{"id":"8cebb186-c424-4989-98e4-467ed1f910f6","arxiv_id":"2607.22324","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Power-α reinforced voting reaches consensus in (1+o(1))·2 log log n / log α rounds for α>1, at least exp(Ω(n)) rounds for α<1, with the winner fixed at round ⌈log 2 / log α⌉.","lead":"A new mathematics paper proves that in a voting model where each candidate's odds grow as the α-th power of their current votes, agreement arrives in about log-log n rounds whenever α>1 — but takes exponentially many rounds when α<1. The result also pins down exactly which round the winner becomes predictable, with sharp thresholds at α=2^{1/k}.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 5.6's lower-bound step is invalid as written: Eq. (5.54) contradicts Lemma 3.8's S, so Lemma 3.8(3) cannot prove (5.28) for the required j.","rationale":"The reader identified Lemma 3.8 as the least verifiable link. My check finds a more concrete, internal algebraic inconsistency: the S used in Lemma 3.8(3) is too large by a factor (log n)^{1+o(1)} for the j range in (5.56). If (3.7) is a typo and the intended S has 1/√log m, the proof may be repairable; if (5.54) is the intended scale, Lemma 3.8 needs a different version. Either way, the current manuscript does not rigorously establish the 1<α<2 branch of Theorems 1.3 and 1.6. I keep the verdict CONDITIONAL rather than REJECT because the flaw may be a fixable typo rather than a false theorem; the proposed test settles which. The α<1 proof and α=1 reduction appear sound, and the α≥2 case is developed separately, so the concern is targeted.","tokens_in":61500,"tokens_out":26575,"duration_ms":205558,"concrete_test":"Take α=1.1, t=1, ℓ=N^-_1. Using the definitions in (5.20)–(5.22), compute log H_1(ℓ) and log|A_1(ℓ)| from the bounds of E_{2,1}, E_{4,1}/Lemma 5.1; evaluate S from (3.7) asymptotically and compare to ½ n^-_2. If S/(½ n^-_2)→∞, the hypothesis of Lemma 3.8(3) fails. Separately, re-derive (5.54) from (3.7) symbolically; if the first inequality is not an identity, one of the two equations is wrong and determines whether the proof can be repaired.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Lemma 3.8(3) requires v≥4S(λ,m), with S(λ,m)=√λ exp(√log λ) √log m (Eq. 3.7). In Lemma 5.6's proof of (5.28), the lemma is applied with λ=H_t(ℓ), m=|A_t(ℓ)|, v=j, where j≥½ n^-_{t+1}=(log n)^{α^t−1+o(1)}. Under E_{2,t}∩E_{4,t}, for ℓ≈N^-_t, log H_t(ℓ)=Θ(α^t log log n) and log m ≥ (log n)^{2−α^t}(log log n)^{−O(1)}, so S(H_t(ℓ),m) ≥ (log n)^{1−o(1)}. Since α^t<2 and may be arbitrarily close to 1, j ≪ S. The displayed inequality (5.54), claiming 2S ≤ e^{√log H}√(H/log m), is algebraically incompatible with (3.7): it corresponds to S ∝ 1/√log m, not √log m. Thus the condition 4S<j fails, and the quoted lower bound for |A'_{t+1}(j)| in (5.56) is unsupported. Because (5.28) is the inductive engine for Proposition 5.5, the sharp constant 2/log α and the phase thresholds for 1<α<2 are not established as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a reinforced voting process on n candidates in which each voter in each round chooses a candidate with probability proportional to the α-th power of that candidate's previous vote count. The main claims are: for α>1 the absorption time satisfies T=(1+o(1))·2 log log n / log α with high probability, with a conditional variant for α>2; for 2^{1/k}<α<2^{1/(k-1)} the round-k leader is the eventual winner with probability 1−o(1), while round-(k−1) leaders are not; for α<1 the absorption time is at least exp(Ω(n)) with high probability; and the empirical vote distributions in constant rounds are close to an explicitly defined sequence of Poisson mixtures. The proof is built around a Poisson-coupling framework, level-set estimates for maxima of i.i.d. Poisson variables, a multi-step induction in Lemma 5.6, and a final 'heuristic zone' analysis in Proposition 4.4.","tokens_in":61806,"tokens_out":11510,"duration_ms":88160,"significance":"If the main theorems hold, the paper gives a sharp and surprising phase transition at α=1, with the same exponent α controlling both the consensus time scale and the round at which the winner is predetermined. The paper is also commendably non-circular: the thresholds λ_k=2^{1/k}, the constants 2/log α and (1+α/2)/log α, and the recursion (μ_k,z_k) are derived rather than fitted, and the simulations in §1.2 are consistent