{"id":"597c69de-f6d7-45ee-94b6-7dc58138bd4f","arxiv_id":"2506.12816","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Mean-field exchange models with any symmetric, non-degenerate redistribution variable exhibit a Wasserstein cutoff at time n log n/(2h) with Gaussian profile set by the size-biased log-variable's first two moments.","lead":"This paper proves that three families of random energy-exchange models all converge to their limiting state through a sharp cutoff at the same explicit time, with the same Gaussian-shaped transition profile. The result covers any symmetric redistribution rule and includes the repeated-averaging process as a special case.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: finite-moment worry is spurious; remaining defects are minor and non-load-bearing.","rationale":"The reader's weakest assumption is incorrect: h and s^2 are automatically finite for any symmetric non-degenerate X on [0,1], because x|log x|^p is bounded on [0,1]. The reader's sign-error observation is correct but ultimately cosmetic, since the gamma term is sent to zero before reaching the cutoff profile. The upper-bound time-shift is a presentation gap rather than a substantive flaw: it is closed either by a diagonal argument over beta or by applying the proof at the earlier time t-r, using the fact that Lemma 4.1 is uniform in t. The central universal-cutoff claim, its explicit entropic constants, and the two-sided estimates from the pile-dynamics identity all appear sound. I therefore see no reason to change the reader's conditional verdict, though a revised manuscript should fix the sign in Corollary 2.2 and state the separate treatment of the s=0 case.","tokens_in":18381,"tokens_out":43987,"duration_ms":415737,"concrete_test":"Re-derive Corollary 2.2 by computing the integral over alpha of phi(alpha) Phi((-alpha - gamma - beta(1+r))/r) and confirm the displayed limit has -gamma in the numerator; then rerun Section 4.2 starting from time T = t - r and verify that K(t) <= 2E||eta^+_{t-r}|| + O(e^{-psi}) follows from the same inequalities and Corollary 2.2 at time T.","verdict_should_be":"UNCHANGED","load_bearing_attack":"As a second-pass reviewer, I found no load-bearing gap in the central claim. The natural finite-moment worry is spurious: for every X in [0,1], the size-biased log moments are automatically finite, since E[|-log X_hat|^p] = 2E[X|log X|^p] <= 2 max_{x in [0,1]} x|log x|^p < infinity. Thus the reader's example of a density near 0 does not make h infinite, and Theorem 1.5 needs no additional moment hypothesis. The genuine defects are minor and fixable: Corollary 2.2 displays the wrong sign in front of gamma (the proof's own integral gives -gamma), and the case s=0 (X identically 1/2) requires a separate binomial CLT because the written proof divides by r. The upper-bound argument also proves the estimate at time t+r rather than at t, but this is closed by a standard diagonal shift of beta, or by rerunning the same estimates from time t-r using Lemma 4.1's uniformity in t. None of these issues affect the gamma-to-0 profile in Theorem 1.5.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves a universal cutoff phenomenon for three families of mean-field exchange models—the Stochastic Redistribution Model (SRM), the Stochastic Equalization Model (SEM), and the Generalized Averaging Model (GAM)—driven by an arbitrary symmetric, non-degenerate redistribution variable X taking values in [0,1]. Starting from a Dirac initial configuration, the worst-case 1-Wasserstein distance to the stationary law converges, along the time t = t_ent + β t_w, to the Gaussian profile 2Φ(−β(1+r)/√(1+r^2)), where t_ent = n log n/(2h), t_w = (1+r)n/2 √(log n/h), h = E[−log X_hat], s^2 = Var(log X_hat), r = s/h, and X_hat is the size-biased version of X. The same statement is shown for the permutation-invariant distance W̄1, and for the distance from a Dirac initial condition. The proof introduces a pile dynamics representation, establishes the exact identity E[∥η^θ_t∥_1] = P(∏_{i=1}^T X_hat_i ≥ θ) with T ~ Bin(t, 2/n), and then combines a central limit theorem (Corollary 2.2) with W2-contraction estimates and truncation arguments to obtain matching lower and upper bounds.","tokens_in":18526,"tokens_out":14241,"duration_ms":127880,"significance":"If correct, this is a significant unification: it extends the cutoff result for the repeated-averaging process of Chatterjee–Diaconis–Sly–Zhang to arbitrary symmetric redistribution laws, including the KMP/flat Kac model, and it gives the first