{"id":"0d868c55-9411-470b-af09-376b6063906f","arxiv_id":"2412.03521","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 parabolic Anderson model in dimension d≥3 with weak multiplicative noise, the solution converges in distribution to a limiting random field, extending earlier ergodicity results to broader noises and initial conditions.","lead":"A mathematics note reviews when stochastic PDEs settle into a steady random state, then proves a new case: the parabolic Anderson model in high dimension converges to a unique long-time limit when the noise is weak. This matters because it sharpens the known phase transition between ergodic behavior and blow-up in a canonical random heat equation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.14's proof substitutes time-one moment data J0(1,y) for the actual short-time moments J0(s+K+1,y) in the S_K estimate (5.56), invalid for rough initial data; this estimate controls the Gronwall closure in Step 7.","rationale":"The paper's central new claim is Theorem 5.14, a convergence-in-law result for the parabolic Anderson model under a weak disorder condition. The proof has one clearly identifiable load-bearing step: the estimate of S_K in Step 4. The reader identified exactly this weakness, and our independent reading confirms it. The substitution of time-one data for short-time moments is not a minor typo: because the initial measure is only rough (e.g., a Dirac mass), J0(t,·) can blow up as t↓0, so the bound (5.21) does not give a uniform-in-s control on the unit interval before the restart. Since this estimate is the sole source of decay for the term I_0^K in (5.58), the Gronwall argument in Step 7 collapses without it. We also noticed the minor index issue in the definition of ~u in Step 1, but that is easily fixed and does not affect the central difficulty. The appropriate resolution is conditional acceptance: the claim is plausible and likely repairable, but the present proof is incomplete. Our concrete test would settle whether the concern is fatal or merely a repairable technical gap.","tokens_in":36526,"tokens_out":14893,"duration_ms":143286,"concrete_test":"Re-derive (5.56) with ||u^*_K(s,y)||_2 (i.e., J0(s+K+1,y)) in place of ||u^*_K(-K,y)||_2. For μ=δ_0 and a Bessel-kernel correlation f in d=3 satisfying (5.9), compute the double integral near s=-K-1; if it is infinite or fails to vanish as K→∞, the proof's S_K estimate is false for an admissible initial condition. Alternatively, check whether a short-time second-moment bound E[u(t,x)^2]≤C J0(t,x)^2 with C uniform in t∈(0,1] exists; if not, the substitution in the displayed estimate cannot be repaired in the present form.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the proof of Theorem 5.14, Step 4 (equations (5.52)–(5.58)) estimates the stochastic term S_K(t,x) = ∫_{-K}^{-K-1}∫ p(t-s,x-y) b(u^*_K(s,y)) W(ds,dy). The displayed bound applies relation (5.6) plus Cauchy-Schwarz, and then invokes the second-moment bound (5.21). Since u^*_K satisfies (5.40) started at time -K-1, the correct moment bound at time s is ||u^*_K(s,y)||_2^2 ≤ J0^2(s+K+1,y)/(L_b^2(1-4L_b^2Υ(0))) with s+K+1∈(0,1). The paper instead uses ||u^*_K(-K,y)||_2 and J0(1,y;|µ|) uniformly in s. No monotonicity or domination justifies this for rough μ: for μ=δ_0, sup_y J0(s+K+1,y) ~ (s+K+1)^{-d/2}, which blows up as s→-K-1. This estimate feeds directly into (5.58), (5.63), and the Gronwall step (5.62); without a valid short-time bound, the claimed decay g_K→0 is unsupported. The issue is not cosmetic: for singular correlation measures allowed by (5.9), the integral with the true J0(s+K+1,y) may diverge near s=-K-1. A correct proof must either control short-time moments under (5.12) or restructure the restart so that only times ≥1 are used.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a survey of ergodic properties of stochastic PDEs, covering the classical Da Prato-Zabczyk framework, reflected stochastic heat equations, degenerate-noise Navier-Stokes equations, and a new result on the parabolic Anderson model in dimension d >= 3. The new contribution is Theorem 5.14, which asserts convergence in distribution in a weighted L^2 space to a limiting random field for solutions of the stochastic heat equation with degenerate coefficient b(u)=u, under a weak-disorder condition and for rough initial measures satisfying a homogeneous-evolution condition. The proof follows the Gu-Li restarting argument and relies on a generalized Gronwall lemma proved in Appendix A.","tokens_in":36801,"tokens_out":15762,"duration_ms":157877,"significance":"If Theorem 5.14 is correct, it is a meaningful extension of Gu and Li's result: it replaces the trace-class noise assumption with the broader condition 4 L_b^2 Upsilon(0) < 1, and it allows rough initial conditions well beyond L^1 cap L^infinity perturbations. The survey portions are accurate and useful, and the