{"id":"a6eb975a-82b7-48e5-bdb6-d60b10c59480","arxiv_id":"1908.04145","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A central limit theorem holds for high-frequency power variations of the stochastic heat equation with multiplicative noise and Riesz spatial covariance of order alpha in (0,1), with no asymptotic bias.","lead":"This paper proves a central limit theorem for power variations of the solution to the stochastic heat equation driven by multiplicative Gaussian noise. It shows that no asymptotic bias appears when the spatial correlation is a Riesz kernel of order alpha in (0,1), due to cancellation of error terms.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proof of Lemma 3.3 uses a uniform initial-increment bound that H4 does not imply; the term is still negligible by total variation, so a repair is needed.","rationale":"The reader flagged H4 as the weakest assumption because the initial-condition contribution must be o(√Δ_n). Close inspection shows the proof's specific uniform estimate at that point is not implied by H4: heat-kernel smoothing gives only Δ_n(iΔ_n)^((γ-1)/2), which, for the small truncation exponent a required later, is not uniformly o(√Δ_n τ_n). A concrete example with compactly supported u0 whose leading singularity is |x|^{1+γ} realizes this bound. However, the same calculation shows the summed contribution is controlled by the total variation of u^(0), which is finite under H4, so the theorem's claim survives. Thus the main risk is not the mathematical truth of Theorem 2.2 but the rigor of the proof as written; a short repair is available. This is why the verdict should be CONDITIONAL rather than REJECT, and why I only partially agree with the reader's identification of H4 as the load-bearing weak spot.","tokens_in":61266,"tokens_out":26884,"duration_ms":251411,"concrete_test":"Re-derive the bound for A_n^1 in Lemma 3.3 without the uniform per-increment estimate: use E[(A_n^1)^*_T] ≤ √Δ_n/τ_n Σ_i |u^(0)(iΔ_n,x)-u^(0)((i-1)Δ_n,x)| ≤ √Δ_n/τ_n TV(u^(0)(·,x); [λ_nΔ_n,T]). With τ_n ≍ Δ_n^{1/2-α/4} and TV finite under H4, this is O(Δ_n^{α/4}) → 0. If this alternative bound is verified, the proof gap is closed and Theorem 2.2 stands; if the total variation were infinite, the CLT would fail.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In Lemma 3.3, the term A_n^1 is dismissed via [5, Remark 2.7] as |u^(0)(iΔ_n,x) - u^(0)((i-1)Δ_n,x)|/τ_n = o(√Δ_n) uniformly in i ≥ λ_n. For u0 satisfying H4 with ∇u0 ∈ C^γ, the heat-kernel bound gives |Δ_i u^(0)| = O(Δ_n (iΔ_n)^((γ-1)/2)). With λ_n = Δ_n^{-a} and a near 1/(2+α), the uniform ratio divided by √Δ_n behaves like Δ_n^{α/4 + (1-a)(γ-1)/2}, which is not o(1) when γ is merely > 1/2 - α/4; for example, α=1/2, γ=0.38, a≈0.4 gives an exponent of about -0.06. Thus the uniform estimate is not a consequence of H4. However, the A_n^1 sum itself is bounded by √Δ_n/τ_n times the total variation of t↦u^(0)(t,x) on [λ_nΔ_n,T]; H4 makes this total variation finite, so the term is O(Δ_n^{α/4}) → 0. Hence the central theorem is not threatened, but the proof as written has a real gap that must be repaired by replacing the uniform estimate with a summation/total-variation argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the high-frequency behavior of normalized variation functionals of the solution to a parabolic stochastic PDE with multiplicative Gaussian noise that is white in time and spatially homogeneous with Riesz-kernel covariance of order α∈(0,1). After recalling a law of large numbers from earlier work, the main result (Theorem 2.2) states a stable central limit theorem for these functionals at the usual √Δ_n rate, with a centered mixed Gaussian limit whose conditional covariance is given explicitly in terms of the evaluation function f and the covariance constants Γ_r. The proof decomposes the variation functional into a martingale part and several error terms, then shows that the apparently non-negligible error terms cancel in the limit; this cancellation is isolated in Proposition 3.23. The paper also discusses why the same statement fails for α=1 in a companion work, making the positive