cutoff statements for the non-reversible SEM and GAM. The constants h and s^2 are explicit, parameter-free functionals of the model input X, so the theorem has strong predictive content and is not a fit of adjustable parameters. Methodologically, the paper is self-contained and relies on probabilistic identities and couplings rather than spectral analysis. The result also yields improved lower bounds for total-variation mixing times as a byproduct. The proofs are presented in enough detail to be checked line by line.","major_comments":[{"comment":"The proof of Corollary 2.2 divides by r = s/h in the standardization step (the expression (log n − ψ − mh)/(s√m) is written as (1/r)(−α−γ−β(1+r))). This is invalid when s = 0, i.e., when X is almost surely 1/2. The theorem’s assumptions include this case (non-degeneracy only excludes Bernoulli {0,1} laws), and the paper explicitly claims to recover the averaging process of [CDSZ22] as a special case. The authors should either handle X ≡ 1/2 by a separate binomial CLT for T (since log X_hat is then constant) or add a continuity argument in distribution. As written, the proof does not cover all cases of Theorem 1.5.","section":"§3.2, Corollary 2.2"}],"minor_comments":[{"comment":"The statement of Corollary 2.2 displays the limit Φ((−β(1+r)+γ)/√(1+r^2)), but the proof’s final integral gives Φ((−β(1+r)−γ)/√(1+r^2)). Since the corollary is used only in the limit γ→0 in the proof of Theorem 1.5, the main result is unaffected, but the sign should be corrected for consistency with the proof.","section":"§3.2, Corollary 2.2"},{"comment":"No finite-moment hypothesis is needed beyond the stated assumptions: because X ∈ [0,1], h = E[−2X log X] ≤ 2/e and E[(log X_hat)^2] = 2E[X(log X)^2] is finite, since x(log x)^2 is bounded on [0,1]. The authors may wish to add a short remark to this effect to preempt a natural concern.","section":"§1.2, Theorem 1.5"},{"comment":"The claim that n^{-1} e^{6ψ} = o(1) is correct for every fixed γ > 0 because e^{6ψ} = exp(O(√(log n))) = o(n). This is fine, but a one-line justification would help the reader avoid the mistaken impression that the argument requires γ to be small.","section":"§4.2, Eq. (4.6)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong and clearly written contribution from well-known authors. The main theorem is correct in substance; the gaps identified are local and fixable within the manuscript’s scope. The reader’s finite-moment concern is unfounded on the actual hypothesis X ∈ [0,1]. I recommend revision to address the s = 0 case in Corollary 2.2 and the sign error in its statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one. The paper proves a universal Wasserstein cutoff for three families of mean-field exchange models — SRM, SEM, GAM — for any symmetric non-degenerate redistribution variable X. The cutoff time is n log n/(2h), the window is order n sqrt(log n), and the profile is 2Φ(−β(1+r)/sqrt(1+r^2)), with h = E[−2X log X], s² = Var(log X̂), and r = s/h. That generalizes the CDSZ22 repeated-averaging cutoff to arbitrary X and to two non-reversible families. The constants are explicit functionals of the model input, not fitted parameters; the proof is a proof. That is the main thing to know.\n\nThe proof strategy is genuinely nice: they track pile dynamics, get an exact identity for the expected mass in piles above a threshold (Proposition 2.1), then use a CLT for the size-biased log and an L² contraction to close lower and upper bounds. The three models are unified through duality and the same pile representation. I did not find a load-bearing gap. The reader's worry about finite moments of h and s² is spurious: for X in [0,1], E[|−log X̂|^p] = 2E[X|log X|^p] is bounded by 2 max_{x∈[0,1]} x|log x|^p, so h and s² are automatically finite under the stated assumptions.\n\nSoft spots, both minor. First, Corollary 2.2 displays the limit Φ((−β(1+r)+γ)/sqrt(1+r²)), but the proof's own integral gives Φ((−β(1+r)−γ)/sqrt(1+r²)). Sign error in the display. It does not affect Theorem 1.5 because γ is later sent to 0, but it should be fixed. Second, the proof of Corollary 2.2 divides by r = s/h. The case X ≡ 1/2 has s = 0, so the CLT step as written does not apply, even though the theorem claims to include that case (and it is the known CDSZ22 case). This needs a separate argument (e.g., a binomial CLT) or an explicit assumption s > 0 with the degenerate case handled by continuity or reference. Not a big deal, but it is a real gap in the written proof. Also, there is a notational clash: r is used both for s/h and for the number of extra steps in the upper bound. The upper-bound step proves the estimate at t+r rather than at t, but the standard diagonal shift in β closes that.