{"id":"39ddd8f3-b1fe-470a-b6d1-135ac7f2ca8e","arxiv_id":"1908.03666","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For time-fractional diffusion with a fractional Brownian motion source and alpha plus H greater than 1, final-time mean and covariance uniquely determine f and |g|, but reconstruction is unstable.","lead":"Researchers prove that a time-fractional diffusion equation with a source driven by fractional Brownian motion has a well-posed direct solution when the fractional order alpha and the Hurst index H satisfy alpha plus H greater than 1, and that final-time mean and covariance data determine the deterministic spatial parts of the source uniquely.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 4.2's H<1/2 lower-bound proof relies on a false monotonicity claim about phi'_k, so the written proof of Theorem 4.3 is unsupported in that regime; a repair may exist.","rationale":"The reader's weakest assumption identifies the same load-bearing concern I find: Lemma 4.2 is the only step that turns nonzero covariance data into recovery of g_k g_l for H < 1/2, and its proof relies on a demonstrably false monotonicity assertion about phi'_k. The failure of the inequality phi'_k(s) >= phi'_k(T) is not cosmetic, because the four-term lower bound in Lemma 4.2 invokes that exact inequality for the mixed terms. Thus the proof of Theorem 4.3 is incomplete as written for 0 < H < 1/2. I agree with the reader's conditional disposition rather than acceptance or rejection, because the lemma's conclusion may still be true: if phi'_k is increasing, the minimum occurs at s = 0, and phi'_k(0) > 0 would give the needed lower bound. The paper would need to supply that corrected argument and verify all four inner products remain positive. Secondary issues, such as the missing gamma factors in the beta formula of Theorem 4.4 and the omitted proof of Theorem 3.2, do not change this assessment, since the central uniqueness claim is already conditional on the Lemma 4.2 repair.","tokens_in":25995,"tokens_out":17546,"duration_ms":165095,"concrete_test":"Analytically re-derive Lemma 4.2 with alpha = 1, H = 1/4 and any fixed k: compute phi'_k(s) at s = 0 and s = T, obtaining phi'_k(s) = lambda_k exp(-lambda_k(T-s)), so phi'_k(0) < phi'_k(T), which directly falsifies the monotonicity premise. Then re-run the lower-bound argument using the correct inequality phi'_k(s) >= phi'_k(0) (and similarly for phi_l), keeping K_H(T,s) > 0, M_H(s) > 0, phi_k(s) >= phi_k(0) > 0; if all four inner products are nonnegative and at least one is positive, Lemma 4.2 survives and Theorem 4.3 stands, otherwise the uniqueness proof fails for H < 1/2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Lemma 4.2, for H in (0,1/2), the proof claims that phi_k(s) = (T-s)^(alpha-1) E_{alpha,alpha}(-lambda_k (T-s)^alpha) has phi'_k > 0 and phi'_k monotonically decreasing, hence phi'_k(s) >= phi'_k(T) > 0. This monotonicity statement is false. For alpha = 1, phi_k(s) = exp(-lambda_k(T-s)), so phi'_k(s) = lambda_k exp(-lambda_k(T-s)) is strictly increasing and phi'_k(s) < phi'_k(T) for s < T, the opposite of the asserted inequality. For alpha < 1, phi'_k(s) = -(T-s)^(alpha-2) E_{alpha,alpha-1}(-lambda_k(T-s)^alpha), which tends to +infinity as s approaches T, so phi'_k(T) is not a finite lower bound and phi'_k(s) >= phi'_k(T) cannot hold. The subsequent four-term decomposition of the covariance inner product I_kl uses phi'_k(T) and phi'_l(T) to make the mixed terms positive; without this, the written proof does not establish C_2 > 0. Since Theorem 4.3 uses Lemma 4.2 to recover g_k g_l from covariance data, the uniqueness proof is incomplete for H < 1/2. The gap may be reparable: phi'_k appears to be increasing, so phi'_k(s) >= phi'_k(0) > 0 would supply the needed lower bound if the four terms are rechecked, but that correction is not present in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers the initial-boundary value problem (2.1) for the time-fractional diffusion equation with a random source f(x)h(t)+g(x)\\dot B_H(t), where B_H is a fractional Brownian motion with Hurst index H. It proves well-posedness of a mild solution when 0<α≤1, 01, with the a priori estimate in Theorem 3.1. The inverse problem uses the expectation and covariance of final-time data u(x,T) to recover f and |g|; uniqueness is claimed in Theorem 4.3 via the lower bounds in Lemmas 4.1 and 4.2, and instability is characterized in Theorem 4.4. The main technical ingredients are Mittag-Leffler function estimates and the stochastic integral representation of fractional Brownian motion.","tokens_in":26220,"tokens_out":13041,"duration_ms":115430,"significance":"If the uniqueness proof is repaired, the paper would be a useful contribution to the inverse random source literature for fractional diffusion, extending the Brownian-motion results of [29] to fractional Brownian motion under