{"id":"e505468e-a9e3-4c5d-a7f4-b8fcac4ffbbf","arxiv_id":"2502.06071","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"The labeled and unlabeled versions of a capacity-constrained congestion dynamics both exhibit abrupt convergence (cutoff), at times (1/2)n log n and (1/2)(1-ρ)n log n respectively.","lead":"Players repeatedly switch resources under capacity limits in a simple random way; the paper proves the system converges abruptly at a sharply predictable time, and that an easier unlabeled version converges even earlier. The result gives exact benchmarks for how quickly congestion-type sampling algorithms mix, relevant to discrete log-concave sampling.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified; the proof is internally consistent for the stated range Q fixed, ρ in (0,1/2], and the only soft spot is the explicitly open case ρ > 1/2.","rationale":"The reader's weakest_assumption identified the ρ ≤ 1/2 constraint in Lemma 3.5, which is exactly the most fragile edge of the proof. However, this is a limitation of the theorem's stated range, not a flaw within that range. My independent pass over the main estimates—the load-profile contraction in Lemma 2.1, the row-wise φ_n bound in Lemma 3.4, the coupling drift and variance lower bounds in Lemmas 2.4 and 3.5, and the diagonal-statistic lower bound in Theorem 3.7—did not reveal an internal inconsistency or a missing argument. The constants grow with Q, but Q is fixed, and the O(n) window absorbs the α-dependence after α is chosen for a given ε. The no-cutoff appendix is a side remark rather than a load-bearing component of the central claim. Since no correctness concern lands, the reader's ACCEPT verdict should remain unchanged.","tokens_in":29625,"tokens_out":52081,"duration_ms":473659,"concrete_test":"Independently verify that the transition prescribed in Lemma 3.5 is a valid coupling: for one of the chains, fix a player i and a target resource w, sum the coupling probabilities over all matched vacancy indices j, and confirm the marginal probability equals (κ−ξ_w)/(1+N−n); if this fails for any case (i)–(iii), the labeled upper-bound proof has a hidden transition-probability error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I read the argument as a complete proof for the claimed regime. The most delicate step is the labeled coupling in Lemma 3.5, whose nonpositive drift uses #V+ ≤ #O+ together with N ≥ 2n, i.e. ρ ≤ 1/2; the paper explicitly leaves ρ > 1/2 open. Rechecking the algebra: in case (ii) the average drift is at most −#O+(N−2n)/(n(N−n)) ≤ 0, case (iii) never increases D, and case (i) provides symmetric ±1 moves with total probability at least 1/(4Q^3). Lemma 3.4's φ_n regularization and Markov bound yield the O(Q^2/R) failure probability needed before the coupling stage. The lower bound in Theorem 3.7 separates a single diagonal entry of the load matrix, with mean excess e^α/√n and variance O(1/n), giving D_TV(·) → 1 before t_n. No hidden circularity or missing step emerged. The remaining limitation is the unproved labeled upper bound for ρ > 1/2, but this is an acknowledged open question, not a defect in the stated theorem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a Glauber dynamics on configurations of n labeled players assigned to Q resources with capacity κ, where N = κQ, n is close to ρN, Q is fixed, and ρ is a fixed constant in (0, 1/2]. Theorem 1.1 establishes cutoff at time (1/2)n log n with an O(n) window for the labeled congestion dynamics, and Theorem 1.2 establishes cutoff at time (1/2)(1−ρ)n log n for the unlabeled load-profile chain, which is a log M-concave sampling chain. The proofs are coupling-based: drift computations drive the normalized load vector or load matrix into an equilibrium neighborhood in time of order N log N, followed by a variance-positive coalescing coupling that finishes in O(N) time; lower bounds use variance estimates to separate a coordinate from its stationary law. Appendix A shows that the analogous natural chain for uniform sampling on an M-convex set need not exhibit cutoff.","tokens_in":29901,"tokens_out":16075,"duration_ms":160919,"significance":"If the results are correct, the paper supplies exact asymptotic mixing-time benchmarks with explicit cutoff constants for a natural family of generalized exclusion processes, including the log M-concave sampling case of the unlabeled chain. The proof is essentially self-contained: Theorem 1.2 is proved directly rather than imported from Lacoin and Leblond, and Theorem 1.1 is a new labeled-process result. The matching upper and lower bounds are derived from the dynamics rather than fitted, the constants are explicit in the fixed-Q, ρ-in-(0,1/2] regime, and the