generalized Gronwall lemma in Appendix A is a self-contained contribution that may be of independent interest. The paper is clearly written and gives credit to the relevant literature, including the prior results of Chen-Kim and Chen-Eisenberg on which the proof relies.","major_comments":[{"comment":"The estimate leading to (5.56) is not justified. In (5.52), S_K(t,x) is a stochastic integral of b(u*_K(s,y)) over s in [-K-1,-K]. The displayed bound replaces ||u*_K(s,y)||_2 by ||u*_K(-K,y)||_2 for all s, and then uses (5.21) at time 1 to replace the resulting factor by J0(1,y;|mu|). However, u*_K is started at time -K-1 from mu, so (5.21) gives E|u*_K(s,y)|^2 <= J0^2(s+K+1,y)/(L_b^2(1-4L_b^2 Upsilon(0))) with s+K+1 in [0,1]. For rough initial data satisfying (5.12), sup_y J0(tau,y) is generally larger for tau<1 than at tau=1 (e.g., for mu=delta_0 it is (2 pi tau)^{-d/2}), so J0(1,y) is not an upper bound. Since (5.56) feeds into (5.57), (5.58), and through (5.62) into the Gronwall step, the claimed decay g_K -> 0, and hence (5.46), is not established. The proof needs either a valid short-time moment estimate under (5.12) or a restructuring that only uses moments at times at least one after the restart.","section":"§5.3, Step 4 (Eqs. (5.52)–(5.57))"},{"comment":"The proof of part (2) sets u = u1 - u2 and ~b(u) = b(u + u2) - b(u2), then states that the arguments in Steps 1–7 still apply. This is not immediate for two reasons. First, the two solutions must be constructed with the same driving noise for u1 - u2 to solve an SPDE; the statement of Theorem 5.14 does not specify this coupling. Second, ~b depends on the solution u2 through its time-space argument, so the coefficient is random and time-space dependent, whereas the moment bound (5.21) and the rest of the proof are stated for a deterministic coefficient b(u). A verification that the Gu-Li argument extends to this setting is missing. Without it, the uniqueness statement (5.33) is unsupported.","section":"§5.3, proof of item (2)"}],"minor_comments":[{"comment":"In condition (ii'), the statement says 'x in R' but the spatial variable should be x in R^d.","section":"§5.3, Theorem 5.14 statement"},{"comment":"The quantity defined in (5.22) as an integral is actually the square of the L^2_rho norm, not the norm itself; this should be clarified to avoid confusion in (5.24) and elsewhere.","section":"§5.1, Eq. (5.22)"},{"comment":"The argument that u(1,·) satisfies (5.12) a.s. from (5.39) is not fully justified: (5.39) shows finiteness of the heat extension of u(1,·), not directly the Gaussian-weighted total variation condition. Since the subsequent construction restarts from mu rather than from u(1,·), this step should be clarified or removed.","section":"§5.3, Step 1"},{"comment":"There are typos such as 'stochasti c' in the abstract and 'homegenous' in condition (ii'); these should be corrected.","section":"Abstract and Theorem 5.14"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the proof of Theorem 5.14, specifically the short-time moment estimate in Step 4 and the coupling/random-coefficient issue in part (2). These are load-bearing for the paper's new claim, but the gaps appear to be potentially repairable with additional technical work. The survey portions are solid and could be published independently if the new theorem cannot be fully fixed in the current format."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a well-written survey of ergodic SPDE results with one new theorem (Theorem 5.14). The new statement is worth having—it extends convergence-to-invariant-measure for the parabolic Anderson model in d≥3 to non-trace-class noises and rough initial data—but the proof as written has a load-bearing gap in Step 4. I would not yet trust the theorem.\n\nWhat's good: Sections 2–4 are reliable restatements of Da Prato–Zabczyk, the reflected SHE, and Hairer–Mattingly's degenerate Navier–Stokes result. The paper's own contribution is clearly framed as an extension of Gu–Li, and the appendix Gronwall lemma is a nice generalization. The heavy reliance on [12], [16], and [36] is not circular; those are independent published results.\n\nThe problem: In Step 4, S_K(t,x) integrates b(u^*_K(s,y)) over s∈[-K-1,-K]. The displayed bound uses ||u^*_K(-K,y)||_2 for every s, and then substitutes the time-one moment J0(1,y;|µ|). For rough initial measures, the correct factor is J0(s+K+1,y;|µ|), which can blow up as s↓-K-1 (e.g. µ=δ_0 gives (s+K+1)^(-d/2)). No monotonicity or domination justifies replacing it by time-one data. Equation (5.56) is therefore invalid, and since (5.56) feeds (5.57), (5.58), and the Gronwall forcing g_K in (5.63), the claimed decay of J(K+t) is unsupported. The same flaw propagates into the proof of part (2). This is not a cosmetic issue; for singular correlations allowed by (5.9), the integral with the true J0 may diverge.