\n\nWho is this for: people working on mixing times, cutoff, mean-field interacting particle systems, and Kac-type models. It deserves a serious referee; I would send it out. The fixes are small and the main result is solid.","headline":"A clean unification of cutoff results for three mean-field exchange families; the main theorem holds up, with two minor fixable statement-level blemishes.","tokens_in":19106,"tokens_out":4322,"would_cite":true,"duration_ms":39063,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","82B20","82C26"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that three mean-field exchange models—stochastic redistribution, stochastic equalization, and generalized averaging—all exhibit the same cutoff phenomenon in Wasserstein distance, with an explicit mixing time and Gaussian…","keywords":["mixing times","cutoff phenomenon","exchange models","Wasserstein distance","mean-field interaction","pile dynamics","size-biased distribution","entropic time"],"falsifier":"Run the SRM with $X\\sim\\mathrm{Beta}(1,1)$ (uniform redistribution) at large $n$, say $10^5$, starting from a point mass, and estimate $W_1$ at $t=t_{\\mathrm{ent}}+\\beta t_w$ for $\\beta=-1,0,1$; here $h=1/2$, $s=1/2$, $r=1$, so Theorem 1.5 predicts $W_1\\approx 2\\Phi(-\\beta\\sqrt2)$, about $1.84$, $1$, and $0.16$ respectively, and a sustained mismatch beyond Monte Carlo error would falsify the universal profile.","tokens_in":18125,"feed_emoji":"🔄","tokens_out":21633,"duration_ms":218464,"temperature":0.7,"pith_summary":"This paper proves that three families of many-particle energy-exchange processes—the stochastic redistribution model, the stochastic equalization model, and the generalized averaging model—all undergo the same sharp cutoff transition in 1-Wasserstein distance, for every symmetric, non-degenerate redistribution rule. Starting from a worst-case state with all energy on one particle, the distance to equilibrium stays near its maximum until time roughly $n\\log n/(2h)$, then drops to zero over a window of order $n\\sqrt{\\log n}$, following an explicit Gaussian curve. The only model-dependent inputs are $h$, the mean of $-\\log \\widehat X$ for the size-biased redistribution variable $\\widehat X$, and the variance $s^2$ of that log-variable. This matters because previous mixing results for such models typically required smooth initial data or used total variation, which is too sensitive here; the Wasserstein cutoff gives a robust and explicit description of how a disordered energy exchange reaches equilibrium.","feed_headline":"Three exchange models mix with one universal cutoff profile","feed_subtitle":"One entropic constant, the mean log of the size-biased rule, sets the mixing time and Gaussian window for all three dynamics.","key_machinery":"The proof is carried by the pile dynamics, a representation in which every particle's energy is decomposed into labeled fragments, or piles, and each interaction at a pair of sites replaces every pile of size $p$ by new piles whose sizes are $Xp$ and $(1-X)p$, with model-specific routing. The key identity, Proposition 2.1, equates the expected total energy in piles of size at least $\\theta$ with $\\mathbb P(\\prod_{i=1}^T \\widehat X_i\\ge\\theta)$, where the $\\widehat X_i$ are i.i.d. size-biased copies of $X$ and $T\\sim\\mathrm{Bin}(t,2/n)$. This identity converts the cutoff problem into a central limit theorem for sums of $\\log\\widehat X_i$, which produces the Gaussian profile; separate $L^2$ contraction bounds for the three models show that the mass carried by many tiny piles is negligible.","core_discovery":"At the center of the paper is Theorem 1.5: for any symmetric, non-degenerate $X\\in[0,1]$ and for all three models (SRM, SEM, GAM), the worst-case Wasserstein distance $W_1(t)$ and its permutation-invariant counterpart $\\overline W_1(t)$ converge as $n\\to\\infty$ to $2\\Phi\\bigl(-\\beta(1+r)/\\sqrt{1+r^2}\\bigr)$ when $t=t_{\\mathrm{ent}}+\\beta t_w$. The entropic time is $t_{\\mathrm{ent}}=n\\log n/(2h)$, the window is $t_w=(1+r)n/2\\,\\sqrt{\\log n/h}$, with $h=\\mathbb E[-2X\\log X]=\\mathbb E[-\\log\\widehat X]$, $s^2=\\mathrm{Var}(\\log\\widehat X)$, and $r=s/h$. The same limit holds for the labelled distance from a point-mass initial state and for the permutation-insensitive distance, so the cutoff is universal across the three dynamics and