of reinforcement processes in which, at each time step, R new elements are added; a random subset joins existing blocks with probability proportional to current block sizes, and the remaining elements start new blocks. The main results are: Theorem 1, almost sure and L^p convergence of d_t^{(i)}/t^\\beta to a strictly positive limit \\xi^{(i)}; Theorem 2, a mixed-normal central limit theorem for the fluctuations; Theorem 3, almost sure existence of a single block with persistent leadership; and Theorem 4, corresponding convergence and CLT results for the maximal block. The proofs combine martingale arguments, exponential tail bounds, Mori supermartingales, an embedded-chain analysis, and a Lyapunov-function argument, with interpretations in terms of balls-and-bins, generalized Chinese restaurant processes, urn models, and preferential attachment graphs.","tokens_in":29537,"tokens_out":28744,"duration_ms":285442,"significance":"The model is a useful meta-model that unifies several reinforced random processes, and the qualitative picture—power-law growth, a phase transition at \\beta=1, and persistent leadership—is appealing and plausible under suitable regularity assumptions. The paper gives detailed martingale-based proofs and correctly isolates the universal parameter \\beta. However, as stated, the central CLT is false under the stated assumptions (A.1)-(A.2), and the proof of the Mori supermartingale contains a false binomial-tail estimate. These issues affect the main claims and require substantive revision. I credit the authors for the structural framework and for developing a martingale and embedded-chain approach that appears repairable once the regularity conditions are strengthened.","major_comments":[{"comment":"The proof of Lemma 6 uses the binomial-tail bound P(\\Delta d_t \\ge j | F_t) \\le 2R(d_t/(Rt))^j, which is false. For example, with R=10, j=2 and d_t/(Rt)=0.1, the true binomial tail is approximately 0.264, while 2R(d_t/(Rt))^j = 0.2. The accompanying bound on the binomial coefficient sum is also not generally valid. Since Lemma 6 underpins Lemma 8 and hence the L^p part of Theorem 1 and Theorem 4, this is a genuine gap. The gap appears repairable: using E[\\binom{\\Delta d_t}{j}|F_t] \\le C(R,j)(d_t/(Rt))^j together with \\binom{d_t+m-1}{m-j}/\\binom{d_t+m-1}{m} \\le m^j/d_t^j gives a contribution O_{m,R}(t^{-j}) \\le O_{m,R}(t^{-2}) for j\\ge 2, so a supermartingale with an enlarged t^{-2} constant in \\phi_t^{(m)} should hold.","section":"Section 2.2, Lemma 6"},{"comment":"The CLT in Theorem 2 is not implied by assumptions (A.1)-(A.2), and in fact is false under them. Writing X_t=d_t/\\phi_t, \\xi=b\\zeta, one has t^{\\beta/2}(d_t/t^\\beta-\\xi) = (\\phi_t/t^\\beta)t^{\\beta/2}(X_t-\\zeta) + \\zeta t^{\\beta/2}(\\phi_t/t^\\beta-b). The second term is uncontrolled by (A.2). Take R=1, g_t(1)=1-t^{-\\varepsilon}, g_t(0)=t^{-\\varepsilon} with 0<\\varepsilon<1/2; then \\beta=1, (A.1) and (A.2) hold, but \\phi_t/t-b\\sim -c t^{-\\varepsilon}. The second term is then of order t^{1/2-\\varepsilon}\\to\\infty almost surely, while the martingale term X_t-\\zeta is o(t^{-1/2+\\varepsilon}) by the martingale law of the iterated logarithm. Since \\zeta>0 almost surely, the claimed convergence fails. The assumptions must be strengthened in a way that guarantees t^{\\beta/2}|\\phi_t/t^\\beta-b|\\to 0 (for instance, \\sum_t |g_t-g_\\infty|<\\infty), and the proof must explicitly handle this deterministic normalization term.","section":"Section 1.2 and proof of Theorem 2, after Eq. (24)"},{"comment":"Assumption (A.2) does not imply pointwise convergence g_t\\to g_\\infty; it only yields Ces\\`aro convergence in the form (1/N)\\sum_{t\\le N}|g_t-g_\\infty|\\to 0. Lemma 15, however, requires convergence of \\hat P_{m_n,t_n}(r) along arbitrary increasing sequences t_n, and Corollary 3 applies it along the random sequence \\sigma_k. Under (A.2), g_t may oscillate on a sparse set of times, and when \\beta=1 each such time has conditional probability bounded below of being an increment time for a given block; hence infinitely many sparse bad times can appear among the \\sigma_k, breaking the asserted limit. The paper should either add pointwise convergence (or a stronger rate condition) to the standing assumptions or replace the embedded-chain convergence arguments by Ces\\`aro/averaged versions.","section":"Section 3.3, Lemma 15 and Corollary 3"}],"minor_comments":[{"comment":"The statement of the Azuma-Hoeffding inequality says \"a sequence of negative real numbers (a_n)\"; it should say positive