\n\nMinor: the definition of ~u around (5.43) seems off by a shift—u_K is only defined from -K onward, so u_K(t-K-1) is not available for t∈[0,1). This is easily fixed but should be clarified.\n\nIf Theorem 5.14 can be repaired—likely by a short-time moment estimate under (5.12) or by restarting after time 1 so only elapsed times ≥1 appear—the result would be a solid contribution. As it stands, I would not cite the theorem, but I would send the paper to a serious referee. The survey alone has value, and the new result is plausible enough to merit referee time rather than desk rejection.","headline":"A reliable survey plus a new convergence-to-invariant-measure theorem whose proof currently has a gap in Step 4; the result is plausible but not yet established.","tokens_in":37459,"tokens_out":3718,"would_cite":false,"duration_ms":36804,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H15","35R60","37L40","35K05"],"pacs":[],"model":"deepseek-v4-flash","headline":"For the parabolic Anderson model in d at least 3, the paper establishes that a weak-noise condition forces convergence in distribution to a limiting random field, whose law depends only on the initial condition's asymptotic homogeneous…","keywords":["ergodicity","invariant measure","parabolic Anderson model","stochastic heat equation","phase transition","rough initial conditions","weak disorder","degenerate noise"],"falsifier":"Compute $\\sup_{x\\in\\mathbb{R}^d}\\sup_{s\\in(0,1)} \\mathbb{E}[u(s,x)^2]/J_0(1,x;\\mu)^2$ for the parabolic Anderson model with $\\mu=\\delta_0$ under condition (5.9) in $d=3$; finiteness of this ratio is exactly what estimate (5.57) needs, and an unbounded ratio would falsify the proof’s control of the stochastic term $S_K$.","tokens_in":1694,"feed_emoji":"🌡️","tokens_out":2309,"duration_ms":106460,"temperature":0.7,"pith_summary":"This paper argues that the parabolic Anderson model — the multiplicative-noise stochastic heat equation — displays an ergodic phase transition in dimensions $d \\ge 3$, controlled by noise intensity. Below a threshold (the “weak disorder” condition $4L_b^2\\Upsilon(0)<1$), solutions converge in distribution, inside a weighted $L^2$ space, to a well-defined limiting random field; above threshold, moments blow up. The new Theorem 5.14 extends convergence to broad Gaussian noises satisfying condition (5.9) and to rough initial conditions (measures, not just functions), beyond earlier settings that required trace-class noise or $L^1 \\cap L^\\infty$ perturbations. The paper also surveys the classical infinite-dimensional ergodic framework, reflected equations, and degenerate-noise Navier–Stokes results, but the phase-transition result is the genuinely new contribution.","feed_headline":"In d≥3, weak-noise Anderson model converges to a limit law","feed_subtitle":"When noise intensity stays below a spectral threshold, solutions settle into a limiting law; above it, moments explode.","key_machinery":"The argument uses the negative-time restart construction: the solution is started at time $-K-1$, evolved one unit, then restarted at $-K$, so that large-time questions about the original system become large-$K$ questions about a sequence of shifted systems driven by the same two-sided noise. The contraction estimate splits the difference between two solutions into terms $I_0, I_1, I_2$; the stochastic terms are controlled by the second-moment bound (5.21), by the weighted space $L^2_\\rho(\\mathbb{R}^d)$, by the spectral decay function $H(t):=\\int_t^\\infty k(s)\\,ds$, and by a new Gronwall-type lemma (Appendix A) that lets the iteration close once $4L_b^2\\Upsilon(0)<1$.","core_discovery":"The central discovery is that, in dimension $d \\ge 3$, the parabolic Anderson model driven by spatially homogeneous Gaussian white-in-time noise has an invariant-measure regime determined by the product of the diffusion coefficient’s Lipschitz constant $L_b$ and the noise spectral mass at zero $\\Upsilon(0)$. Under $4L_b^2\\Upsilon(0)<1$, for any rough initial measure $\\mu$ whose free-heat evolution $J_0(t,\\cdot;\\mu)$ has finite long-time supremum, the random field $u(t,\\cdot)$ converges in distribution in $L^2_\\rho(\\mathbb{R}^d)$ to a limiting field $Z$; and if two initial measures have the same asymptotic homogeneous evolution, then the limits share the same law.","pith_inferences":["The equivalence classes of initial data defined by (5.34) suggest a natural ergodic state space: the limiting random field is a functional of the asymptotic homogeneous evolution, generalizing the flat-initial-condition stationary field to nonstationary settings.","The weak-disorder threshold is the same type of condition that governs central limit theorems for directed polymers in random environments; a testable extension is that polymer measures in these general Gaussian environments should exhibit diffusive fluctuations exactly when $4L_b^2\\Upsilon(0)<1$.","A numerical