main limitations (ρ ≤ 1/2 for the labeled process, Q fixed) are stated honestly in the paper. The appendix's negative example is a useful caveat for the broader M-convex sampling program. No code is supplied, but the proofs are detailed enough to be checkable line by line.","major_comments":[],"minor_comments":[{"comment":"In the displayed expectation bounds, the quantity 'ρ/N' should be 'ρ/Q'; the same typo appears in the stationary-distribution paragraph where the bound '|E_{π_n} ξ_v/N − ρ/N| ≤ C/N' should have 'ρ/Q'. The correct stationary value is the constant vector (ρ/Q)_{v∈S}.","section":"§2, proof of Theorem 2.6"},{"comment":"The sentence 'after that we consider two independent load matrix chains up to time t_n + T_{α,N}' is inconsistent with the total time t_n + T_{α,N} + ⌈αn⌉ in the final display; the argument and the surrounding equations show that the load-matrix chains should first be run up to time t_n and only then coupled for ⌈αn⌉ further steps.","section":"§3, proof of Theorem 3.6"},{"comment":"In the last line of the proof, 'first clam' should read 'first claim'.","section":"§3, Lemma 3.2"},{"comment":"The variance lower bounds used before invoking [LP17, Proposition 17.20] are asserted from the one-step probability bounds P(Δ = ±1) ≥ p and E[Δ] ≤ 0. A reader must verify that these imply Var(Δ) ≥ 2p when p ≤ 1/4; this is true but is not shown, and a one-line derivation would improve readability.","section":"§2, Theorem 2.5 and §3, Theorem 3.6"},{"comment":"The abstract's wording that 'a family of Markov chains for uniform sampling on M-convex sets does not necessarily exhibit cutoff' is supported by the Q = 2 example in Remark A.2, but Proposition A.1 only gives an upper bound for general Q; the scope of the negative claim could be stated more explicitly to avoid appearing stronger than proved.","section":"Appendix A"}],"recommendation":"accept","confidential_remarks":"For the editor: the paper is a solid contribution to the mixing-time literature. The proof is long but internally consistent in the stated regime, and the open case ρ > 1/2 for the labeled process is explicitly acknowledged rather than hidden. The self-citation [PTZ20] is used only as a source for a Lyapunov-function lemma and is appropriate. I see no citation or scope concerns for a probability journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a good paper and the main theorem is real. Theorem 1.1 — labeled Glauber dynamics cutoff at (1/2)n log n for fixed Q and ρ ≤ 1/2 — is new, and the load-matrix coupling in Section 3 is the actual contribution. I read it as a proof, not a sketch. The drift calculation in Lemma 3.5 is the delicate part, and it holds: #V+ ≤ #O+ plus N ≥ 2n gives nonpositive drift, and the variance bound supplies the 1/(4Q^3) factor needed for the hitting-time argument. Lemma 3.4's φ_n regularization is a nice way to get the O(Q^2/R) failure probability without a Nash-type inequality.\n\nThe unlabeled Theorem 1.2 is not new — the author says it follows from Lacoin–Leblond via a time change — but the proof is self-contained, clean, and the load-profile lemmas feed the labeled argument. The no-cutoff example in Appendix A is small but useful: it shows uniform M-convex chains do not automatically have cutoff, which keeps the paper's claims honest.\n\nSoft spots, in proportion: the labeled theorem requires ρ ≤ 1/2 and fixed Q. The ρ ≤ 1/2 restriction is exactly where the drift in Lemma 3.5 uses N ≥ 2n, and the ρ > 1/2 case is left open, explicitly. The fixed-Q restriction appears in the coupling probabilities (1/Q^2 and 1/Q^3), so nothing here addresses Q growing with n. Those are limitations, not defects. I also noticed the small typo in the proof of Theorem 2.6 ('ρ/N' vs 'ρ/Q') and the final step of Theorem 3.7 leaves implicit the choice of R proportional to e^α, which is what turns the O(1/R^2) bounds into the claimed 1 − O(e^{-2α}). Minor.\n\nThe citation pattern looks fine. The only self-citation with author overlap is PTZ20 Lemma 3.6, used as a Lyapunov-function lemma, and it is appropriately credited. No fitted parameters, no invented entities.