{"id":"4f4b158f-880a-4830-a545-94236d482861","arxiv_id":"1909.00016","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For nonsmooth initial data, the authors derive optimal spatial and temporal error bounds for a semilinear fractional diffusion equation using a new Gronwall inequality.","lead":"This paper proves error estimates for finite element approximations of a semilinear fractional diffusion equation when the initial data is rough, not smooth. It introduces a new Gronwall inequality that lets the analysis proceed without smooth initial data, and tests the predicted rates with numerical experiments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 4.1 Eq. (48), the h^2 t^{-alpha} linear error estimate on which Theorem 4.1's headline pointwise bound rests, is stated without proof; the cited analogue is weaker, so the main spatial error estimate is not fully established.","rationale":"The reader's conditional verdict is appropriate. The central claim's headline spatial estimate (42) depends on Lemma 4.1 (48), an unproved linear nonsmooth-data error bound. The estimate is likely true, since it is standard for linear fractional subdiffusion, so I do not recommend rejection; but the paper explicitly omits its proof and the cited result in Remark 4.2 only supplies a weaker version with an extra logarithmic factor. Because (42) contains no such factor on the t^{-alpha} term, the omitted proof is load-bearing. The numerical experiments in Section 6 support the predicted rates for many parameter combinations, but they use a smooth nonlinearity and a reference solution, so they cannot substitute for the missing analytic bound. The second gap in Lemma 3.5's Gronwall proof is also real and should be addressed. Conditional acceptance, pending a complete proof of (48) and the Gronwall step, is the correct disposition.","tokens_in":26278,"tokens_out":25556,"duration_ms":219965,"concrete_test":"Independently derive (48) from the resolvent representation (19), (46): prove the uniform resolvent estimate ||(r^alpha e^{pm i alpha pi} - Delta)^{-1} - (r^alpha e^{pm i alpha pi} - Delta_h)^{-1} P_h||_{L(L2)} <= C h^2 for all r >= 0, then compute the convolution with D^alpha_{0+}v. The check is decisive: if the sharp bound is h^2 ln(1/h) t^{-alpha} rather than h^2 t^{-alpha}, then Theorem 4.1's estimate (42) must be weakened; if the proof goes through with no logarithmic loss, the omitted lemma should be written out in a revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 4.1 (48) claims for every v in L2(Omega) that ||(S D^alpha_{0+}v - S_h D^alpha_{0+}P_h v)(t)||_{L2} <= C h^2 t^{-alpha} ||v||_{L2}. This is the sole estimate that produces the O(h^2 t^{-alpha}) part of the pointwise error (42): Lemma 4.2's estimate (51) and Theorem 4.1 both apply it with v = u0. The proof is not given, only a note that it is 'similar to [17, Theorem 2.1]', and Remark 4.2 cites [5, Theorem 3.7] only for the weaker bound h^2 ln(1/h) t^{-alpha}. Because (42) has no logarithmic factor on the t^{-alpha} term, (48) is genuinely stronger than what the paper cites. If (48) fails or carries an extra log factor, the headline pointwise error (42) and the spatial rates (43)-(44) degrade. A second, related gap is in Lemma 3.5's proof, which uses an unproved inequality ||D^{-beta}_{t-} y|| <= C ||D^{-beta}_{0+} y||; that step is also load-bearing for the energy-norm estimates in Theorem 4.1. Both issues must be settled before the central spatial claim can be relied upon.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops a priori error analysis for the semilinear fractional diffusion equation D^alpha_{0+}(u-u0) - Delta u = f(u) on a convex polytope, with alpha in (0,1) and u0 in L2(Omega). The authors prove existence, uniqueness, and temporal regularity of the weak solution, introduce a new continuous Gronwall inequality and a discrete analogue, and use them to derive spatial finite-element error estimates for a semidiscrete scheme (Theorem 4.1) and temporal error estimates for a piecewise-constant discontinuous Galerkin discretization on graded grids (Theorem 5.2). The headline results are the pointwise L2 spatial error O(h^2(t^{-alpha}+ln(1/h))) and temporal rates O(J^{-(1-alpha)/2} sqrt(ln J)) for nonsmooth data. Numerical experiments in one dimension verify the predicted orders.","tokens_in":26559,"tokens_out":9327,"duration_ms":80455,"significance":"If