the natural condition α+H>1. The direct problem estimates are detailed, the instability exponents are explicit, and the paper is generally self-contained in its use of the fBm stochastic calculus. However, the key lower bound for H∈(0,1/2) in Lemma 4.2 rests on a monotonicity claim about the derivative of the Mittag-Leffler kernel that is false for α=1 and unproved for α<1; this gap is load-bearing for Theorem 4.3. The remaining issues are local and appear reparable.","major_comments":[{"comment":"The proof asserts that φ_k(s)=(T-s)^{α-1}E_{α,α}(-λ_k(T-s)^α) has φ'_k(s)>0 and φ'_k monotonically decreasing, leading to φ'_k(s)≥φ'_k(T)>0. This is false for α=1, where φ_k(s)=exp(-λ_k(T-s)) and φ'_k(s)=λ_k exp(-λ_k(T-s)) is strictly increasing in s. For 0<α<1, φ'_k(s)=-(T-s)^{α-2}E_{α,α-1}(-λ_k(T-s)^α) tends to +∞ as s↑T because E_{α,α-1}(0)=1/Γ(α-1)<0, so φ'_k(T) is not a finite lower bound and the monotone-decrease statement cannot hold in the form used. Since this lower bound is used to make all four inner-product terms in I_kl positive, Lemma 4.2 is not established for H<1/2, and Theorem 4.3, which relies on this lemma to recover g_k g_l from the covariance data, is unsupported in that regime. A repair may be possible using φ'_k(s)≥φ'_k(0)>0 after proving monotone increase, but the present proof is incomplete.","section":"Section 4, Lemma 4.2, H∈(0,1/2) case"}],"minor_comments":[{"comment":"Substituting t_*=λ_k^{-γ} in (4.13) gives the term t_*^{2H}=λ_k^{-2γH}, so the third entry in the minimum for β should be 2γH rather than 2H; please correct the displayed formula.","section":"Theorem 4.4, case 0 1 as the dividing line. It also answers the natural Brownian-motion question negatively for alpha <= 1/2. The direct problem estimates are mostly coherent, and the uniqueness strategy—positivity of the Mittag-Leffler kernel plus fBm covariance—is a natural extension of the Brownian case.\n\nThe soft spot is exactly where the reader flagged. Lemma 4.2 for H in (0,1/2) asserts that phi'_k is monotonically decreasing and uses phi'_k(s) >= phi'_k(T) > 0 as a lower bound. That is false for alpha = 1, where phi'_k is increasing; for alpha < 1 the claim is not established by the cited lemmas and the argument appears to go the other way. This is load-bearing because it converts covariance data into positivity of g_k g_l. The good news is that the gap is likely reparable: in fact phi'_k appears to be increasing, so phi'_k(s) >= phi'_k(0) > 0 would supply the lower bound, provided the four-term inner-product decomposition is rechecked with that bound.\n\nThe other problems are minor but real. Theorem 4.4's beta for H < 1/2 lists a term 2H; substituting t_* = lambda^{-gamma} into (4.13) gives 2 gamma H. The proof of Theorem 3.2 is deferred with \"details omitted,\" and some series/integral interchanges are not justified. None of these are fatal; they are the kind of thing a serious referee can ask to be fixed.\n\nIf the H < 1/2 uniqueness proof is repaired, the paper delivers what it promises: a clean parameter condition and a clear statement of uniqueness and instability for a niche but legitimate class of stochastic inverse problems. The citation pattern is appropriate; there are no fitted constants or self-citations to inflate the result.\n\nRecommendation: send it to peer review. It should not be desk-rejected. It needs a competent referee to push on Lemma 4.2 and the beta formula, but the core idea is sound and the contribution is new. If I were refereeing, I would ask for the fix and then accept.","headline":"First fBm extension for this inverse source problem, with a real but likely fixable gap in the H<1/2 uniqueness proof; send to peer review.","tokens_in":26895,"tokens_out":8910,"would_cite":true,"duration_ms":79863,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35R30","35R60","65M32"],"pacs":[],"model":"deepseek-v4-flash","headline":"Final-time statistics uniquely determine a fractional diffusion random source.","keywords":["time-fractional diffusion equation","inverse random source problem","fractional Brownian motion","Mittag-Leffler function","uniqueness","ill-posedness","Hurst index","mild solution"],"falsifier":"Compute numerically, or analytically, the sign and minimum of $\\varphi_k'(s)$ on $[0,T]$ for one eigenvalue $\\lambda_k$ with $\\alpha=0.9$ and $H=0.3$; if $\\inf_s\\varphi_k'(s)\\le0$ or if the covariance integral $I_{kk}$ in Lemma 4.2 is zero or negative, the uniqueness proof collapses for that parameter range. A simpler probe is $\\alpha=1$, where $\\varphi_k(s)=e^{-\\lambda_k(T-s)}$ has $\\varphi_k'$ increasing, so the asserted lower bound $\\varphi_k'(s)\\ge\\varphi_k'(T)$ is reversed.","tokens_in":25658,"feed_emoji":"🎲","tokens_out":7910,"duration_ms":77274,"temperature":0.7,"pith_summary":"This paper studies a fractional diffusion equation whose random source is the time-derivative of a fractional