{"id":"c574f782-8ac6-40c0-90b3-aa827f002daa","arxiv_id":"2412.16952","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 p-spin Curie-Weiss model, Glauber dynamics has three mixing-time scales determined by the number and curvature of local maxima of a function H, and a restricted version mixes fast even in the slow regime.","lead":"This paper maps the convergence speed of Glauber dynamics for the p-spin Curie-Weiss model, a network spin system with higher-order interactions, into three regimes: fast (N log N), critical (N^{3/2}), and exponentially slow, covering the whole parameter space except one curve. The result gives practitioners a practical guide to when standard simulation converges quickly and includes a fast restricted sampler for the slow metastable regime.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Upper-bound proof for the S_p regime jumps from (3.12) to an O(N^{3/2}) hitting-time claim that the displayed cubic-decay inequality does not imply; the missing diffusion estimate (and a similar LLP citation for λ′≤−1) needs to be supplied.","rationale":"The reader's weakest assumption points to Mukherjee-Son-Bhattacharya (2021); that is a real dependency but it is published work and its use in the R_p lower bound and C_p bottleneck appears legitimate. My reading found a more immediate soft spot inside the paper's own argument: the S_p upper bound is the least self-contained step of Theorem 2.1. The displayed inequality (3.12) has the wrong scaling if pushed all the way to 1/N; the appeal to LLP Step 1 is doing essential work that is not shown. This is not an accusation that the theorem is false, nor a claim that the authors are careless; it is a precise request for the missing diffusion estimate. The lower bound for S_p and the rest of the R_p/C_p arguments appear sound given the cited prior work. Therefore the conditional verdict remains appropriate: accept after the missing hitting-time/diffusion estimate is supplied, and after the minor exp(Ω(N)) versus e^{Ω(√N)} inconsistency in Section 1 is fixed. A separate wording issue in Theorem 2.2 (the definition of m− as the largest stationary point not exceeding m+ literally includes m+ itself) should also be clarified, but it does not bear on Theorem 2.1 and is not the primary concern here.","tokens_in":31630,"tokens_out":43925,"duration_ms":398374,"concrete_test":"Independently derive the hitting-time estimate for Section 3.3: starting from θ_T = O(1), use (3.12) to reach θ ≈ cN^{−1/4} in O(N^{3/2}), then bound the probability that |e_t| falls to 1/N within an additional O(N^{3/2}) using the same martingale/diffusion argument as the paper's lower bound (3.14)-(3.15). If the argument only works with τ0 redefined as min{t : |e_t| ≤ cN^{−1/4}}, the text must be corrected; additionally, write out the LLP-based argument for λ′(c∗) ≤ −1 in Section 3.1, since Lemma 3.4 requires |λ′|<1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing soft spot is the upper-bound half of the p-special case (Theorem 2.1(3), Section 3.3). After the burn-in, the proof derives the cubic-drift inequality (3.12), θ_{t+1} ≤ θ_t − (c/N)θ_t^3, and then asserts, by analogy with Step 1 of Theorem 4.1 in Levin-Luczak-Peres (2007), that P(τ0 > aN^{3/2}) → 0 with τ0 = min{t : |e_t| ≤ 1/N}. This does not follow from (3.12) alone: for θ_T = O(1), the deterministic inequality only forces |e| to fall to order N^{−1/4} in O(N^{3/2}); reaching 1/N requires a separate diffusion/absorption estimate that is not written out. The same LLP citation is also used in Section 3.1 for the λ′(c∗) ≤ −1 subcase of R_p, where the path-coupling contraction of Lemma 3.4 is unavailable because |λ′| > 1. If the LLP analogy is valid the theorem survives; if the missing diffusion estimate cannot be supplied with the stated τ0, the O(N^{3/2}) upper bound is unproved and Theorem 2.1(3) would need revision. This is an omitted proof rather than a demonstrated error, so it supports the existing conditional verdict.