the proofs are completed, the paper would supply one of the first numerical analyses for a semilinear fractional diffusion problem with nonsmooth initial data, extending linear subdiffusion theory to the semilinear setting. The proposed Gronwall-type inequalities are a useful technical contribution, the spatial rates (42)-(44) are sharp with respect to the solution regularity, and the graded-grid temporal rates in Theorem 5.2 go beyond uniform-grid analyses. The numerical section is honest in that errors are measured against a fine reference solution rather than calibrated to the theory, and the experiments reproduce the predicted rates in most regimes, with the authors themselves noting the discrepancy for alpha=0.2 in Table 5. However, two load-bearing estimates are not fully proved: Lemma 4.1(48) and an inequality appearing in Lemma 3.5. Until those arguments are supplied, the central spatial and full-discretization claims are conditionally established rather than rigorously proven.","major_comments":[{"comment":"The estimate (48) is the only ingredient that produces the O(h^2 t^{-alpha}) term in the pointwise error (42), since Lemma 4.2 applies it with v=u0 and Theorem 4.1 uses (51). Its proof is omitted ('similar to that of [17, Theorem 2.1]'), and the only supporting citation given in Remark 4.2, [5, Theorem 3.7], provides the strictly weaker bound O(h^2 ln(1/h) t^{-alpha}). If (48) holds only with the logarithmic factor, then the headline bound (42) and the spatial rates (43)-(44) degrade unless the theorems are restated. Please supply a complete proof of (48), or weaken the statement and propagate the change through Lemma 4.2 and Theorem 4.1.","section":"Section 4, Lemma 4.1, Eq. (48)"},{"comment":"In the proof of Lemma 3.5, the displayed chain of inequalities contains the step ||D^{-beta}_{t-}y||_{L2(0,t)} <= C_beta ||D^{-beta}_{0+}y||_{L2(0,t)}, justified by 'Lemma 2.2'. Lemma 2.2, however, concerns positive-order fractional derivatives and does not, as stated, imply a comparison of negative-order right- and left-fractional integrals. This inequality is load-bearing: Lemma 3.5 is invoked in the uniqueness proof in Theorem 3.1 and in the derivation of (54)-(55) in Theorem 4.1, and its discrete version underlies Lemma 5.1 and therefore (64)-(65). Please add a proof of this inequality or replace Lemma 3.5 with a version whose every step is justified.","section":"Section 3, Lemma 3.5"},{"comment":"Several results that are used as theorems are not proved in the text: the cases 0<alpha<=1/2 of Theorem 3.1 ('the rest... can be proved analogously'), the semidiscrete regularity bound (41) ('Following the proof of Theorems 3.1 and 3.2, we can easily prove'), the proof of (83) in Lemma B.1 ('Since the proof of (83) is similar, it is omitted'), and estimate (64) in Theorem 5.2 ('Since (64) can be proved analogously'). The omitted cases carry the spatial rates for alpha<=1/2 in (43)-(44) and the temporal rate (64), so these are not purely presentational omissions. Please expand these proofs or state explicitly which existing argument is being modified, where, and why the modifications do not change the estimates.","section":"Sections 3-5, delegated proofs"}],"minor_comments":[{"comment":"The text 'l = 1, 2.3, 4' should read 'l = 1, 2, 3, 4'.","section":"Section 2, conventions"},{"comment":"The headers 'J = 2^{-16}' and the reference-solution description should involve J = 2^{16} (or 65536), since J is the number of temporal intervals and must be a positive integer.","section":"Section 6, Tables 1-3"},{"comment":"The statement of Lemma 5.1 has the exponent (tau^* - tau)^{-1/2} in the exponential, while the proof yields (tau^* - tau)^{-1}; please reconcile the statement with the proof.","section":"Section 5, Lemma 5.1"},{"comment":"The notation '2kbeta' is ambiguous; it should be written as '2^k beta' to distinguish powers from products.","section":"Section 3, proof of Lemma 3.5"},{"comment":"The phrase 'optimal with respect to the regularity of the solution' is used without a definition of optimality or any lower-bound discussion; consider rephrasing to 'sharp up to logarithmic factors' where that is what is proved.