{"id":"156845dd-2614-4fc8-85ed-5029b5af3fde","arxiv_id":"2502.06429","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For the Curie-Weiss model at β>1, the magnetization conditioned on staying positive converges uniformly in time to the positive mean-field steady-state trajectory, with polynomial-in-n error.","lead":"A new theorem shows that in the low-temperature Curie-Weiss model, the magnetization of n spins, conditioned to never fall below a small positive threshold, stays close to the deterministic mean-field trajectory uniformly in time, with error shrinking polynomially in n. This is a proof of concept for using quasi-stationary distributions to obtain uniform propagation of chaos in metastable mean-field systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the central argument is sound; the issues raised by the reader are typos that do not affect the theorem.","rationale":"The reader's verdict is CONDITIONAL, driven by three localized errors or typos. My pass confirms those inaccuracies exist: Lemma 2.1's support statement uses a larger set than K_n, Lemma 2.5's formula for r'_β(m+) is off by a factor of 2 (though only its sign is used), and the exponent in the (4.12) computation is written without a √n. None of these changes the conclusion of the theorem. The reader's identified weakest assumption, non-degeneracy of g at m+, is actually proved in Lemma 2.2 and is not an unverified assumption; it is a property of the Curie-Weiss model, and the paper only claims a proof of concept for this model. I therefore do not regard it as the most load-bearing concern. The most delicate step I found is the unproved uniform bound on P^n_s V_n for small s in the proof of Theorem 1, but this follows from the existing Lyapunov estimates and does not appear to be a real gap. Since the central theorem is supported and the remaining issues are fixable typos, I see no reason to move the reader's verdict; CONDITIONAL remains appropriate until the typos are corrected.","tokens_in":30436,"tokens_out":47125,"duration_ms":405925,"concrete_test":"Independently re-derive the uniform bound in Theorem 1's proof for δ_m0 P^n_s V_n with m0 ∈ [η,1] and s ∈ [0,τ], using the semigroup inequality M^n_s g ≤ e^{-γs}g + C/(γn) and the uniform lower bound of h_n on [η,1]. If this bound fails for some n and s, the prefactor in Theorem 1 must be revised; if it holds, Theorem 2's conclusion is unchanged.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I read the proof of Theorem 2 as a chain: Proposition 1.1 gives finite-time chaos, Lemma 2.4 gives bn ≤ C/√n and the survival estimates, Lemmas 3.1 and 3.2 feed the Harris argument for Theorem 1, and Section 4.3 splits time into short, intermediate, and long ranges. I found no step where the argument breaks. The non-degeneracy g''(m+) > 0 is proved in Lemma 2.2 and is used only for the specific Curie-Weiss potential, so it is not a hidden assumption. The terse assertion in Theorem 1's proof that δ_m0 P^n_s V_n is uniformly bounded for s ∈ [0, τ] is justified by the continuous-time version of Lemma 3.2, namely M^n_s g ≤ e^{-γs}g + C/(γn), together with uniform boundedness of h_n^{-1} on [η, 1]. The derivation of (4.12) contains a small exponent inaccuracy (an intermediate term is written as n^{-1/2+(c1+c2)/(2S)} instead of n^{-1/2+(c1+c2)/(2S√n)}), but the correct term is n^{-c3/(2S)+o(1)}, so the polynomial bound still holds. The genuine defects are the typos the reader lists; they do not threaten the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the continuous-time Curie-Weiss model at inverse temperature beta > 1, where the mean-field ODE (1.7) has two stable equilibria +/-m_+ while the finite-n magnetization is ergodic with a unique symmetric invariant law. The authors condition the magnetization to stay above a level epsilon > 0 and prove two results: Theorem 1 establishes convergence of the conditioned law to a quasi-stationary distribution at rate C n e^{-ct}; Theorem 2 establishes uniform-in-time propagation of chaos for the conditioned process, sup_{t>=0} |E_{m0}(f(m^n_t)|tau_n>t) - f(m_t)| <= C(||f||_inf + ||f||_lip)/n^alpha, where m_t is the solution of the mean-field ODE started at m0. The proof uses Perron-Frobenius eigen-elements of the killed semigroup, a Doob h-transform, a Meyn-Tweedie Harris argument with