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies heat-bath Glauber dynamics for the p-spin Curie-Weiss model with external field h, for p ≥ 3. It partitions the parameter space into three sets: R_p, where H_{β,h,p} has a unique local maximizer with negative second derivative and no other stationary point; C_p, where H has multiple local maximizers; and S_p, where H has a unique local maximizer with zero second derivative. Theorem 2.1 asserts mixing times Θ(N log N), e^{Ω(N)}, and Θ(N^{3/2}) in these three regions, leaving only a one-dimensional boundary curve B_p unresolved. Theorem 2.2 constructs a restricted Glauber dynamics in the metastable case that mixes in time Θ(N log N). The proofs use drift and burn-in estimates, a path-coupling contraction argument, a bottleneck lower bound, and a convex-minorant drift analysis, leaning on fluctuation results from Mukherjee, Son and Bhattacharya (2021).","tokens_in":31884,"tokens_out":7744,"duration_ms":80338,"significance":"If the main theorem is fully established, the paper gives an almost complete scaling classification of Glauber mixing times for the p-spin Curie-Weiss model, extending the classical results of Levin, Luczak and Peres (2007) from p = 2 to p ≥ 3 in the presence of a field. The explicit phase geometry in Appendix A, the matching lower bounds, and the restricted-dynamics mixing result in Theorem 2.2 are substantive contributions. The proofs are structured and mostly credible, and the paper makes good use of prior rigorous results on magnetization fluctuations. The main weaknesses are two upper-bound steps that are delegated to sketched analogies with existing proofs rather than written out; these steps are load-bearing for the central theorem.","major_comments":[{"comment":"The proof of the S_p upper bound jumps from the cubic drift inequality θ_{t+1} ≤ θ_t − (c/N)θ_t^3 to the conclusion that P(τ0 > aN^{3/2}) → 0 uniformly in N, where τ0 = min{t : |e_t| ≤ 1/N}. This implication is not justified by the displayed inequality alone: the deterministic flow dθ/dt = −(c/N)θ^3 starting from θ = O(N^{−1/4}) reaches order N^{−1} only on a time scale of order N^3, not N^{3/2}. The O(N^{3/2}) hitting-time claim therefore depends on the diffusive fluctuations of the magnetization chain, and requires a separate submartingale or second-moment estimate for the process e_t. Moreover, the target scale 1/N is far below the N^{−1/4} stationary fluctuation scale given by Mukherjee-Son-Bhattacharya Theorem 3.1(3), so the analogy with Theorem 4.1 of Levin-Luczak-Peres (2007) is not immediate. This missing estimate is needed before the upper bound in Theorem 2.1(3) is rigorous.","section":"Section 3.3, Eq. (3.12)"},{"comment":"In the p-locally regular case, the paper handles the case λ′(c∗) ≤ −1 by first showing that e_t hits the small neighborhood |e_t| ≤ 1/N in time O(N log N), and then saying that 'the same arguments' as in the proof of Theorem 4.1 of Levin-Luczak-Peres (2007) complete the upper bound. This is not automatic: Lemma 3.4 requires sup_c |λ′(c)| < 1, and the burn-in reduction only guarantees contraction inside a neighborhood of c∗, not a global path-coupling contraction when |λ′(c∗)| > 1. Theorem 4.1 of LLP concerns the case λ′(0) = 2β ∈ (−1,1), so the cited argument does not by itself cover the oscillatory drift case. The manuscript should either provide a coupling or contraction argument that works when |λ′(c∗)| > 1 after the hitting time τ0, or identify precisely which part of LLP applies and why.","section":"Section 3.1, proof for λ′(c∗) ≤ −1"}],"minor_comments":[{"comment":"The introduction states that the mixing time for p-locally critical points is e^{Ω(√N)}, while Theorem 2.1 and the abstract state e^{Ω(N)}; this should be corrected to e^{Ω(N)}.","section":"Introduction, Section 1"},{"comment":"There are several typos, such as 'Theoorem' in Section 3 and 'exitence' in Section 3.3, which should be fixed in a revision.","section":"Throughout"},{"comment":"The simulation figure would be more informative if the axes and the precise stopping rule for the capped mixing times were described, and if the number of repetitions or error bars were reported.","section":"Figure 2"},{"comment":"The proof of Proposition D.1 cites an unnamed displayed inequality after the definition of A_t; the displayed inequality should be numbered or referenced explicitly