real numbers.","section":"Appendix A.1, Theorem 5"},{"comment":"In the final display of the proof, the expression \\xi(n) should be \\xi(I); the argument otherwise compares the CLT for the empirical maximizer with the limit block I.","section":"Section 5, proof of Theorem 4 (CLT part)"},{"comment":"Inside the sums over s, the terms use X_t instead of X_s; this appears to be a typographical error, but it makes the displayed asymptotics harder to follow.","section":"Section 3.2, equations (28)-(30)"},{"comment":"The inequality B(n,p,n/2)<1/2 is not strict for n=2 and p=1/2, where the probability equals 1/2; changing the conclusion to \\le 1/2 is sufficient for the argument.","section":"Section 4.1, Lemma 19"},{"comment":"The citations \"favaro2015,\" \"favaro2018,\" and \"OPR20\" appear as unresolved placeholders and should be completed.","section":"Section 1.3"},{"comment":"The proof of Lemma 21 is omitted; a one-sentence justification using (52) and the continuous mapping theorem would make the section self-contained.","section":"Section 5, Lemma 21"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is that Theorem 2 is false under the stated assumptions (A.1)-(A.2), not merely unproved; I would ask the authors to state the strengthened rate condition prominently and then re-check every place where pointwise convergence of g_t or a rate for \\phi_t/t^\\beta is used. The paper is likely salvageable, but it needs a substantive revision before consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is a real contribution, not a repackaging. The multiple-entry process with R arrivals per step and time-dependent g_t is new, and the results—t^beta growth for every block, mixed-normal CLT, persistent leadership, max CLT—go beyond single-entry Chinese restaurant and urn results. The phase transition at beta=1 is clean, and the leadership theorem is strong. The paper also earns credit for framing the model so it translates to balls-and-bins, GCRP, urns with immigration, and preferential attachment graphs; that is useful even if each translation is mostly terminological.\n\nThe main theorems are supported by a coherent martingale strategy: Freedman for exponential tails, Mori's martingale for moment bounds, a generating-function argument for positivity, and the Hall-Heyde CLT for fluctuations. I checked the key steps, and apart from one gap the derivations are detailed and mostly correct.\n\nThe soft spot is Lemma 6. The proof bounds the j≥2 terms using P(Δd_t ≥ j | F_t) ≤ 2R (d_t/(Rt))^j, which is false—for R=10, j=2, p=0.1 the true binomial tail is about 0.264, larger than 2R p^2 = 0.2. The companion bound C(r,j)≤2R is also not true as written. Lemma 6 is load-bearing: it feeds Lemma 8, which gives the uniform moment bounds, which give the L^p part of Theorem 1 and the L^p convergence in Theorem 4. So this is a genuine gap in the written proof.\n\nThat said, the gap looks patchable. Replacing the false bound with E[C(Δd,j)|F_t] ≤ C(R,j)(d/(Rt))^j and using C(d+m-1, m-j)/C(d+m-1, m) ≤ m^j/d^j gives a j≥2 contribution of order O_{m,R}(t^{-j}) ≤ O_{m,R}(t^{-2}), so the Mori supermartingale holds with a larger t^{-2} constant in φ_t^{(m)}. I would want the authors to fix that and re-check the constants, but I do not see a reason the main theorems fail. The omitted proof of Lemma 21 is minor; it is a routine consequence of (52) and tightness of the μ_n^*.\n\nAssumption (A.2) is strong but reasonable; without it the clean β scaling could break down, and the authors use it where they say they do. No circularity, no fitted parameters. The paper is honest about its limitations and positions the model correctly relative to the literature.\n\nBottom line: this deserves a serious referee. I would condition acceptance on a corrected Lemma 6 proof and a cleanly stated replacement for the binomial tail bound, but I expect the results to survive. Worth bringing to reading group.","headline":"A genuinely new meta-model for reinforced partitions with strong results; one repairable gap in Lemma 6's proof but the architecture holds.","tokens_in":30053,"tokens_out":1617,"would_cite":true,"duration_ms":15263,"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":"A broad class of reinforcement processes is governed by one number β: every block grows like t^β times a positive random variable, and one block permanently takes the lead.","keywords":["reinforcement processes","random partitions","generalized Chinese restaurant process","preferential attachment","balls and bins","central limit theorem","persistent leadership","Pólya