result for α∈(0,1) surprising.","tokens_in":61538,"tokens_out":8444,"duration_ms":91402,"significance":"If the theorem is correct, it is a substantial contribution to the statistical theory of SPDEs: it extends second-order limit theorems for power variations from additive or smooth-coefficient cases to the multiplicative case, where the coefficient σ(u) has only the low regularity inherited from the solution. The result is non-trivial because the coefficient is not 1/2-Hölder in time, and the paper identifies an interesting cancellation mechanism that removes the bias one might expect. The normalization τ_n and the covariance constants Γ_r are derived explicitly from the Riesz kernel, and the covariance formula in (2.15) is parameter-free. The proof is detailed and contains a direct verification of the decisive cancellation in Proposition 3.23. The sharp contrast with the α=1 companion paper [7] makes the positive result credible. The paper should be of interest to researchers in SPDE inference and high-frequency statistics.","major_comments":[{"comment":"The proof of Lemma 3.3 contains a gap in the treatment of the term A_n^1. It is asserted that hypothesis H4 implies the uniform estimate |u^(0)(iΔ_n,x)-u^(0)((i-1)Δ_n,x)|/τ_n = o(√Δ_n) for all i≥λ_n, referring to [5, Remark 2.7]. This uniform bound is not a consequence of H4. If ∇u_0 is Hölder continuous with exponent γ>1/2−α/4, semigroup smoothing gives |u^(0)(t,x)-u^(0)(t−Δ_n,x)| = O(Δ_n t^{(γ−1)/2}) for t≥Δ_n. For t in the summation range, the worst case is t≈λ_nΔ_n=Δ_n^{1−a}, which yields |u^(0)(iΔ_n,x)-u^(0)((i−1)Δ_n,x)|/(τ_n√Δ_n) = O(Δ_n^{α/4+(1−a)(γ−1)/2}). This exponent can be negative under the assumptions of H4; for example, α=1/2, γ=0.38, and a≈0.4 give an exponent of about −0.06. The gap is repairable: the sum defining A_n^1 can be bounded by (√Δ_n/τ_n) times the total variation of t↦u^(0)(t,x) on [λ_nΔ_n,T], which is finite under H4, so A_n^1=O(Δ_n^{α/4})→0. However, the proof as written relies on a false uniform estimate and must be corrected.","section":"Section 4, proof of Lemma 3.3, treatment of A_n^1"}],"minor_comments":[{"comment":"The core martingale approximation in Lemma 3.4 is not proved self-containedly: the decomposition C^{n,m}=C_1^{n,m}+C_2^{n,m} is obtained by reference to [5, Eq. (3.11)] and [6, Eq. (D.8)] with the comment that the proofs are identical and do not use Hölder regularity of σ(u). Since this lemma supplies the limiting Gaussian process, the manuscript should either reproduce the argument or state the relevant results precisely enough that a reader can verify the claimed independence from the missing regularity assumptions.","section":"Section 3.1, Lemma 3.4"},{"comment":"The notation a and a (with and without underline) in (3.5) and subsequent lemmas is extremely easy to confuse, especially in print. Renaming one of the two exponents, e.g., using b for the underlined quantity, would considerably improve readability.","section":"Throughout Section 3"},{"comment":"The theorem states that the infinite series in the covariance formula converge in L1, but I did not find an explicit proof of this convergence in the text. A short justification or a precise reference to where it is established would be helpful.","section":"Theorem 2.2, Eq. (2.15)"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically impressive and the central cancellation mechanism is convincing, but the gap in Lemma 3.3 must be fixed before publication. I believe the fix is local and does not threaten the main theorem. The manuscript relies heavily on the author's companion works [5] and [6]; the editor may wish to confirm that these are publicly available and that the quoted results are exactly the ones needed. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper proves a real CLT — for power variations of the stochastic heat equation with multiplicative noise and Riesz spatial covariance alpha in (0,1), the normalized variation converges stably to a mixed Gaussian process, and there is no asymptotic bias. The novelty is genuine: prior CLTs needed sigma(u) to be 1/2-Holder in time and C^2 in space, which fails here. The no-bias result is traced to cancellation among four error terms (B_2, D_2, D_3, H), and Proposition 3.23 — the cancellation itself — is proved with explicit estimates rather than hand-waved. The covariance formula is natural and the statistical motivation (confidence intervals for volatility in the parabolic Anderson model) is well explained.