Lyapunov function V_n = h_n^{-1} g and a local Doeblin condition on a shrinking interval around m_+, and a two-scale time decomposition: short times via finite-time propagation of chaos (Proposition 1.1), intermediate times via a generation-of-chaos estimate for an auxiliary process (Lemma 4.3 and (4.2)), and long times via the QSD convergence and the convergence of the QSD to delta_{m_+} (Theorem 1 and (4.12)).","tokens_in":30673,"tokens_out":32798,"duration_ms":220572,"significance":"If the results hold, the paper provides a quantitative uniform-in-time propagation of chaos in a metastable mean-field setting by conditioning on survival, overcoming the generic obstruction that the unconditioned particle system and the non-linear limit have incompatible long-time behaviors. The proof is self-contained: Proposition 1.1 is proved in the paper, the eigen-elements (b_n, h_n) are defined via Perron-Frobenius, and their asymptotics (Lemmas 2.4 and 2.5) are derived from the dynamics. The method, combining Harris/Meyn-Tweedie theory with n-dependent Lyapunov and Doeblin estimates, is a promising template for other metastable mean-field models, and the authors correctly state that the non-degeneracy g''(m_+) > 0 is a property of the specific Curie-Weiss potential, so they do not overclaim generality. The paper also clearly acknowledges that the convergence rate n^{-alpha} is not optimized. The main strengths are the explicit two-scale argument, the careful handling of the n-dependence of the spectral gap via the Harris approach rather than a bare Perron-Frobenius bound, and the transparent presentation of the auxiliary-process construction.","major_comments":[],"minor_comments":[{"comment":"In the generator expansion for the scaled gap process, the second term reads f(m - 4/sqrt(n)) - f(sqrt(n) m); the second argument should be m, not sqrt(n) m, so that the bracket is f(m - 4/sqrt(n)) - f(m). This is clearly a typo and does not affect the subsequent drift/diffusion computation, but it should be corrected.","section":"Lemma 2.1, proof"},{"comment":"The proof reuses the symbol K_n for a larger interval than in the statement: the statement defines K_n = [m_+ - omega/sqrt(n), m_+ + omega/sqrt(n)], while the proof writes K_n = [m_+ - 3omega/sqrt(n), m_+ + 3omega/sqrt(n)] and later states Delta_0 ≈ 2omega/sqrt(n), which is inconsistent with the extremes of that larger set. Please introduce a separate symbol for the larger set (e.g., K'_n) and reconcile the constants in Delta_0 and alpha_n.","section":"Lemma 2.1, proof"},{"comment":"The condition 'for all m in E^epsilon_n such that |m - eta| < eta' should read '|m - m_+| < eta'; the letter eta is used for the margin, not for the minimum point.","section":"Lemma 2.4, proof"},{"comment":"The display P(sup_{0<=t<=t0} |M_t| > r) <= C E(|M_{t0}|) is a use of Doob's weak (1,1) maximal inequality and should include the factor 1/r or explicitly state that the constant C depends on the fixed r; as written, the inequality is formally missing the 1/r factor.","section":"Lemma 2.4, proof"},{"comment":"In the intermediate-time range t1 <= t <= t2, the text says 'for some tilde C2, alpha1 > 0' but the exponent should be alpha2, consistently with the notation alpha2 introduced earlier in the same paragraph.","section":"Section 4.3, proof of Theorem 2"},{"comment":"The line 'Plugging (4.7), (4.7) and (4.8) back into (4.6)' contains a duplicate reference; it should be 'Plugging (4.7) and (4.8) back into (4.6)'.","section":"Section 4.3, proof of Theorem 2"},{"comment":"In the estimate after choosing t = ln(n)/(2(c1+c2+c3)), the displayed simplification can be made more transparent: the first term e^{c2 t/sqrt(n) + c1 t}/sqrt(n) is bounded by n^{-c3/(2S)} after using e^{c2 t/sqrt(n)} <= n^{c2/(2S)}, and the equality with 2 n^{-c3/(2S)} then follows; the current text is correct but slightly compressed.","section":"Section 4.3, proof of Theorem 2"}],"recommendation":"minor_revision","confidential_remarks":"I concur with the reader and the skeptic that the central proof is sound and that the issues are typos and notational sloppiness. The paper is a solid contribution to the metastable propagation-of-chaos literature and is within the scope of math.PR. The K_n notation in Lemma 2.1 should be fixed, and the small typos listed in the minor comments