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a numerical-analysis journal and the results would be valuable if the missing proofs can be supplied. My main concern is completeness rather than an identified false claim; in particular, Eq. (48) and the Lemma 3.5 inequality need full proofs before the central spatial and temporal bounds can be relied upon. If the authors provide those arguments, I would be willing to review a revised version favorably."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThis is a genuinely useful paper, but its headline spatial error estimate is not fully proven. The new content — regularity for semilinear fractional diffusion with u0 in L2, a Gronwall-type inequality and its discrete analogue, and graded-grid temporal error estimates — is real and likely reusable. The regularity results extend Jin–Li–Zhou to rough data, and the experiments mostly confirm the predicted rates.\n\nThe main soft spot is exactly where the stress-test points: Lemma 4.1(48), the O(h^2 t^{-alpha}) estimate on which Theorem 4.1's pointwise bound (42) rests, is stated without proof. The cited analogue [17] gives only h^2 ln(1/h) t^{-alpha} (as the authors note in Remark 4.2). That is a genuine gap. The result may be true, but as written the central claim is not established.\n\nThe second soft spot is Lemma 3.5. The specific inequality flagged in the stress-test — ||D^{-beta}_{t-} y|| <= C||D^{-beta}_{0+} y|| — is actually true by reflection, so that is not the problem. The problem is that the descent from D^beta to lower-order fractional integrals ('repeating the above argument k-1 times') is too compressed to verify, and the discrete version Lemma 5.1 has the same issue. These lemmas carry the energy-norm estimates, so they matter.\n\nSmaller issues: Theorem 3.1 proves selected ranges of alpha and leaves other cases 'analogously'; Experiment 3's alpha=0.2 rates fall short of the predicted orders, which the authors concede without a full explanation. These are minor relative to the two gaps above.\n\nWho this is for: numerical analysts working on fractional diffusion, especially nonlinear problems with rough initial data. It deserves peer review. I would send it out, but the referee report should demand a full proof of (48) or a revised theorem statement, and a more careful write-up of the Gronwall arguments. If those are patched, this becomes a solid contribution.","headline":"Worth refereeing, but the headline pointwise spatial estimate and the new Gronwall tool are not fully proven as written.","tokens_in":27087,"tokens_out":7753,"would_cite":false,"duration_ms":61749,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M60","35B65"],"pacs":[],"model":"deepseek-v4-flash","headline":"Finite element error bounds are proved for rough-data semilinear fractional diffusion.","keywords":["semilinear fractional diffusion equation","nonsmooth initial data","finite element method","graded temporal grid","fractional Grönwall inequality","fractional Sobolev spaces","error estimates","Riemann-Liouville fractional derivative"],"falsifier":"Evaluate the left side of Eq. (48) numerically for $v(x)=x^{-0.49}$ on $\\Omega=(0,1)$ at a fixed $t>0$: if $\\|(S D^\\alpha_{0+}v - S_h D^\\alpha_{0+}P_h v)(t)\\|_{L^2(\\Omega)}$ is not bounded by a constant times $h^2 t^{-\\alpha}\\|v\\|_{L^2(\\Omega)}$ on a sequence of uniformly refined meshes, the omitted lemma fails and Theorem 4.1 collapses.","tokens_in":26068,"feed_emoji":"📐","tokens_out":10754,"duration_ms":97728,"temperature":0.7,"pith_summary":"This paper proves the first numerical error estimates for a semilinear fractional diffusion equation in which the initial data is only in $L^2(\\Omega)$ rather than in a smoother Sobolev space. The central spatial result is a pointwise $L^2(\\Omega)$ bound $\\|(u-u_h)(t)\\|_{L^2(\\Omega)} \\lesssim h^2(t^{-\\alpha}+\\ln(1/h))\\|u_0\\|_{L^2(\\Omega)}$ for piecewise-linear finite elements, and the central temporal result is an $O(J^{-(1-\\alpha)/2}\\sqrt{\\ln J})$ error bound for a discontinuous Galerkin time discretization on graded grids. Both estimates are new because earlier analyses required $u_0\\in\\dot H^2(\\Omega)$. The paper attributes the gain to a new fractional Grönwall inequality and its discrete analogue, which convert fractional-derivative energy bounds into true error control. A sympathetic reader should care because these are the first rigorous convergence rates for a class of problems that are known to develop time singularities even from smooth data.","feed_headline":"Error bounds proved for rough-data