check of the short-time moment ratio for Dirac-delta initial data in $d=3$ with a Bessel-kernel noise would directly probe the proof’s load-bearing estimate and could indicate whether the convergence statement requires an additional short-time moment hypothesis."],"forward_implications":["In $d \\ge 3$, the law of the multiplicative-noise heat equation has a well-defined large-time limit for every rough initial measure satisfying (5.31), establishing ergodicity for a broad class of non-trace-class noises.","Two initial data with the same asymptotic homogeneous evolution produce the same limiting random field, so the limit law depends only on an equivalence class of the initial condition.","The threshold $4L_b^2\\Upsilon(0)<1$ separates the ergodic regime from the intermittency regime where second moments grow exponentially, making the phase transition quantitative.","Convergence holds in $L^2_\\rho(\\mathbb{R}^d)$ for any nonnegative integrable weight $\\rho$, not only the admissible weights of the prior framework.","Under the strengthened Dalang condition (5.29), the paper notes that the same convergence can be upgraded to weighted Hölder spaces, giving a spatially regular limit field."],"supporting_citations":[{"why":"Supplies the negative-time restart construction and the contraction framework that Theorem 5.14 adapts to broader noises and initial data.","marker":"[36]"},{"why":"Provides conditions (5.9) and (5.10), the invariant-measure Theorem 5.10, and kernel examples distinguishing trace-class from admissible noises.","marker":"[12]"},{"why":"Gives the second-moment bound (5.21) under $4L_b^2\\Upsilon(0)<1$ for rough initial data, used in Steps 4-6.","marker":"[16]"},{"why":"Sets the infinite-dimensional ergodic framework, Krylov–Bogoliubov tightness, and invariant-measure criteria that organize the review sections.","marker":"[21]"},{"why":"Defines the Walsh martingale-measure stochastic integral and Dalang’s condition (5.5) under which equation (5.1) is solved.","marker":"[24]"},{"why":"Introduces admissible weight functions and the weighted spaces $L^2_\\rho(\\mathbb{R}^d)$ used for the convergence statement.","marker":"[50]"}],"fun_headline_variants":["d≥3 Anderson: subcritical noise locks in a limit law","Weak-noise Anderson model in d≥3 settles to a limit law","Subcritical noise gives Anderson d≥3 an invariant law","d≥3 Anderson: below-threshold noise converges to invariant law"],"cache_read_input_tokens":39296,"weakest_assumption_plain":"The proof rests on the assumption that the second-moment bound derived from time-one data also controls the solution during the first unit of time from a rough initial measure; if short-time moments can instead explode, the contraction estimate (5.57) would not follow.","fun_headline_variants_meta":{"raw":{"variants":["d≥3 Anderson: subcritical noise locks in a limit law","Weak-noise Anderson model in d≥3 settles to a limit law","Subcritical noise gives Anderson d≥3 an invariant law","d≥3 Anderson: below-threshold noise converges to invariant law"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000563,"raw_usage":{"total_tokens":2611,"prompt_tokens":825,"completion_tokens":1786,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":441,"completion_tokens_details":{"reasoning_tokens":1713}},"tokens_in":441,"tokens_out":1786,"duration_ms":12570,"temperature":1.0,"reasoning_tokens":1713,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:19:50.192099+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $\\sup_{x\\in\\mathbb{R}^d}\\sup_{s\\in(0,1)} \\mathbb{E}[u(s,x)^2]/J_0(1,x;\\mu)^2$ for the parabolic Anderson model with $\\mu=\\delta_0$ under condition (5.9) in $d=3$; finiteness of this ratio is exactly what estimate (5.57) needs, and an unbounded ratio would falsify the proof’s control of the stochastic term $S_K$.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides conditions (5.9) and (5.10), the invariant-measure Theorem 5.10, and kernel examples distinguishing trace-class from admissible noises."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the second-moment bound (5.21) under $4L_b^2\\Upsilon(0)<1$ for rough initial data, used in Steps 4-6."},{"cited_title":"Da Prato and J","cited_arxiv_id":null,"evidence_quote":"Sets the infinite-dimensional ergodic framework, Krylov–Bogoliubov tightness, and invariant-measure criteria that organize the review sections."},{"cited_title":"Dalang, Extending the martingale measure stochastic integral with a pplica- tions to spatially homogeneous s.p.d.e.’s , Electron","cited_arxiv_id":null,"evidence_quote":"Defines the Walsh martingale-measure stochastic integral and Dalang’s condition (5.5) under which equation (5.1) is solved."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces admissible weight functions and the weighted spaces $L^2_\\rho(\\mathbb{R}^d)$ used for the convergence statement."}],"review_version":1}