\n\nWhat to be careful about. The paper leans heavily on two companion papers [5] and [6] for several technical lemmas, including Lemma 3.4 and the differentiability of mu_f. That is a self-containedness problem, not a correctness problem, but it makes independent verification harder. The more specific issue is in the proof of Lemma 3.3. The term A_n^1 is dismissed using [5, Remark 2.7] with a uniform bound on |u^(0)(i Delta_n,x)-u^(0)((i-1) Delta_n,x)|/tau_n = o(sqrt(Delta_n)). Under H4 alone, that uniform bound is not justified: for t near lambda_n Delta_n, the increment behaves like Delta_n t^{(gamma-1)/2}, and after dividing by tau_n sqrt(Delta_n) the exponent is alpha/4 + (1-a)(gamma-1)/2, which can be negative when gamma > 1/2 - alpha/4 is chosen close to the bound. The stress-test note is right about that. It is also right that the damage is repairable: the sum of those increments is bounded by the total variation of u^(0)(.,x) on [lambda_n Delta_n,T], which H4 makes finite, and the resulting estimate is O(Delta_n^{alpha/4+(1-a)(1+gamma)/2}) -> 0. So the theorem still stands, but the proof as written contains a real gap that ought to be fixed before publication.\n\nBottom line: worth sending to a serious referee. The main result is new, the cancellation mechanism is enlightening, and the flawed step is localized and fixable. I'd cite this in my own work and bring it to the reading group, but I'd ask the author to patch Lemma 3.3 and to think about making the companion-paper dependency less opaque.","headline":"Genuinely new CLT for multiplicative-noise SPDEs with a surprising no-bias cancellation; proof has one repairable gap in Lemma 3.3.","tokens_in":62065,"tokens_out":5020,"would_cite":true,"duration_ms":46580,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H15","60F05","62M40","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that for the stochastic heat equation driven by multiplicative Gaussian noise with Riesz spatial correlation of order alpha in (0,1), normalized power variations satisfy a stable central limit theorem with no asymptotic…","keywords":["power variations","stochastic heat equation","multiplicative noise","Riesz kernel","central limit theorem","parabolic Anderson model","parameter estimation","mixed Gaussian process"],"falsifier":"Run a high-frequency Monte Carlo simulation of the parabolic Anderson model with $\\alpha=\\tfrac12$, $\\sigma(x)=\\sigma_0 x$, and a known smooth $u_0$, and compute the sample version of $\\Delta_n^{-1/2}(V_4^n(u,t)-V_4(u,t))$; if its distribution does not approach the zero-mean mixed Gaussian with covariance (2.15) as $\\Delta_n\\to0$, the no-bias claim fails. A second check is to take $u_0$ with derivative Hölder exponent exactly at $\\frac12-\\frac{\\alpha}{4}$ and observe whether a bias of order $\\Delta_n^{-1/2}$ appears.","tokens_in":61037,"feed_emoji":"📈","tokens_out":8363,"duration_ms":82261,"temperature":0.7,"pith_summary":"The paper proves a central limit theorem (CLT) for high-frequency power variations and related functionals of the solution to the stochastic heat equation with multiplicative Gaussian noise. Earlier CLTs for such functionals needed the noise coefficient to be almost 1/2-Hölder in time and twice differentiable in space, assumptions that the coefficient $\\sigma(u)$ here does not satisfy. The paper shows that when the spatial correlation is a Riesz kernel of order $\\alpha\\in(0,1)$, the normalized variation converges stably to a continuous mixed Gaussian process with an explicit covariance and with no asymptotic bias. The result matters because second-order limits of this kind are the step that turns consistent estimators of volatility-type parameters into tools for confidence intervals and tests. The absence of bias is explained by