fractional diffusion equations","feed_subtitle":"A new inequality makes finite element error estimates rigorous without smooth initial data.","key_machinery":"The argument is carried by two interlocking devices. The first is a new fractional Grönwall inequality, Lemma 3.5: if $\\|D^\\beta_{0+} y\\|^2_{L^2(0,t)} \\le \\epsilon + A\\|y\\|^2_{L^2(0,t)}$ for all $0<t<T$ with $0<\\beta<1/2$, then $\\|D^\\beta_{0+}y\\|_{L^2(0,t)} \\le C_{\\beta,A,T}\\sqrt{\\epsilon}$; Lemma 5.1 is its discrete, piecewise-constant analogue on graded grids. These inequalities let the authors turn the usual energy identity, which controls a fractional derivative of the error, into control of the error itself without assuming smoothness of $u_0$. The second device is the integral representation $Sg(t)=\\int_0^t E(t-s)g(s)\\,ds$ of the solution operator in terms of the kernel $E$ built from the resolvent of $-\\Delta$, together with the spatial projection estimate Lemma 4.1, Eq. (48), $\\|(S D^\\alpha_{0+}v-S_h D^\\alpha_{0+}P_h v)(t)\\|_{L^2(\\Omega)} \\lesssim h^2 t^{-\\alpha}\\|v\\|_{L^2(\\Omega)}$, which transfers the nonsmooth linear estimate to the semidiscrete setting.","core_discovery":"The paper's central claim is Theorem 4.1: for each $0<t\\le T$, the semidiscrete piecewise-linear finite element solution $u_h$ satisfies $\\|(u-u_h)(t)\\|_{L^2(\\Omega)} \\lesssim h^2(t^{-\\alpha}+\\ln(1/h))\\|u_0\\|_{L^2(\\Omega)}$ with $u_0\\in L^2(\\Omega)$. It also proves Sobolev-norm error estimates in ${}_0H^{\\alpha/2}(0,T;L^2(\\Omega))$, $L^2(0,T;\\dot H^1(\\Omega))$, and $L^2(0,T;L^2(\\Omega))$, with rates that change at the thresholds $\\alpha=1/3$ and $\\alpha=1/2$ and are optimal with respect to the regularity of the solution. For the full discretization, Theorem 5.2 gives $O(J^{-(1-\\alpha)/2}\\sqrt{\\ln J})$ convergence for $L^2$ data on graded grids and sharper rates, expressed through $\\eta_1(\\alpha,\\sigma,J)$ and $\\eta_2(\\alpha,\\sigma,J)$, for data in $\\dot H^1(\\Omega)$. The proof isolates the nonsmooth-data difficulty in the linear operator $S$ and its discrete counterpart $S_h$, controls the difference $u-\\tilde u_h$ by projection estimates, and then uses the new Grönwall inequality to control the nonlinear feedback $u_h-\\tilde u_h$.","pith_inferences":["The $t^{-\\alpha}$ factor in the pointwise bound suggests that on a fixed spatial mesh the early-time error dominates; a time-adaptive finite element code that refines the first moments more finely should be able to recover uniform $O(h^2)$ accuracy in $t$, a design question the paper does not address.","The abstract Grönwall lemma is not tied to the specific nonlinearity $f(u)$; the same two-step semidiscrete argument should carry over to locally Lipschitz nonlinearities, time-dependent coefficients, or systems of fractional diffusion equations, giving testable analogues of Theorems 4.1 and 5.2.","If the omitted projection estimate, Eq. (48), is proved for higher-degree finite elements, the identical architecture should yield $O(h^{r+1}(t^{-\\alpha}+\\ln(1/h)))$ pointwise accuracy for $P_r$ elements with only cosmetic changes.","The numerical results for small $\\alpha$ (for example, $\\alpha=0.2$ in Table 5) show rates below the predicted ones, and the paper attributes this to instability at large $J$; a clean way to separate a proof gap from a numerical artifact is to rerun those cases with a stable solver on substantially larger graded grids and check whether $E_1$ approaches $J^{-1/2}$."],"forward_implications":["Piecewise-linear finite element approximations of semilinear fractional diffusion converge in $L^2(\\Omega)$ at the rate $h^2$ away from the initial singularity, even when $u_0$ is merely square-integrable.","Near $t=0$ the spatial error is instead $O(h^2 t^{-\\alpha})$, so the estimate quantifies exactly how much the initial layer costs and where graded or adaptive meshes are needed.","The full discontinuous Galerkin time discretization on graded grids is guaranteed to converge at rate $J^{-(1-\\alpha)/2}\\sqrt{\\ln J}$ in the fractional and $H^1$ norms for $L^2$ data, with the better graded-grid rates of Theorem 5.2 for $H^1$ data.","The new Grönwall inequality and its discrete version are reusable tools: any fractional evolution equation whose energy identity yields the same fractional-derivative bound now has a route from that bound to a convergence theorem.","The