\n\nBottom line: this paper is for researchers in mixing times and interacting particle systems. It deserves a serious referee, and I would accept after a light revision. I would bring it to reading group.","headline":"A solid, honest paper: the labeled cutoff theorem is new and the load-matrix coupling is the real contribution; the unlabeled half is a self-contained reproof of a known result, with limitations clearly stated.","tokens_in":30409,"tokens_out":2495,"would_cite":true,"duration_ms":24762,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J10","60K35","82C22"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that a congestion dynamics and its unlabeled load profile both exhibit abrupt cutoff, at times (1/2)n log n and (1/2)(1−ρ)n log n, with windows of order n.","keywords":["cutoff phenomenon","congestion dynamics","Glauber dynamics","generalized exclusion processes","log M-concave distributions","mixing time","total variation","Bernoulli-Laplace urn"],"falsifier":"A concrete test: take Q = 2, κ = 100, n = 150 (so ρ = 3/4) and estimate the labeled chain's T_mix(ε) by simulation or exact computation; if the crossing time is not (1/2)n log n + O(n), then Theorem 1.1 cannot extend to all ρ in the way the authors conjecture. A second test: fix ρ = 1/3 but let Q grow with n (say Q = $n^{{1/2}}$), where the coupling constants 1/Q² and 1/Q³ vanish; if cutoff disappears or the window changes, the fixed-Q assumption is load-bearing.","tokens_in":29446,"feed_emoji":"🎲","tokens_out":7560,"duration_ms":73715,"temperature":0.7,"pith_summary":"This paper studies a congestion dynamics in which n players repeatedly choose among Q resources of capacity κ, with n = ⌊ρN⌋ and N = κQ, in the regime where n grows, Q is fixed, and ρ ≤ 1/2. It proves that the labeled Glauber dynamics has a total-variation cutoff at time (1/2)n log n, while the unlabeled load profile (counting only how many players use each resource) has a cutoff at (1/2)(1−ρ)n log n. These are exact asymptotic benchmarks, not just bounds: the mixing time differs from the cutoff time by only O(n), and the distance to stationarity drops abruptly across that window. Because the unlabeled chain is a natural sampler for a log M-concave distribution, the result supplies one of the few sharp mixing-time laws for such sampling chains. The paper also constructs a uniform-sampling chain on an M-convex set that provably does not exhibit cutoff, showing the phenomenon is not automatic in this discrete-convexity setting.","feed_headline":"Exact mixing cutoff found for congestion dynamics","feed_subtitle":"Labeled players mix at (1/2)n log n; removing labels shifts the cutoff to (1/2)(1−ρ)n log n.","key_machinery":"The argument runs through two objects. For the unlabeled chain, the load profile $\\xi(t)$ is normalized to $\\Delta(t) = \\xi(t)/N - (\\rho/Q) \\in \\mathbb{R}^Q$, and a one-step calculation gives $\\mathbb{E}[\\|\\Delta(t+1)\\|^2 \\mid F_t] \\le (1 - 2\\gamma/N)\\|\\Delta(t)\\|^2 + C/N^2$ with $\\gamma = 1/((1-\\rho)\\rho)$; this contraction predicts the time $(2\\gamma)^{-1}N\\log N = \\tfrac12\\rho(1-\\rho)N\\log N$ needed to approach the typical profile, and a coupling with drift at most $0$ and variance at least $1/(2Q^2)$ finishes the coalescence. For the labeled chain, the paper tracks the load matrix $\\xi^{\\sigma_0}_{u,v}(t)$, the number of players in resource $u$ at time $0$ and resource $v$ at time $t$; Lemma 3.1 shows total-variation distance of the full chain equals that of this matrix chain. The matrix rows contract at rate $1 - 1/n$, giving time $\\tfrac12 n\\log n$, and a second coupling on paired occupancy/vacancy sequences has downward drift that uses $N \\ge 2n$ through the inequality $\\#V^+ \\le \\#O^+$.","core_discovery":"On the paper's own terms, the discovery is a pair of theorems. Theorem 1.1: for fixed Q ≥ 2 and ρ ∈ (0, 1/2], with N = κQ and n = ⌊ρN⌋, the worst-case total-variation mixing time $T_{\\mathrm{mix}}(\\varepsilon)$ of the labeled Glauber dynamics on $S_n$ satisfies $T_{\\mathrm{mix}}(\\varepsilon) = \\tfrac12 n\\log n + O_\\varepsilon(n)$, so there is cutoff at $\\tfrac12 n\\log n$ with window $O(n)$. Theorem 1.2: the unlabeled load-profile chain on $\\Xi_{n,S,\\kappa}$ satisfies $T_{\\mathrm{mix}}(\\varepsilon) = \\tfrac12(1-\\rho)n\\log n + O_\\varepsilon(n)$, equivalently $\\tfrac12\\rho(1-\\rho)N\\log N + O_\\varepsilon(N)$, so it also has cutoff, at a different time. The unlabeled statement is proved self-contained and also follows as a time change of known exclusion-process cutoffs on complete graphs; the labeled statement extends the analysis to track, for each resource $u$ at time $0$, where those players have moved by time $t$. The paper leaves open whether the labeled cutoff holds for $\\rho > 1/2$.","pith_inferences":["Editorial inference: if the $\\rho \\le 1/2$ restriction is only technical, one would expect the labeled cutoff time to become symmetric under $\\rho \\leftrightarrow 1-\\rho$ by complementing vacancies; the paper's own remark says no such equivalence is apparent, so a direct test for $\\rho > 1/2$ (for example $Q=2$, $\\rho=3/4$) would either confirm the same $(1/2)n\\log n$ law or reveal a genuinely dif","Editorial inference: the constants in the couplings degrade as powers of $Q$ (at least $Q^{-2}$ in