urn with immigration"],"falsifier":"With R=2, set g_t(2)=1−1/t, g_t(1)=1/t, g_t(0)=0, so β=1 and assumptions (A.1)–(A.2) hold; simulate many trajectories and for a fixed block estimate the empirical asymptotic density of times at which it gains 0, 1, or 2 elements. If these densities do not converge to the binomial probabilities B(2, ξ^(1)/2, r), or if $t^{{1/2}}$(d_t^(1)/t−ξ^(1)) does not converge to a mixed normal with variance ξ^(1)(1−ξ^(1)/2), then the central CLT mechanism is wrong.","tokens_in":29032,"feed_emoji":"📈","tokens_out":10397,"duration_ms":100118,"temperature":0.7,"pith_summary":"This paper studies a general class of reinforced random partition processes in which R new items arrive at each time t; a random number of them join existing blocks with probability proportional to block size, and the rest form new blocks. The central claim is that a single parameter β — the limiting expected proportion of arrivals that join old blocks — determines the whole asymptotic picture. For every fixed block i, its cardinality at time t behaves like t^β times a strictly positive random variable ξ^(i), and the fluctuations around this limit are mixed normal after scaling by $t^{{β/2}}$. Moreover, almost surely one block eventually has strictly larger cardinality than every other block, and its lead grows without bound. This unifies predictions for balls-and-bins, generalized Chinese restaurant, Pólya-urn-with-immigration, and preferential-attachment graph interpretations.","feed_headline":"One block permanently outgrows all others","feed_subtitle":"In a broad reinforcement model, every block grows like t^β and one leader's gap goes to infinity.","key_machinery":"The load-bearing object is the normalizing sequence φ_t=∏_{s=2}^t(1+β_s/(s−1)), where β_s is the expected proportion of time-s arrivals joining existing blocks; Lemma 3 shows φ_t/t^β converges to a positive constant b under Assumption (A.2). The normalized block sizes X_t^(i)=d_t^(i)/φ_t form a submartingale, and become a martingale under the conditional measure given the block's birth time; auxiliary supermartingales built from binomial coefficients control all moments and prove the limits ξ^(i) are strictly positive. The central limit theorem is obtained from a martingale CLT by controlling the total squared increments, with the case β=1 requiring asymptotic-density results for the times a block receives exactly r new elements. Persistent leadership is proved by studying the difference in size between two blocks only at times when either one grows, showing this difference process has the drift lower bound E[ΔZ_k | G_k] ≥ (3/4)Z_k/k eventually and is transient, so Z_k → ∞.","core_discovery":"The paper proves four theorems. Theorem 1 says that for each block i, d_t^(i)/t^β converges almost surely and in L^p to a strictly positive finite random variable ξ^(i). Theorem 2 gives the finer fluctuation statement $t^{{β/2}}$(d_t^(i)/t^β − ξ^(i)) → W Z^(i) in distribution, where W is a standard Gaussian independent of Z^(i) and (Z^(i))^2 is ξ^(i) when β<1 and ξ^(i)(1−ξ^(i)/R) when β=1. Theorem 3 establishes persistent leadership: almost surely some block I satisfies d_t^(I) − max_{i≠I} d_t^(i) → ∞. Theorem 4 transfers the convergence and CLT to the maximum block size, whose limit is ξ^(I)=sup_i ξ^(i). A consequence is a phase transition at β=1: when β<1 a fixed block's share of all items vanishes, while when β=1 that share converges to a strictly positive random limit.","pith_inferences":["The paper's results suggest a universality class: any time-inhomogeneous reinforcement process whose drift converges to the same β should share the t^β growth and t^{β/2} fluctuation scaling, even though the law of ξ^(i) may depend on finer details of the arrival distribution.","If Assumption (A.2) is weakened to mere convergence g_t→g_∞, one would expect logarithmic corrections to φ_t and possibly non-Gaussian fluctuations; constructing such an example would delimit the boundary of the phase diagram.","The cardinality distribution of all block sizes is not proved here, but the paper sketches a power-law exponent 1+1/β for β<1; checking this by simulation would be a direct extension of the same machinery.","The transfer argument used for the maximum — once leadership holds, the maximum is just a single block — could also yield fluctuation results for the second-largest block or for the gap between leader and runner-up, though the paper does not pursue this."],"forward_implications":["Every fixed block, bin, table, or vertex grows to infinity at the same