should be corrected. I see no need for a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick read of Journel & Le Bris (arXiv:2502.06429). The genuinely new thing here is the first uniform-in-time propagation of chaos result for a conditioned mean-field particle system in a metastable setting: they show that for β>1, the Curie-Weiss magnetization, conditioned to stay above ε, stays within polynomial distance of the deterministic positive-well trajectory, uniformly for all t. The rate is n^{-α} with explicit α. That is a real step, not a repackaging. The strategy — QSD for the killed process, Doob transform, Harris/Meyn-Tweedie on the h-transform — is coherent and self-contained. Proposition 1.1 is proved in the paper; the eigen-elements h_n and b_n are constructed via Perron-Frobenius and estimated from the dynamics. No circularity that I can see. The paper also proves a quantitative convergence to the QSD (Theorem 1) that is interesting in its own right.\n\nWhat's soft: the written proof is not as clean as the mathematics. The reader flagged a typo in Lemma 2.1 (the generator expansion has f(√n m) in the second term, should be f(m−4/√n) or equivalent), a similar slip in Lemma 2.4's Doob inequality, and a sloppy exponent in the derivation of (4.12). I checked (4.12): the intermediate expression should contain n^{-1/2+(c1+c2)/(2S√n)} rather than n^{-1/2+(c1+c2)/(2S)}, but the final bound n^{-c3/(2S)+o(1)} still holds, so the theorem survives. The typos are genuine but local; they make verification harder than it should be. Also worth noting: the proof uses g''(m+)>0 (Lemma 2.2), which is proved for Curie-Weiss. That non-degeneracy is not a hidden assumption, but it does mean the method, as written, leans on a property that won't hold in every metastable McKean–Vlasov model. The paper frames itself as a proof of concept, which is fair; the extension to double-well SDEs is a hope, not a result.\n\nBottom line: the main theorem is new and, as far as I can tell, correct. The typos should be fixed before this becomes a reference, but they don't break the argument. I'd send it to a serious referee. It's worth citing for anyone working on metastable mean-field limits or QSD-based approaches to uniform-in-time chaos.\n\nBest.","headline":"Solid new result on conditioned propagation of chaos for Curie-Weiss; the typos are real but fixable.","tokens_in":31266,"tokens_out":1803,"would_cite":true,"duration_ms":16034,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","60J27","82C22"],"pacs":[],"model":"deepseek-v4-flash","headline":"For the low-temperature Curie-Weiss model, conditioning on survival makes propagation of chaos hold uniformly in time, with an explicit polynomial error bound.","keywords":["Curie-Weiss model","uniform in time propagation of chaos","metastability","quasi-stationary distribution","Glauber dynamics","mean-field limit","Doob h-transform","magnetization"],"falsifier":"Run the killed Glauber dynamics at $\\beta=2$ with $n$ up to $10^4$, start at $m_0=0.9$, and estimate $\\sup_{t\\ge 0}|E(m^n_t|\\tau_n>t)-m_t|$ for $f(x)=x$; if this quantity fails to decay as $n^{-\\alpha}$ for some $\\alpha>0$, Theorem 2 is false. A more structural test: modify the flip rates so the effective potential has a quartic minimum at $m_+$, i.e. $g(m)\\sim(m-m_+)^4$, and measure the stationary conditional fluctuations around $m_+$; if their standard deviation decays as $n^{-1/4}$ rather than $n^{-1/2}$, the non-degeneracy assumption that carries the proof has failed exactly as predicted.","tokens_in":30143,"feed_emoji":"🧲","tokens_out":8960,"duration_ms":75564,"temperature":0.7,"pith_summary":"The paper establishes that, in the Curie-Weiss model at inverse temperature $\\beta>1$, the magnetization process conditioned to stay above a small positive level remains uniformly close, for all times, to the deterministic solution of the mean-field ODE started from the same point: the error is at most $C(\\|f\\|_\\infty+\\|f\\|_{\\mathrm{lip}})n^{-\\alpha}$ for any Lipschitz observable. This matters because, unconditioned, the particle system eventually flips to the symmetric negative well, so ordinary uniform-in-time propagation of chaos is impossible. The mechanism is to show that the conditioned chain converges to a quasi-stationary distribution, that this distribution