cancellations among error terms that depend crucially on $\\sigma$ being evaluated along the solution itself.","feed_headline":"Power variations of SPDE solutions obey a zero-bias CLT","feed_subtitle":"Normalized power-variation fluctuations converge to a mixed Gaussian when spatial noise correlation has order below 1.","key_machinery":"The central object is the decomposition of the normalized variation into a martingale array plus four 'bad' terms $B_2^{n,m}$, $D_2^n$, $D_3^n$, and $H^n$ that individually do not vanish at the $\\sqrt{\\Delta_n}$ rate. The load-bearing mechanism is Proposition 3.23, which shows that certain simplified versions of these bad terms cancel: $B_2^{n,8}+D_2^{n,6}+D_3^{n,6}+H^n\\xrightarrow{L^1}0$. The cancellations use the special algebraic fact that $\\sigma(u(s,y))$ is a function of the solution $u$ itself, not an arbitrary rough random field; this creates identities between Taylor expansions of $f$, of $\\sigma$, and of the solution increments. The proof is carried by two quantitative tools: 'standard size estimates' for stochastic integrals against the rough coefficient and 'martingale size estimates' that exploit conditional independence, together with Wiener-chaos orthogonality to kill terms of wrong parity. A block-splitting argument isolates conditionally independent pieces so that Jacod's martingale CLT applies.","core_discovery":"Under hypotheses H1–H4, for each $\\alpha\\in(0,1)$, the normalized variation functional satisfies $\\frac{1}{\\sqrt{\\Delta_n}}(V_f^n(u,t)-V_f(u,t))\\xrightarrow{\\text{st}}Z$, where $Z$ is a continuous process that, conditionally on the original $\\sigma$-field, is centered Gaussian with independent increments and covariance matrix (2.15). In words: the fluctuation of the power variation around its law-of-large-numbers limit is asymptotically Gaussian at the usual $\\sqrt{\\Delta_n}$ rate, with zero mean. This holds although $\\sigma(u(t,x))$ is only $(1/2-\\alpha/4-\\epsilon)$-Hölder in time and $(1-\\alpha/2-\\epsilon)$-Hölder in space, i.e. far below the regularity previously required. The companion analysis at $\\alpha=1$ shows the positive result is sharp: when the noise is space-time white, a genuine asymptotic bias appears for $p\\ge4$, while $p=2$ remains unbiased.","pith_inferences":["The sharp change at $\\alpha=1$ suggests a phase transition in the statistical theory: estimators based on fourth or higher power variations should behave differently in the white-noise case than for any $\\alpha<1$, and this distinction should be detectable in simulation studies.","The cancellation mechanism likely transfers to other parabolic SPDEs with multiplicative noise, provided the coefficient is a smooth function of the solution and the solution has enough moments; testing this on stochastic reaction-diffusion equations would be a natural next step.","If the open conjecture for $\\alpha\\in(1,2)$ in dimension $d\\ge2$ is correct, the $\\sqrt{\\Delta_n}$ rate itself would fail there, so the present $\\alpha\\in(0,1)$ window may be the only one with standard CLT behavior; this is an explicit, testable boundary."],"forward_implications":["For the parabolic Anderson model, the CLT turns the consistent estimator of $\\sigma_0^p$ into an asymptotically normal estimator, enabling confidence intervals and tests from high-frequency observations at one spatial point.","For $\\alpha\\in(0,1)$, no bias correction is needed in the CLT for any $p>0$; the companion result at $\\alpha=1$ shows bias correction is required for $p\\ge4$ when the noise is space-time white.","The covariance formula (2.15) gives explicit asymptotic variances for power, multipower, and signed multipower variations, so it can be used to choose among functionals for efficiency.","The CLT applies to power and multipower variations at a fixed spatial point; functionals that mix different spatial points within one coordinate are excluded by hypothesis H2 and are not covered by this theorem."],"supporting_citations":[{"why":"Supplies the law of large numbers for the variation functionals and the high-regularity CLT framework whose hypotheses are relaxed here; its