for readability.","section":"Appendix D"}],"recommendation":"major_revision","confidential_remarks":"The two major comments concern omissions in the written proof rather than demonstrated errors, and they are likely fixable within the manuscript's scope. I did not find evidence of circularity: the dependence on Mukherjee-Son-Bhattacharya (2021) is a normal reliance on published prior results, and the central mixing-time claims are derived from the model and Markov-chain arguments. The novelty disclosure regarding Mikulincer and Piana (2024) appears adequate. If the missing diffusion and contraction estimates cannot be supplied, the upper-bound halves of Theorem 2.1(3) and part of Theorem 2.1(1) would need to be revised; hence my recommendation is major revision rather than acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is the first essentially complete mixing-time phase diagram for p-spin Curie-Weiss with external field, and the restricted fast sampler for metastable states is a genuine addition. The p=2 trichotomy is prior work, but the p≥3 geometry, with slow mixing starting before the magnetization phase transition, is new. The proof structure is credible: drift and coupling for R_p, a bottleneck lower bound for C_p, and a convex-minorant drift argument for S_p. The dependence on Mukherjee-Son-Bhattacharya (2021) is real but not circular; those theorems are about stationary fluctuations and phase geometry, not mixing times, so this is a reliance on prior results, not a circular derivation.\n\nThe softest spot is the upper-bound half of the S_p case. The proof derives θ_{t+1} ≤ θ_t − (c/N)θ_t^3 and then jumps to the conclusion that τ0 = min{t: |e_t| ≤ 1/N} satisfies P(τ0 > aN^{3/2}) → 0, citing Step 1 of Theorem 4.1 in LLP. That deterministic inequality only gets |e| down to order N^{−1/4} in N^{3/2} steps; reaching 1/N needs a separate diffusion or absorption estimate that is not written out. The LLP analogy may supply it, but as it stands the O(N^{3/2}) upper bound is not fully demonstrated. This is an omitted argument, not a known counterexample, but it is load-bearing.\n\nTwo smaller issues. The abstract and Theorem 2.1 state exp(Ω(N)) for C_p while the introduction says e^{Ω(√N)}; those should be consistent. And the claim that the parallel-restricted sampler produces 'very nearly a sample' from the Curie-Weiss measure is too loose; either quantify the total variation distance or soften the phrase.\n\nThe boundary curve Bp is left open, which is an honest limitation. The crude N^{4/3} lower bound there is fine as a start. For a reader working on mean-field Glauber dynamics or metastability, this paper is worth the time. I would send it to a serious referee, with the request to either supply the missing diffusion estimate in the S_p upper bound or adjust the statement. If that step is fixed, the paper is a solid contribution.","headline":"Solid new phase diagram for p-spin Curie-Weiss Glauber dynamics, but the S_p upper bound needs a missing diffusion estimate spelled out.","tokens_in":32491,"tokens_out":5184,"would_cite":true,"duration_ms":46517,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","60J10"],"pacs":[],"model":"deepseek-v4-flash","headline":"For the p-spin Curie-Weiss model with $p \\geq 3$, the mixing time of the Glauber dynamics takes exactly three orders of magnitude — $\\Theta(N \\log N)$, $\\Theta(N^{3/2})$, or $\\exp(\\Omega(N))$ — depending on the local maximizers of a…","keywords":["Glauber dynamics","mixing time","Curie-Weiss model","p-spin model","metastability","phase transition","mean-field Ising model","Markov chain Monte Carlo"],"falsifier":"Take a point on the boundary curve $B_p$, for example for $p=4$ the curve $h=U(\\beta)$ where $H$ has a stationary inflection point alongside a local maximizer, simulate the Glauber dynamics for increasing $N$, and estimate $t_{\\mathrm{mix}}(0.35)$; the paper guarantees only $\\Omega(N^{4/3})$ there, so observing $\\Theta(N\\log N)$ or $\\exp(\\Omega(N))$ would show the three-regime classification misses a fourth regime, while observing $\\Theta(N^{3/2})$ would support the natural