is polynomially close to $\\delta_{m_+}$, the minimizer of the positive well, and that the short, intermediate, and long time scales can be bridged. The authors present the result as a proof of concept for metastable mean-field systems beyond this single model.","feed_headline":"Surviving spins track their mean-field limit forever","feed_subtitle":"In the low-temperature Curie-Weiss model, conditioning on positive magnetization restores uniform-in-time propagation of chaos.","key_machinery":"The load-bearing object is the killed semi-group $M^n_t f(m)=E_m(f(m^n_t)\\mathbf{1}_{\\tau_n>t})$ and its Doob $h$-transform $P^n_t f=e^{b_n t}h_n^{-1}M^n_t(h_n f)$, built from the Perron-Frobenius eigenpair $(b_n,h_n)$ of the killed generator. The transformed process is Markov, and the paper verifies for it a local Doeblin condition on the small set $K_n=[m_+-\\omega/\\sqrt n,m_++\\omega/\\sqrt n]\\cap E^\\varepsilon_n$ and a Lyapunov condition with $V_n=h_n^{-1}g$. The non-degeneracy $g''(m_+)>0$ is what makes the sublevel sets of $V_n$ sit inside $K_n$ with width $C/\\sqrt n$ and makes the ratio $(g')^2/g$ tend to $2g''(m_+)>0$, so the contraction argument closes. A Harris-type contraction theorem in a weighted total-variation norm then gives Theorem 1's explicit rate, and the quasi-stationary distribution estimates feed the three-regime proof of Theorem 2.","core_discovery":"The central claim is Theorem 2: for $\\beta>1$ and any $\\eta\\in(\\varepsilon,m_+)$, there exist $C,\\alpha>0$ such that for every Lipschitz $f:[0,1]\\to\\mathbb{R}$ and every initial magnetization in $[\\eta,1]\\cap E^\\varepsilon_n$, one has $\\sup_{t\\ge 0}|E_{m_0}(f(m^n_t)|\\tau_n>t)-f(m_t)|\\le C(\\|f\\|_\\infty+\\|f\\|_{\\mathrm{lip}})n^{-\\alpha}$, where $m_t$ is the solution of the mean-field ODE with initial condition $m_0$. The proof splits time into three regimes: short times use the classical finite-time propagation of chaos; intermediate times use a generation-of-chaos estimate that exploits convexity of the potential near $m_+$ and a Lyapunov decay; long times use Theorem 1, which gives an explicit $C n e^{-ct}$ convergence of the conditioned law to a unique quasi-stationary distribution, together with a polynomial closeness of that distribution to $\\delta_{m_+}$. Thus the paper establishes that metastable conditioning restores uniformity in time, with an explicit though deliberately non-sharp rate.","pith_inferences":["The same mechanism should extend to the McKean-Vlasov SDE with double-well confining potential and quadratic interaction named as a next target, provided each well's minimum is non-degenerate; a flat quartic well would change the fluctuation scale and likely require a different small set.","The polynomial rate in $n$ is probably far from optimal; optimizing the cutoffs $t_1\\approx \\ln n$ and $t_2\\approx n^{1/4}$ could improve the exponent, and the proof itself suggests the rate is limited by the coupling rather than by the phenomenon.","A quantitative prediction of the framework is that the conditional fluctuation scale around $m_+$ is of order $n^{-1/2}$, inherited from the central limit theorem; this could be tested by measuring $\\mathrm{Var}(m^n_t|\\tau_n>t)$ under the quasi-stationary distribution at large $t$.","For models whose basins of attraction are not explicit, the paper's proof of concept suggests a two-step program: first construct a metastable subset of the basin, then repeat the killed-process, quasi-stationary-distribution argument inside it."],"forward_implications":["The finite-$n$ conditional law $\\nu^n_t$ converges to a unique quasi-stationary distribution at rate $C n e^{-ct}$, uniformly over initial points in $[\\eta,1]$.","The quasi-stationary distribution itself converges to $\\delta_{m_+}$ at polynomial rate $n^{-\\alpha}$ in Lipschitz observables.","Consequently the killed magnetization process is indistinguishable, uniformly in time and with explicit error, from the deterministic mean-field trajectory as long as it has not hit the death level.","The three-regime proof yields explicit, if non-optimal, polynomial exponents that can in principle be computed from $\\beta$, $\\eta$, and $\\varepsilon$.","This is the first uniform-in-time propagation of chaos obtained through quasi-stationary distributions for a conditioned particle