block-splitting and martingale-CLT structure is reused.","marker":"[5]"},{"why":"Companion paper showing that at alpha=1 the CLT acquires an asymptotic bias for p>=4, which makes the alpha in (0,1) no-bias theorem sharp.","marker":"[7]"},{"why":"Provides existence, uniqueness, and moment bounds for the mild solution of the stochastic heat equation with spatially homogeneous noise.","marker":"[12]"},{"why":"Supplies the Holder-continuity estimates for the solution that quantify the roughness of sigma(u) and drive the size estimates throughout the proof.","marker":"[24]"},{"why":"Jacod's martingale central limit theorem is the engine for stable convergence in law of the approximating martingale arrays.","marker":"[17]"},{"why":"Supplement containing the technical estimates for differentiability of mu_f, the Pi^n measure estimates, and earlier variants of the approximation steps.","marker":"[6]"}],"fun_headline_variants":["Zero-bias CLT for SPDE power variations","Unexpected zero-bias CLT for SPDE power variations","No bias in SPDE power variation CLT","SPDE power variations obey zero-bias Gaussian limit","Low-regularity SPDE variations still give zero-bias CLT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is hypothesis H4, that the initial condition $u_0$ is bounded and differentiable with derivative Hölder continuous of exponent larger than $\\frac12-\\frac{\\alpha}{4}$, together with $\\alpha\\in(0,1)$; if $u_0$ is rougher, the deterministic contribution to the increments need not be $o(\\sqrt{\\Delta_n})$, and at $\\alpha=1$ an additional bias appears for $p\\ge4$.","fun_headline_variants_meta":{"raw":{"variants":["Zero-bias CLT for SPDE power variations","Unexpected zero-bias CLT for SPDE power variations","No bias in SPDE power variation CLT","SPDE power variations obey zero-bias Gaussian limit","Low-regularity SPDE variations still give zero-bias CLT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000258,"raw_usage":{"total_tokens":1532,"prompt_tokens":842,"completion_tokens":690,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":609}},"tokens_in":458,"tokens_out":690,"duration_ms":6865,"temperature":1.0,"reasoning_tokens":609,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:49:39.338775+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a high-frequency Monte Carlo simulation of the parabolic Anderson model with $\\alpha=\\tfrac12$, $\\sigma(x)=\\sigma_0 x$, and a known smooth $u_0$, and compute the sample version of $\\Delta_n^{-1/2}(V_4^n(u,t)-V_4(u,t))$; if its distribution does not approach the zero-mean mixed Gaussian with covariance (2.15) as $\\Delta_n\\to0$, the no-bias claim fails. A second check is to take $u_0$ with derivative Hölder exponent exactly at $\\frac12-\\frac{\\alpha}{4}$ and observe whether a bias of order $\\Delta_n^{-1/2}$ appears.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the law of large numbers for the variation functionals and the high-regularity CLT framework whose hypotheses are relaxed here; its block-splitting and martingale-CLT structure is reused."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Companion paper showing that at alpha=1 the CLT acquires an asymptotic bias for p>=4, which makes the alpha in (0,1) no-bias theorem sharp."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides existence, uniqueness, and moment bounds for the mild solution of the stochastic heat equation with spatially homogeneous noise."},{"cited_title":"Sanz-Solé and M","cited_arxiv_id":null,"evidence_quote":"Supplies the Holder-continuity estimates for the solution that quantify the roughness of sigma(u) and drive the size estimates throughout the proof."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Jacod's martingale central limit theorem is the engine for stable convergence in law of the approximating martingale arrays."},{"cited_title":"High-frequency analysis of parabolic stochastic PDEs","cited_arxiv_id":"1806.06959","evidence_quote":"Supplement containing the technical estimates for differentiability of mu_f, the Pi^n measure estimates, and earlier variants of the approximation steps."}],"review_version":1}