across all symmetric redistribution laws.","pith_inferences":["The same pile-threshold mechanism plausibly yields the same Gaussian profile for multi-particle equalization dynamics, where an interaction splits or averages several piles at once; the paper does not treat that case.","At the edges of the window, the Gaussian profile should give way to large-deviation tails controlled by the cumulant generating function of $\\log\\widehat X$; the paper does not explore this regime.","The explicit profile invites a finite-size expansion: the first correction should be governed by the third moment of $\\log\\widehat X$, which could be measured numerically.","The lower-bound technique could be pushed to prove total-variation cutoff for the SRM, which the paper explicitly leaves open."],"forward_implications":["For any symmetric non-degenerate $X$, all three processes mix from the worst initial state at time $n\\log n/(2h)$ with a window of order $n\\sqrt{\\log n}$; the profile depends on $X$ only through $h$ and $r=s/h$.","The permutation-invariant distance $\\overline W_1$ has the same cutoff as the labelled distance $W_1$, so the result is insensitive to particle labels.","When $X\\sim\\mathrm{Beta}(\\alpha,\\alpha)$, the constants are explicit digamma functions, so the mixing time and window are computable for every $\\alpha$; letting $\\alpha\\to\\infty$ recovers the averaging process $X\\equiv 1/2$.","For the SRM with $\\mathrm{Beta}(\\alpha,\\alpha)$ and $\\alpha\\in(0,1)$, the lower bound improves the coupon-collector total-variation bound from earlier work.","The duality between SRM and SEM transfers the cutoff statement to the SEM, whose stationary state is a random flat configuration rather than a deterministic one."],"supporting_citations":[{"why":"Analyzes the averaging-process special case $X\\equiv 1/2$, which the theorem generalises and recovers as a limit.","marker":"[CDSZ22]"},{"why":"Introduces the pile/splitting process that serves as the paper's main proof device.","marker":"[CQS24]"},{"why":"Establishes the order-$n\\log n$ total-variation mixing time for the SRM with Beta redistribution, providing the coupon-collector bound the paper improves.","marker":"[Smi14]"},{"why":"Supplies the 'entropic time' cutoff paradigm and the Gaussian-profile technology adapted here.","marker":"[BCS19]"},{"why":"Provides the averaging-process background and the $L^2$ contraction estimates used to control small piles.","marker":"[AL12]"},{"why":"Introduces the heat-flow exchange model that motivates the redistribution class.","marker":"[KMP82]"},{"why":"Discusses the duality between SRM and SEM used to identify the SEM's random flat stationary state.","marker":"[DMFG24]"}],"fun_headline_variants":["Universal cutoff: all symmetric exchange rules mix alike","One entropic constant fixes mixing time for all exchange models","Gaussian cutoff profile emerges universally in exchange dynamics","Exchange models share a single cutoff time and profile","Mean-field exchange: one mixing law, three dynamics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes the redistribution variable actually mixes the two energies with positive probability and that the logarithm of its size-biased version has finite variance; the stated hypotheses on $X\\in[0,1]$ guarantee both, but the theorem does not say so.","fun_headline_variants_meta":{"raw":{"variants":["Universal cutoff: all symmetric exchange rules mix alike","One entropic constant fixes mixing time for all exchange models","Gaussian cutoff profile emerges universally in exchange dynamics","Exchange models share a single cutoff time and profile","Mean-field exchange: one mixing law, three dynamics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000256,"raw_usage":{"total_tokens":1538,"prompt_tokens":868,"completion_tokens":670,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":596}},"tokens_in":484,"tokens_out":670,"duration_ms":7236,"temperature":1.0,"reasoning_tokens":596,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:10:00.029087+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the SRM with $X\\sim\\mathrm{Beta}(1,1)$ (uniform redistribution) at large $n$, say $10^5$, starting from a point mass, and estimate $W_1$ at $t=t_{\\mathrm{ent}}+\\beta t_w$ for $\\beta=-1,0,1$; here $h=1/2$, $s=1/2$, $r=1$, so Theorem 1.5 predicts $W_1\\approx 2\\Phi(-\\beta\\sqrt2)$, about $1.84$, $1$, and $0.16$ respectively, and a sustained mismatch beyond Monte Carlo error would falsify the universal profile.","supporting_citations":[],"review_version":1}