system, and the paper frames it as a template for other metastable mean-field models."],"supporting_citations":[{"why":"Supplies the Harris-type contraction theorem used as Proposition 3.1 to convert the Doeblin and Lyapunov conditions into the explicit exponential convergence of Theorem 1.","marker":"[26]"},{"why":"Gives the quasi-stationary distribution survival identity $E_{\\nu_\\infty^n}(\\mathbf{1}_{\\tau_n>t})=e^{-b_n t}$, which is used to compare the QSD with $\\delta_{m_+}$ at polynomial rate.","marker":"[37]"},{"why":"Provides the generator-convergence theorem used in the coupling construction that proves the local Doeblin condition of Lemma 2.1.","marker":"[30]"},{"why":"Supplies the Curie-Weiss potential picture and the Eyring-Kramers exponential exit-time scale that motivate and frame the metastable conditioning.","marker":"[7]"},{"why":"A non-conservative Harris ergodic theorem that motivates the Doob h-transform route for studying the long-time behavior of the killed semi-group.","marker":"[2]"}],"fun_headline_variants":["Conditioned spins keep uniform propagation of chaos","Positive magnetization makes chaos uniform in time","Metastable spins: conditioned chaos holds for all time","Taming metastability with a positive magnetization condition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument leans on the fact that the potential well at $m_+$ is non-degenerate, meaning the potential curves upward quadratically at the bottom of the well; if the well were flat, the fluctuations of the conditioned magnetization would be wider than $n^{-1/2}$ and the small-set and Lyapunov mechanism would not close.","fun_headline_variants_meta":{"raw":{"variants":["Conditioned spins keep uniform propagation of chaos","Positive magnetization makes chaos uniform in time","Metastable spins: conditioned chaos holds for all time","Taming metastability with a positive magnetization condition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000611,"raw_usage":{"total_tokens":2872,"prompt_tokens":1001,"completion_tokens":1871,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":1813}},"tokens_in":617,"tokens_out":1871,"duration_ms":14113,"temperature":1.0,"reasoning_tokens":1813,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T15:26:54.633609+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the killed Glauber dynamics at $\\beta=2$ with $n$ up to $10^4$, start at $m_0=0.9$, and estimate $\\sup_{t\\ge 0}|E(m^n_t|\\tau_n>t)-m_t|$ for $f(x)=x$; if this quantity fails to decay as $n^{-\\alpha}$ for some $\\alpha>0$, Theorem 2 is false. A more structural test: modify the flip rates so the effective potential has a quartic minimum at $m_+$, i.e. $g(m)\\sim(m-m_+)^4$, and measure the stationary conditional fluctuations around $m_+$; if their standard deviation decays as $n^{-1/4}$ rather than $n^{-1/2}$, the non-degeneracy assumption that carries the proof has failed exactly as predicted.","supporting_citations":[{"cited_title":"Mattingly","cited_arxiv_id":null,"evidence_quote":"Supplies the Harris-type contraction theorem used as Proposition 3.1 to convert the Doeblin and Lyapunov conditions into the explicit exponential convergence of Theorem 1."},{"cited_title":"Quasi-stationary distributions and population processes","cited_arxiv_id":null,"evidence_quote":"Gives the quasi-stationary distribution survival identity $E_{\\nu_\\infty^n}(\\mathbf{1}_{\\tau_n>t})=e^{-b_n t}$, which is used to compare the QSD with $\\delta_{m_+}$ at polynomial rate."},{"cited_title":"Foundations of modern probability","cited_arxiv_id":null,"evidence_quote":"Provides the generator-convergence theorem used in the coupling construction that proves the local Doeblin condition of Lemma 2.1."},{"cited_title":"Metastability, volume 351 of Grundlehren der mathematis- chen Wissenschaften [Fundamental Principles of Mathematical Sciences]","cited_arxiv_id":null,"evidence_quote":"Supplies the Curie-Weiss potential picture and the Eyring-Kramers exponential exit-time scale that motivate and frame the metastable conditioning."},{"cited_title":"A non-conservative Harris er- godic theorem","cited_arxiv_id":null,"evidence_quote":"A non-conservative Harris ergodic theorem that motivates the Doob h-transform route for studying the long-time behavior of the killed semi-group."}],"review_version":1}