power-law rate t^β, so no single component is starved or monopolized.","There is a sharp phase transition at β=1: fixed-block shares of all items tend to zero for β<1 but to positive random limits for β=1.","In the graph interpretation, the maximum degree is of order t^β, linear when β=1 and sublinear when β<1, and from some random time onward there is a unique vertex of maximum degree whose lead increases without bound.","The same central-limit scaling t^{β/2} applies to the maximum as to a fixed block, with the variance of the mixed normal involving ξ^(I) and the correction factor 1−ξ^(I)/R only in the β=1 regime.","The leadership result holds without knowing the distribution of the ξ^(i)'s, so it applies uniformly across all sequences g_t satisfying the assumptions."],"supporting_citations":[{"why":"Provides the higher-moment martingale argument used to prove L^p convergence and positivity of the limiting block sizes.","marker":"[13]"},{"why":"Supplies the martingale central limit theorem (Corollary 3.5) that yields the Gaussian fluctuation results.","marker":"[9]"},{"why":"Provides the Lyapunov-function and transience argument for a drift process, used to prove persistent leadership.","marker":"[12]"},{"why":"Supplies the generating-function estimate for exponential moments that proves the limit ξ^(i) is almost surely positive.","marker":"[19]"},{"why":"Gives Freedman's martingale tail inequality used for the exponential concentration of normalized block sizes.","marker":"[8]"},{"why":"Gives the Azuma–Hoeffding inequality used to prove asymptotic-density statements about increments and drifts.","marker":"[3]"}],"fun_headline_variants":["One block forever outgrows all others","Reinforcement: a single block takes permanent lead","Almost surely, one block leads with infinite gap","Asymptotic dominance: one block beats all","Reinforcement process: leader gap goes to infinity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing regularity condition is (A.2): the arrival distributions g_t must approach g_∞ quickly enough that the weighted sum ∑_{t≥2} $t^{{-1}}$∑_r |g_t(r)-g_∞(r)| is finite, because this is what makes the normalizer φ_t behave like a clean power t^β and what keeps the martingale drifts under control.","fun_headline_variants_meta":{"raw":{"variants":["One block forever outgrows all others","Reinforcement: a single block takes permanent lead","Almost surely, one block leads with infinite gap","Asymptotic dominance: one block beats all","Reinforcement process: leader gap goes to infinity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000302,"raw_usage":{"total_tokens":1745,"prompt_tokens":956,"completion_tokens":789,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":717}},"tokens_in":572,"tokens_out":789,"duration_ms":8545,"temperature":1.0,"reasoning_tokens":717,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:48:55.041298+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"With R=2, set g_t(2)=1−1/t, g_t(1)=1/t, g_t(0)=0, so β=1 and assumptions (A.1)–(A.2) hold; simulate many trajectories and for a fixed block estimate the empirical asymptotic density of times at which it gains 0, 1, or 2 elements. If these densities do not converge to the binomial probabilities B(2, ξ^(1)/2, r), or if $t^{{1/2}}$(d_t^(1)/t−ξ^(1)) does not converge to a mixed normal with variance ξ^(1)(1−ξ^(1)/2), then the central CLT mechanism is wrong.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the higher-moment martingale argument used to prove L^p convergence and positivity of the limiting block sizes."},{"cited_title":"Martingale limit theory and its application","cited_arxiv_id":null,"evidence_quote":"Supplies the martingale central limit theorem (Corollary 3.5) that yields the Gaussian fluctuation results."},{"cited_title":"Urn-related random walk with drift ρxα/tβ","cited_arxiv_id":null,"evidence_quote":"Provides the Lyapunov-function and transience argument for a drift process, used to prove persistent leadership."},{"cited_title":"Time-dependent P\\'olya urn","cited_arxiv_id":"1807.04844","evidence_quote":"Supplies the generating-function estimate for exponential moments that proves the limit ξ^(i) is almost surely positive."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives Freedman's martingale tail inequality used for the exponential concentration of normalized block sizes."},{"cited_title":"Chung and L","cited_arxiv_id":null,"evidence_quote":"Gives the Azuma–Hoeffding inequality used to prove asymptotic-density statements about increments and drifts."}],"review_version":1}