with, not calibrating, the theory. The α<1 proof in §6 is short and convincing, and the α=1 identification with Wright-Fisher/Kingman coalescent is standard. However, the entire 1<α<2 branch, including Theorems 1.3 and 1.6 for all k≥2, rests on Lemma 5.6 and the level-set estimates of Lemma 3.8; one displayed inequality in the proof of Lemma 5.6 is algebraically inconsistent with the definition of S(λ,m), and the affected lower-bound estimate is load-bearing. The sharp results for 1<α<2 are therefore not established as written.","major_comments":[{"comment":"The proof of the lower bound in (5.49) applies Lemma 3.8(3) to the i.i.d. Pois(H_t(λ)) variables on A_t(λ), requiring 4S(H_t(λ),|A_t(λ)|) ≤ j. The displayed inequality (5.54) claims 2S(H,|A|) ≤ exp(√log H)·√(H/log|A|). But by the definition (3.7), S(H,m)=√H exp(√log H)√log m, so this inequality is equivalent to 2√H exp(√log H)√log m ≤ exp(√log H)√H/√log m, i.e. 2 log m ≤ 1, which is false in the regime where |A_t(λ)| is large. In the intended application, log|A_t(λ)| ≈ (log n)^{2−α^t+o(1)} and √H_t(λ) ≈ (log n)^{α^t/2+o(1)}, so S ≈ (log n)^{1+o(1)}, while j ≈ (log n)^{α^t−1+o(1)} with α^t−1<1 for α<2. Thus the required condition 4S≤j fails badly, and Lemma 3.8(3) cannot be used to obtain (5.56). Since (5.28) is the inductive engine of Lemma 5.6 and Proposition 5.5, the proofs of Theorems 1.3 and 1.6 for 1<α<2 are incomplete as written.","section":"§5.2.4, Eq. (5.54) and Lemma 3.8(3)"},{"comment":"The transfer from the Poisson levels A'_{t+1} to the multinomial levels A_{t+1} is asserted via (5.31)–(5.32) and the relation A_{t+1}(λ)⊂∪_{−8≤u≤8} A'_{t+1}(λ+u). This relation is plausible but is not proved in detail, and the constants 8, 1/2 are not derived from the coupling. More importantly, the lower-bound part of the transfer uses the invalid inequality (5.54); without a valid lower bound on |A'_{t+1}(λ,j)|, the lower bound in the definition of E'_{4,t+1} is unsupported. This is the same load-bearing failure as in the previous comment, but I flag it separately because it is the exact step that feeds the induction for all t≤m_0.","section":"§5.2.4, around (5.52)–(5.56)"}],"minor_comments":[{"comment":"The title contains a spacing typo: 'CONCLA VE' should be 'CONCLAVE'.","section":"Title"},{"comment":"The assertion that the two-thirds-stopping-rule variant has absorption time (1+o(1))E[T] for α>1 is stated without proof. If this claim is not needed for the main theorems, it should be labeled as a conjecture or moved to an 'open problems' remark; as written it appears to be an unproved additional result.","section":"Remark 1.2"},{"comment":"Even if the inequality in (5.54) were reversed, it would not help the proof, because the required condition for Lemma 3.8(3) is a lower bound on j, not an upper bound. The authors should re-derive the condition carefully and either correct the display or replace the argument.","section":"Eq. (5.54)"},{"comment":"The superscript notation n^-_t, n^+_t, N^-_t, N^+_t is easy to confuse with exponents and with negative numbers. A different notation, such as n_t^- and n_t^+, or n_{\\mathrm{low},t}, n_{\\mathrm{high},t}, would improve readability.","section":"Notation, §5.2"}],"recommendation":"major_revision","confidential_remarks":"The identified error in §5.2.4 is not a cosmetic typo: it invalidates the lower-bound estimate that drives Lemma 5.6, and hence the whole 1<α<2 branch of Theorems 1.3 and 1.6. I could not verify the rest of that long chain line-by-line, but the contradiction between (5.54) and (3.7) is unambiguous. If the authors can replace (5.54) with a valid argument or adjust the induction, the paper may be salvageable; otherwise the advertised sharp results for 1<α<2 should be withdrawn or substantially weakened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely interesting process and the paper has the right shape — a sharp phase diagram and a winner-determination ladder — but the proof of the 1<α<2 branch has a load-bearing error that I don't see how to patch from the text. If you pick this up expecting the full theorem, you will find the core lemma (5.6) applies Lemma 3.8(3) in a regime where its hypotheses fail.