extrapolation.","tokens_in":31384,"feed_emoji":"🧲","tokens_out":7682,"duration_ms":64182,"temperature":0.7,"pith_summary":"This paper proves that the speed at which the heat-bath Glauber dynamics converges to equilibrium in the $p$-spin Curie-Weiss model is controlled entirely by the local maximizers of a one-dimensional free-energy function $H_{\\beta,h,p}(x)=\\beta x^p+hx-I(x)$, where $I$ is the binary entropy. For $p \\geq 3$, it identifies three regions of the $(\\beta,h)$ parameter space: where $H$ has a unique stable maximizer, mixing takes $\\Theta(N \\log N)$; where $H$ has multiple local maximizers, mixing takes at least $\\exp(\\Omega(N))$; and where the unique maximizer has zero curvature, mixing takes $\\Theta(N^{3/2})$. Almost the whole parameter plane is covered, with only a curve of stationary inflection points left open. The paper also constructs a restricted Glauber dynamics that, even in the metastable multi-maximizer region, mixes in $\\Theta(N \\log N)$, allowing approximate simulation in polynomial time.","feed_headline":"Three mixing-time regimes found for p-spin Curie-Weiss","feed_subtitle":"Beyond a boundary curve, the spin dynamics runs in fast, critical, or exponentially slow time, set by the shape of one free-energy curve.","key_machinery":"The central object is the mean-field free-energy function $H_{\\beta,h,p}(x)=\\beta x^p+hx-I(x)$ on $[-1,1]$, together with the mean-magnetization update map $\\lambda(c)=\\tanh(p\\beta c^{p-1}+h)$; fixed points of $\\lambda$ are exactly stationary points of $H$, and the sign of $\\lambda'(c_*)-1$ equals the sign of $H''(c_*)$. The argument runs through the drift identity $\\mathbb{E}(c_{t+1}-c_t \\mid c_t=c)=(\\lambda(c)-c)/N$, burn-in epochs with exponential-in-$\\sqrt{N}$ concentration, a coupling that contracts expected Hamming distance when $\\lambda'$ is bounded below $1$ near the unique fixed point, a bottleneck-ratio bound for the exponential lower bound, and a Taylor expansion at the zero-curvature fixed point for the $N^{3/2}$ critical law. The restricted dynamics in the metastable region is analyzed as a birth-and-death chain on the sum $S_t^+$ with rejection at the basin boundary, controlled through expected hitting times.","core_discovery":"The central claim is Theorem 2.1: for every $\\varepsilon \\in (0,1/2)$, $p \\geq 3$ and $(\\beta,h) \\in \\Theta$, if the point is $p$-locally regular then $t_{\\mathrm{mix}}(\\varepsilon)=\\Theta_\\varepsilon(N \\log N)$; if it is $p$-locally critical then $t_{\\mathrm{mix}}(\\varepsilon) \\geq \\exp(\\Omega_\\varepsilon(N))$; and if it is $p$-special then $t_{\\mathrm{mix}}(\\varepsilon)=\\Theta_\\varepsilon(N^{3/2})$. The regions are defined by the number of local maximizers of $H_{\\beta,h,p}$ and the sign of $H''$ at the maximizer, and a direct corollary is that non-global local maximizers — metastable states — are what cause exponentially slow mixing, not the global maximizer structure alone. The complement of the three regions is a one-dimensional curve on which $H$ has a stationary inflection point together with a local maximizer; the paper leaves the exact mixing order there open and proves only a crude polynomial lower bound $\\Omega(N^{4/3})$. Theorem 2.2 adds that, in the metastable region, a restricted chain that rejects moves crossing the basin boundary of the largest local maximizer mixes in $\\Theta(N \\log N)$ to its stationary distribution.","pith_inferences":["[Editorial inference] The boundary curve likely carries its own mixing exponent, plausibly $N^{4/3}$ up to logarithmic factors, because the crude lower bound saturates a Taylor expansion around the stationary inflection point; this extrapolation is not claimed in the paper.","[Editorial inference] The criterion 'mixing is fast exactly when the free-energy function has one stable local maximizer' should transfer to other mean-field spin systems such as the $p$-spin Curie-Weiss Potts model, where the paper names a counting question as the natural next step.","[Editorial inference] The restricted-dynamics construction yields a concrete sampling recipe — run parallel chains in each basin for $C N \\log N$ steps and combine them with weights proportional to $(m_i^2-1)H''(m_i)^{-1/2}$ — that could be benchmarked numerically against naive Glauber dynamics in the metastable region."],"forward_implications":["Away from the boundary curve, the computational cost of exact simulation from the $p$-spin Curie-Weiss model is now known to three orders: fast $N \\log N$, critical $N^{3/2}$, and exponentially slow.","In the metastable region, starting the chain near a non-global local maximizer causes an exponentially long delay, so the practical prescription is to initialize inside the concave basin of a global maximizer.","For $p \\geq 3$ and $h=0$, the mixing transition threshold $\\beta'_p$ is strictly smaller than the threshold $\\tilde{\\beta}_p$ for the asymptotics of the magnetization, so there is a window where magnetization looks Gaussian but mixing is already exponentially slow.","The restricted dynamics of Theorem 2.2 lets one approximate samples from a $p$-locally critical model in $O(N \\log N)$ time by running parallel restricted chains around each global maximizer and mixing their outputs with Gaussian weights.","The boundary curve $B_p$ is a genuine gap in the classification: the paper proves only $t_{\\mathrm{mix}}=\\Omega(N^{4/3})$ there, and the exact order remains open."],"supporting_citations":[{"why":"Supplies the stationary magnetization fluctuation widths ($O_P(N^{-1/2})$ and $N^{1/4}$ at special points) and the explicit shape of $H''$ used to construct the regions $R_p$, $C_p$ and $S_p$, via their Theorem 3.1 and Lemmas F.2, F.3 and 3.7.","marker":"Mukherjee, Son and Bhattacharya (2021)"},{"why":"Provides the $p=2$ baseline regimes $\\Theta(N\\log N)$, $\\Theta(N^{3/2})$ and $\\exp(\\Omega(N))$, and the critical-power-law and restricted-chain proof templates that this paper adapts.","marker":"Levin, Luczak and Peres (2007)"},{"why":"Supplies the coupling theorem and bottleneck-ratio theorems used to convert Hamming-distance contraction and conductance estimates into mixing-time bounds.","marker":"Levin, Peres and Wilmer (2006)"},{"why":"Contributes the attractor/burn-in drift framework for the magnetization chain that underlies Lemmas 3.2 and 3.3.","marker":"DeMuse, Easlick and Yin (2019)"}],"fun_headline_variants":["Three mixing speeds for p-spin Curie-Weiss","Fast, critical, or slow: p-spin Glauber mixing","Metastable states set Glauber mixing time","Free-energy curvature dictates mixing speed","p-spin Curie-Weiss: three mixing-time regimes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole three-regime phase diagram rests on previously established facts about the stationary magnetization — its fluctuations are of order $N^{-1/2}$ in the regular and critical regions, its scaling is $N^{-1/4}$ at special points, and the second-derivative shape of $H$ is exactly as assumed for every $p \\geq 3$; if any of those facts fails at some parameter values, the claimed exponents and phase boundaries would need revision.","fun_headline_variants_meta":{"raw":{"variants":["Three mixing speeds for p-spin Curie-Weiss","Fast, critical, or slow: p-spin Glauber mixing","Metastable states set Glauber mixing time","Free-energy curvature dictates mixing speed","p-spin Curie-Weiss: three mixing-time regimes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00048,"raw_usage":{"total_tokens":2549,"prompt_tokens":1292,"completion_tokens":1257,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":908,"completion_tokens_details":{"reasoning_tokens":1178}},"tokens_in":908,"tokens_out":1257,"duration_ms":11043,"temperature":1.0,"reasoning_tokens":1178,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:57:43.073480+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a point on the boundary curve $B_p$, for example for $p=4$ the curve $h=U(\\beta)$ where $H$ has a stationary inflection point alongside a local maximizer, simulate the Glauber dynamics for increasing $N$, and estimate $t_{\\mathrm{mix}}(0.35)$; the paper guarantees only $\\Omega(N^{4/3})$ there, so observing $\\Theta(N\\log N)$ or $\\exp(\\Omega(N))$ would show the three-regime classification misses a fourth regime, while observing $\\Theta(N^{3/2})$ would support the natural extrapolation.","supporting_citations":[],"review_version":1}