the unlabeled coupling and $Q^{-3}$ in the labeled one), so the proof gives no information as $Q$ grows with $n$; a natural extension to test is whether cutoff persists when $Q = n^a$, $0 < a < 1$, where the stationary profiles are less concentrated.","Editorial inference: the unlabeled chain's exact log M-concave cutoff time suggests a possible general principle, that for product-form log M-concave measures with a linear drift the cutoff time should be governed by $(2\\lambda)^{-1}n\\log n$ where $\\lambda$ is the smallest nonzero contraction rate of the mean field; one could test this on other explicit potentials such as the uniform sampling case"],"forward_implications":["If Theorem 1.1 is right, the labeled congestion dynamics is an explicit example where Glauber sampling from a log-concave potential has a sharp cutoff with an exact constant, $(1/2)n\\log n$, not merely an $O(n\\log n)$ bound.","If Theorem 1.2 is right, the load-profile chain, a generalized exclusion process on a complete graph with capacities, inherits and sharpens the Bernoulli-Laplace cutoff picture, now at $\\tfrac12\\rho(1-\\rho)N\\log N$.","The two theorems together show that erasing labels changes the cutoff time by the factor $1-\\rho$; for small $\\rho$ the unlabeled chain mixes almost as fast, and for $\\rho$ close to $1/2$ it mixes half as fast.","The no-cutoff example in Appendix A implies that uniform M-convex samplers can mix in $O(n^2)$ without an abrupt transition, so the cutoff property is tied to the non-uniform log M-concave structure of the load-profile model, not to M-convexity alone."],"supporting_citations":[{"why":"Supplies the exclusion-process cutoff on the complete graph from which the unlabeled load-profile result follows by a time change.","marker":"[LL11]"},{"why":"Gives the classical Bernoulli-Laplace cutoff benchmark for the half-full case that this paper generalizes to capacities κ and many resources.","marker":"[DS87]"},{"why":"Provides the O(n log n) mixing-time bound for log-concave samplers that motivates the search for exact cutoff and sets the order benchmark.","marker":"[ALOG+21]"},{"why":"Connects Glauber dynamics for congestion games to sampling from Gibbs distributions, placing the model in the sampling context.","marker":"[Kle23]"},{"why":"Defines log M-concavity and M-convex sets, the discrete-convexity framework that frames the unlabeled chain and the Appendix A example.","marker":"[Mur03]"},{"why":"Used for the standard mixing-time and cutoff definitions and for the coupling hitting-time estimate that converts drift and variance into coalescence.","marker":"[LP17]"},{"why":"Supplies the convex test function φ_n with the one-step Taylor inequality used to prove Lemma 3.4 for the labeled load-matrix chain.","marker":"[PTZ20]"},{"why":"Places the load-profile process among generalized exclusion processes and supplies the particle-system viewpoint.","marker":"[KL99]"}],"fun_headline_variants":["Labeled congestion dynamics cut off at (1/2)n log n","Unlabeled congestion dynamics cut off at (1/2)(1−ρ)n log n","Congestion and exclusion processes show mixing cutoff","Cutoff proven for Glauber dynamics on congestion","Mixing cutoff at half n log n for labeled players"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Both theorems assume the system is at most half full (ρ ≤ 1/2, so N ≥ 2n) and Q is fixed; the labeled coupling's drift stays nonpositive only when that inequality holds, and all constants grow with Q.","fun_headline_variants_meta":{"raw":{"variants":["Labeled congestion dynamics cut off at (1/2)n log n","Unlabeled congestion dynamics cut off at (1/2)(1−ρ)n log n","Congestion and exclusion processes show mixing cutoff","Cutoff proven for Glauber dynamics on congestion","Mixing cutoff at half n log n for labeled players"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000454,"raw_usage":{"total_tokens":2291,"prompt_tokens":962,"completion_tokens":1329,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":1242}},"tokens_in":578,"tokens_out":1329,"duration_ms":12159,"temperature":1.0,"reasoning_tokens":1242,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T16:53:56.674032+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: take Q = 2, κ = 100, n = 150 (so ρ = 3/4) and estimate the labeled chain's T_mix(ε) by simulation or exact computation; if the crossing time is not (1/2)n log n + O(n), then Theorem 1.1 cannot extend to all ρ in the way the authors conjecture. A second test: fix ρ = 1/3 but let Q grow with n (say Q = $n^{{1/2}}$), where the coupling constants 1/Q² and 1/Q³ vanish; if cutoff disappears or the window changes, the fixed-Q assumption is load-bearing.","supporting_citations":[],"review_version":1}