\n\nWhat's new and good: Definition 1.1 (power-α multinomial reinforcement) is natural and, as far as the citations go, new. The α=1 reduction to Wright-Fisher/Kingman is standard and correctly credited. The α<1 exponential lower bound (Theorem 1.5) is clean and easily checked — the p_i ≤ 1/(1+3^{-α}) < 3/4 bound does the job. For α≥2, the two-step analysis via occupancy and Poisson coupling is plausible and has no obvious gaps. The absence of fitted constants or self-citations is a real plus; the thresholds 2^{1/k} and constants 2/log α, (1+α/2)/log α are derived, not tuned. The simulations are honest.\n\nNow the soft spot, and it's a big one. In Lemma 5.6, proof of (5.28), Case α^t<2, the authors need Lemma 3.8(3) for v=j. That lemma requires v ≥ 4S(λ,m) with S = √λ exp(√log λ) √log m. In the intended application, λ=O((log n)^{α^t+o(1)}), log m = (log n)^{2-α^t-o(1)}, so S = (log n)^{1-o(1)}. But j = (log n)^{α^t-1+o(1)}. Since α^t<2, α^t-1<1, so j ≪ S for any α^t bounded away from 2? Actually even if α^t close to 2, j ~ (log n)^{1-o(1)} as well; compare to S ~ (log n)^{1-o(1)} times exp(sqrt(log H))? Actually S ~ (log n)^{1-o(1)} so j could be comparable if α^t close to 2, but they claim 4S < j in (5.54). The algebra (5.54) is simply wrong: the right side should have √log m in the numerator, not the denominator. So the hypothesis fails.\n\nMaybe this can be fixed: perhaps the authors meant to use a different S or a different level-set lemma. But as written, the lower bound for |A'_{t+1}(j)| and hence the induction (5.28) is unsupported. Since Proposition 5.5 and the sharp constant 2/log α for 1<α<2 rest on this step, the main theorem is not established for that range. That's not a minor gap.\n\nIf the authors can correct (5.54) or find an alternative bound, the paper would be significant. As it stands, my advice: send to a competent referee for α≥2 and α<1 parts, but tell the referee to focus on Lemma 5.6 and not let the polished intro hide the error. The α=1 literature summary is worth citing.","headline":"A genuinely interesting new process, but the 1<α<2 regime rests on a false inequality in Lemma 5.6 that I don't see how to patch from the text.","tokens_in":62449,"tokens_out":5117,"would_cite":false,"duration_ms":38225,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J10","60K35"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that in a voting model where the probability of voting for a candidate is proportional to the α-th power of that candidate's previous vote count, the absorption time is (1+o(1))·2 log log n / log α for every α>1, and that","keywords":["reinforced voting process","power-law reinforcement","absorption time","log-log consensus time","phase transition","Poisson approximation","winner selection","level-set method"],"falsifier":"Run the voting process with α=1.9 for n=10^4,10^5,10^6 and track whether the eventual winner was among the first-round leaders; Theorem 1.6 predicts this probability tends to 0, so any flattening above 0 refutes the phase transition.","tokens_in":61251,"feed_emoji":"🗳️","tokens_out":8644,"duration_ms":71622,"temperature":0.7,"pith_summary":"The paper studies a stochastic voting process in which n cardinals vote repeatedly and the probability that a candidate receives a vote is proportional to the α-th power of that candidate's current tally. It establishes a sharp transition at α=1: for any α>1, consensus is reached in about 2 log log n / log α rounds with probability tending to 1, whereas for α<1 the process typically needs at least exp(Ω(n)) rounds. It also identifies exactly when the eventual winner becomes predetermined: if 2^{1/k}<α<2^{1/(k-1)}, the unique leader after round k is the eventual winner with probability 1−o(1), while the round-(k−1) leaders are not. A sympathetic reader cares because the model captures a general reinforcement mechanism — success begets success — and the results show that such reinforcement produces ultra-fast consensus and a sharp predictability transition.","feed_headline":"Reinforced voting reaches consensus in log-log n rounds","feed_subtitle":"Above α=1, a handful of rounds settles even huge electorates; below α=1, consensus requires exponential time.","key_machinery":"The load-bearing object is the pair (maximum vote count M_t, relative gap δ_t between maximum and second maximum) plus, for 1<α<2, the level-set counts |A_t(j)| = #{i : X_i^{(t)} = M_t − j}. The core mechanism is the deterministic-looking recursion 1−δ_{t+1}≈(1−δ_t)^α once the leader has a large enough lead, together with a level-set lemma that controls how many candidates sit at each distance below