Brownian motion with Hurst index $H$. It tries to show that the direct problem is well posed exactly when $\\alpha+H>1$, and that the expectation and covariance of the solution at the final time uniquely determine the spatial profiles $f$ and $|g|$ of the source. This matters because fractional diffusion models anomalous transport in heterogeneous media, and real sources carry uncertainty; knowing which noise statistics can be recovered from terminal measurements is a prerequisite for using such models. The paper also shows that the recovery is unstable, with data Fourier modes decaying only like powers of $\\lambda_k^{-1}$, so regularization is unavoidable. The main advance over earlier work is replacing Brownian noise by fractional Brownian noise and extending the admissible fractional order to the full range $0<\\alpha\\le1$ under the combined condition $\\alpha+H>1$.","feed_headline":"Final-time statistics identify random source in fractional diffusion","feed_subtitle":"Uniqueness holds when fractional order plus Hurst index exceeds one, but recovery of the noise amplitude is unstable.","key_machinery":"The central object is the Mittag-Leffler kernel $G_{\\alpha,k}(t)=t^{\\alpha-1}E_{\\alpha,\\alpha}(-\\lambda_kt^\\alpha)$, which plays the role of the exponential heat kernel in fractional diffusion. The argument separates variables in the Laplacian eigenbasis, producing scalar stochastic fractional differential equations per mode. The fractional Brownian motion enters through its covariance kernel: for $H>1/2$ the quadratic variation of the stochastic integral is an explicit double integral over $|p-q|^{2H-2}$, while for $H<1/2$ it is expressed through the square-integrable kernel $K_{H,T}$ and the associated It\\^o isometry. The positivity of these integrals, established in Lemma 4.2, is what lets products $g_kg_\\ell$ be recovered from the covariance data. The instability comes from splitting the time integral at $t_*=\\lambda_k^{-\\gamma}$ and bounding the separated pieces by powers of $\\lambda_k^{-1}$.","core_discovery":"The paper's central claim is Theorem 4.3: under Assumption 1, with $f,g\\in L^2(D)$, $g\\not\\equiv0$, and $h$ positive and bounded below, the collection of expected values and covariances of the Fourier modes of the final-time field, $\\{\\mathbb E(u_k(T,\\omega)), \\operatorname{Cov}(u_k(T,\\omega),u_\\ell(T,\\omega))\\}_{k,\\ell\\in\\mathbb N}$, determines $f$ and $|g|$ uniquely. The direct problem is claimed to be well posed for $0<\\alpha\\le1$, $01$, with the second-moment bound $\\mathbb E(\\|u\\|^2_{L^2(D\\times[0,T])}) \\lesssim \\|h\\|^2\\|f\\|^2 + T^{2\\alpha+2H-1}\\|g\\|^2$. The inverse problem is claimed to be unstable: the mode-$k$ data decay at least like $\\lambda_k^{-1}$ for the mean and $\\lambda_k^{-\\beta}$ for the variance, where $\\beta$ is a positive exponent that can be made small, so small data perturbations can produce large source errors. These results are established by separation of variables, the Mittag-Leffler representation of the fractional evolution, and the It\\^o isometry for fractional Brownian motion.","pith_inferences":["Inference: Because the instability exponent $\\beta$ can be made arbitrarily small by choosing $\\gamma$ close to $0$ or $1$, practical reconstruction of $|g|$ will require strong regularization or additional data, not just more samples.","Inference: A direct numerical check of the covariance lower bound for $H<1/2$ and $\\alpha$ near 1 would either confirm Lemma 4.2 or expose a gap; this is a tractable one-mode calculation.","Inference: The same statistic-based strategy should generalize to other random drivers with stationary increments and to nonlinear functionals of the field, since only the covariance kernel of the noise enters the argument.","Inference: The uniqueness result distinguishes $|g|$ but not the sign of $g$, so any reconstruction can only recover the magnitude of the random amplitude, matching the physics of second-order statistics."],"forward_implications":["For any fractional order $\\alpha\\in(0,1]$ and Hurst index $H\\in(0,1)$ with $\\alpha+H>1$, the expectation and covariance of the final-time field determine the deterministic profiles $f$ and $|g|$ uniquely.","The condition $\\alpha+H>1$ is exactly the regime where the mild solution has finite second moments; when it fails, the singular integrals used to define the solution need not converge.","Reconstruction is unstable: mode-$k$ data decay no faster than $\\lambda_k^{-1}$ for $f$ and $\\lambda_k^{-\\beta}$ for the variance, so arbitrarily small data noise can produce large source errors.","The same separation-of-variables arguments carry over to the fractional Laplacian, giving the same uniqueness and instability results in that setting.","The Brownian-motion special case with $1/2<\\alpha<1$ is recovered as $H=1/2$, and the extension here covers all $0