earlier smooth-data analyses become a special case of the same argument, and the new proof does not require an extra square of the logarithmic factor in the pointwise rough-data estimate."],"supporting_citations":[{"why":"Supplies the regularity framework for smooth data and the previous pointwise estimate $O(h^2(\\ln(1/h))^2)$ that this paper extends to $L^2$ initial data.","marker":"[7]"},{"why":"The linear nonsmooth-data estimate that Lemma 4.1, Eq. (48), is asserted to follow from; the proof of (48) is omitted as similar to this result.","marker":"[17]"},{"why":"Provides the semidiscrete finite element projection estimates and the Mittag-Leffler representation used in Lemma 4.1 and Appendix A.","marker":"[5]"},{"why":"Supplies the abstract theorem used to convert the fractional integral inequality into the pointwise bound on $u_h-\\tilde u_h$ in the proof of Theorem 4.1.","marker":"[28]"},{"why":"Provides the interpolation space tools used throughout the regularity proofs and in the discrete Grönwall argument.","marker":"[26]"},{"why":"Gives the Mittag-Leffler growth estimate that underlies the regularity bounds in Lemma A.1 and Theorem 3.1.","marker":"[24]"},{"why":"The time-stepping analysis for linear fractional problems with nonsmooth data that informs the full discretization's graded-grid error estimates.","marker":"[9]"}],"fun_headline_variants":["New Grönwall inequality sharpens fractional diffusion error bounds","Optimal FEM error rates proven for rough-data fractional diffusion","Sharp temporal estimates on graded grids for fractional diffusion","Pointwise L2 error O(h^2) proved for nonsmooth fractional data","Novel discrete Grönwall inequality enables rigorous FEM analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the unproved spatial projection estimate $\\|(S D^\\alpha_{0+}v - S_h D^\\alpha_{0+}P_h v)(t)\\|_{L^2(\\Omega)} \\lesssim h^2 t^{-\\alpha}\\|v\\|_{L^2(\\Omega)}$ for all $v\\in L^2(\\Omega)$; if that transfer from a known linear result fails, the central pointwise error bound of Theorem 4.1 does not follow.","fun_headline_variants_meta":{"raw":{"variants":["New Grönwall inequality sharpens fractional diffusion error bounds","Optimal FEM error rates proven for rough-data fractional diffusion","Sharp temporal estimates on graded grids for fractional diffusion","Pointwise L2 error O(h^2) proved for nonsmooth fractional data","Novel discrete Grönwall inequality enables rigorous FEM analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000267,"raw_usage":{"total_tokens":1631,"prompt_tokens":976,"completion_tokens":655,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":571}},"tokens_in":592,"tokens_out":655,"duration_ms":6338,"temperature":1.0,"reasoning_tokens":571,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:05:59.482024+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the left side of Eq. (48) numerically for $v(x)=x^{-0.49}$ on $\\Omega=(0,1)$ at a fixed $t>0$: if $\\|(S D^\\alpha_{0+}v - S_h D^\\alpha_{0+}P_h v)(t)\\|_{L^2(\\Omega)}$ is not bounded by a constant times $h^2 t^{-\\alpha}\\|v\\|_{L^2(\\Omega)}$ on a sequence of uniformly refined meshes, the omitted lemma fails and Theorem 4.1 collapses.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the regularity framework for smooth data and the previous pointwise estimate $O(h^2(\\ln(1/h))^2)$ that this paper extends to $L^2$ initial data."},{"cited_title":"Lubich, I","cited_arxiv_id":null,"evidence_quote":"The linear nonsmooth-data estimate that Lemma 4.1, Eq. (48), is asserted to follow from; the proof of (48) is omitted as similar to this result."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the semidiscrete finite element projection estimates and the Mittag-Leffler representation used in Lemma 4.1 and Appendix A."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the abstract theorem used to convert the fractional integral inequality into the pointwise bound on $u_h-\\tilde u_h$ in the proof of Theorem 4.1."},{"cited_title":"Podlubny","cited_arxiv_id":null,"evidence_quote":"Gives the Mittag-Leffler growth estimate that underlies the regularity bounds in Lemma A.1 and Theorem 3.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The time-stepping analysis for linear fractional problems with nonsmooth data that informs the full discretization's graded-grid error estimates."}],"review_version":1}