the maximum. These estimates feed a recursion for the number of candidates at each level set; the balance between enumeration gain and expected-value loss determines the critical round k and the initial gap δ_0=(log n)^{-1+o(1)}, which via Proposition 4.4 yields the constant 2/log α.","core_discovery":"The central claim is a complete asymptotic classification of the absorption time and winner identity in the power-law reinforced voting process. For α>1 the process reaches full consensus in (1+o(1))·2 log log n / log α steps; the proof decomposes the dynamics into a pre-heuristic phase in which leaders at successive rounds migrate through level sets of the vote distribution, and a heuristic zone in which the maximum and second maximum evolve deterministically via the map 1−δ'≈(1−δ)^α. For 1<α<2, the level-set analysis shows the winner is selected at round k exactly when α lies between 2^{1/k} and 2^{1/(k-1)}, with all earlier rounds' leaders eliminated; for α>2 the winner is fixed already a","pith_inferences":["An immediate testable extension is to replace x^α by a general increasing reinforcement function μ and ask whether the thresholds become the points where an iterated map μ^{∘k}(m)/m crosses 2; the paper's techniques suggest such a generalization.","The log-log law implies that the same model could serve as a stylized explanation for fast consensus in winner-take-all elections or in ranking algorithms, since a few rounds of proportional-to-power updates concentrate an enormous field into a single leader.","The paper's conjecture that at the exact critical α=2^{1/k} the winner is decided at round k+1 rather than k could be tested numerically where the bounds are tight; a Monte Carlo estimate of P(W∈L_k) at α=2^{1/k} would settle it."],"forward_implications":["For any α>1, consensus time does not grow with n beyond log-log; even astronomically large electorates finish in a handful of rounds, matching the observation that real conclaves end quickly despite n≈100.","The threshold sequence λ_k=2^{1/k}: for α between λ_k and λ_{k-1}, the event of the eventual winner is determined in round k, so the later a leader emerges, the smaller α is; round k−1 gives no information about the winner.","At α>2, the conditional constants differ: if the first-round leader is unique, T≈2 log log n/log α; if not, T≈(1+α/2) log log n/log α.","For α<1, the process is essentially frozen at exponentially long time scales; any majority is exponentially unlikely, so absorption time ≥ exp(Ω(n)).","The empirical vote distribution after k rounds is close to a universal Poisson-mixture law μ_k, with partition function close to n z_k, so early-round dynamics are asymptotically distributionally tractable."],"fun_headline_variants":["Phase transition in voting: α>1 gives log-log n consensus time","Reinforced voting: sharp transition at α=1, consensus time drops to log-log n","Winner chosen in rounds for α>1; α<1 needs exponential time","Consensus in double-log rounds: the shift at α=1","α>1: voting reaches unanimity in O(log log n) rounds"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The exact thresholds 2^{1/k} and the constant 2/log α rest on the level-set estimates of Lemma 3.8 being accurate to a (log n)^{o(1)} factor; if those Poisson count estimates fail by more than that, the phase boundaries and log-log constant shift.","fun_headline_variants_meta":{"raw":{"variants":["Phase transition in voting: α>1 gives log-log n consensus time","Reinforced voting: sharp transition at α=1, consensus time drops to log-log n","Winner chosen in rounds for α>1; α<1 needs exponential time","Consensus in double-log rounds: the shift at α=1","α>1: voting reaches unanimity in O(log log n) rounds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000963,"raw_usage":{"total_tokens":3963,"prompt_tokens":797,"completion_tokens":3166,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":3064}},"tokens_in":541,"tokens_out":3166,"duration_ms":17335,"temperature":1.0,"reasoning_tokens":3064,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T05:08:13.294907+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the voting process with α=1.9 for n=10^4,10^5,10^6 and track whether the eventual winner was among the first-round leaders; Theorem 1.6 predicts this probability tends to 0, so any flattening above 0 refutes the phase transition.","supporting_citations":[],"review_version":1}