{"id":"06f463d7-952b-4b9b-8e7a-4a0ddec08d91","arxiv_id":"2506.22779","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A semiparametric PDE model with a parametric physical part and a neural-network unknown mechanism is estimated by profiling maximum likelihood, with proofs of optimal nonparametric rates and root-n efficient inference on the physical parameters.","lead":"This paper proposes a statistical framework that combines known physical laws (a parametric PDE) with a flexible neural-network term to model noisy data, and constructs estimators with proven convergence and confidence intervals. The method promises to let scientists estimate unknown 'mechanisms' in PDE-governed data while still getting root-n inference on physical parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 3 (Lipschitz/stability/sensitivity) is unverified for the paper's PDE examples and may fail for parabolic smoothing maps; all rates and efficiency results depend on it.","rationale":"The reader identified Assumption 3 as the weakest link, and my reading agrees: the paper's headline claim of parametric efficiency for complex PDE models is conditional on stability and sensitivity conditions that are neither proved nor verified for any genuine PDE in the paper. The authors themselves flag this gap in Section 3.1, and Example 3 is a one-dimensional linear ODE, not a PDE. For the reaction-diffusion and Navier-Stokes examples emphasized in the simulations, the solution map from the nonparametric term F to u is a smoothing operator, and the required lower bound (9) is a nontrivial observability/stability condition that cannot be taken for granted. The assumption r = r1 is also a substantive restriction, and the paper provides no evidence that this equality holds for the complex models in the title. I do not see an internal inconsistency in the main argument beyond this gap, but the gap is load-bearing: if Assumption 3 fails for the examples, the stated rates and efficiency do not follow. The reader's conditional verdict is therefore appropriate; no change is needed. I also note a secondary concern: the efficient variance formula in Theorem 3 appears to have a factor/constant discrepancy with the later expression in (12), but the primary blocker remains the unverified Assumption 3.","tokens_in":20432,"tokens_out":11758,"duration_ms":132539,"concrete_test":"Choose the linearized reaction-diffusion SemiPDE from Example 1: let theta = theta0, f0(u) = u(1-u), and linearize around the true solution u0 to obtain the parabolic equation for the perturbation delta_u: partial_t delta_u = theta0 Delta delta_u + (1-2u0) delta_u + h(x,t), where h is a perturbation in F. For h_k(x) = sin(kx) (taking t fixed, or integrating over the time interval), compute the L2 norm of the solution delta_u via the semigroup formula and compare it with ||Gamma_{r1} h_k||_{L2} for r1 = 1 and r1 = 2. If for some sequence k the ratio ||delta_u||_{L2}^2 / ||Gamma_{r1} h_k||_{L2}^2 tends to 0, then the stability condition (9) fails for that r1, directly invalidating Assumption 3 for the paper's flagship reaction-diffusion example. If instead the ratio is bounded below for an appropriate r1, the concern is weakened and the theory may apply.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claims (Theorem 1's optimal nonparametric rate and Theorem 3's root-n efficiency for theta) rest entirely on Assumption 3. The paper explicitly states in Section 3.1 that 'verifying these conditions for specific SemiPDE models remains an open and challenging problem,' and the only worked verification, Example 3 (eq. 10), is a first-order linear ODE, not a PDE. For the reaction-diffusion, Nernst-Planck, and Navier-Stokes examples used in the simulations, no argument is given that Assumption 3 holds. This matters because the solution map F maps to u for typical parabolic/elliptic PDEs is smoothing/compact: high-frequency perturbations in the source term F can produce L2 perturbations in u that are much smaller than the perturbation itself. If the lower bound in (9) fails, Theorem 1's rate n^{-2s/(2(s+r)+d)}, which is the well-posed nonparametric derivative-estimation rate, is unattainable, and the root-n efficiency in Theorem 3 collapses. The main text also assumes r = r1 as 'commonly satisfied,' but the correct smoothing order of the PDE may differ from the order appearing in the stability condition; this is not established for any of the complex models in the headline claim. Since the proofs are deferred to the supplementary, the load-bearing premise of all four theorems is currently unsupported for exactly the 'complex PDE models' highlighted in the abstract.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a semiparametric PDE (SemiPDE) model that combines a parametric physical operator L(u;θ) with a nonparametric unknown mechanism F(u), and proposes a profiling penalized M-estimation procedure in which the solution u is replaced by a numerical PDE solve for each (θ,F), with F represented by an overparameterized feedforward neural network. The main theoretical claims are (i) Theorem 1: an L2 convergence rate for the estimated mechanism f̂ of order n^{-2s/(2(s+r)+d)} log n under Assumptions 1-4; (ii) Theorem 2: a faster rate for the estimated solution û; (iii) Theorem 3: root-n consistency, asymptotic normality, and—under Assumption 6—semiparametric efficiency of the estimator θ̂ of the parametric part; and (iv) Theorem 4: an asymptotically valid confidence interval based on a numerically estimated efficient covariance. The paper also presents simulation comparisons against parametric, nonparametric, and PINN baselines, and two real-data applications (cell scratch assays and wave propagation through vegetation). The central mechanism is a set of Lipschitz, stability, and sensitivity conditions on the PDE solution map (Assumption 3), which the paper explicitly states remains open to verify for general SemiPDE models.","tokens_in":20782,"tokens_out":2241,"duration_ms":27621,"significance":"If the theoretical results are correct and the key assumptions are verifiable for practically relevant PDEs, this would be a substantial contribution: it would be the first semiparametric efficiency result for deep M-estimation in PDE settings, and it offers a concrete algorithmic template (profiling with numerical solvers) that avoids explicit score construction. The paper is also honest about the difficulty of its core assumptions, and it provides a useful taxonomy of how the PDE smoothing order r enters the nonparametric rate. However, the significance is currently conditional: the headline theorem (Theorem 3) rests on Assumption 3, which is not verified for any of the PDEs used in the simulations or real-data analyses, and the proofs are relegated to a supplementary file that is not part of the arXiv version under review. The paper's own Section 3.1 acknowledges that 'verifying these conditions for specific SemiPDE models remains an open and challenging problem,' and the only worked example (Example 3, Eq. (10)) is a first-order linear ODE, not a PDE.","major_comments":[{"comment":"Assumption 3 is the premise for all four theorems, but it is not verified for any PDE example in the paper. The only concrete verification, Example 3, is a first-order linear ODE (B_x u = θu + F(u) on [0,1]), not a partial differential equation. The reaction-diffusion, Nernst-Planck, and Navier-Stokes cases used in Section 5 are not analyzed with respect to the Lipschitz, stability, and sensitivity conditions. Since parabolic/elliptic solution maps can be smoothing in ways that may violate the lower bound in Eq. (9), the claimed rates in Theorems 1-3 would break down if the margin condition fails. The paper's own statement in Section 3.1 that verifying these conditions remains open is therefore a critical gap: the central claims are conditional on an assumption that is not demonstrated for any example in the manuscript.","section":"§3.1, Assumption 3 and Example 3 (Eq. (10))"},{"comment":"The paper restricts the main theory to the case r = r1, asserting that this is 'commonly satisfied,' but no justification is provided for the reaction-diffusion, Nernst-Planck, or Navier-Stokes models. The smoothing order of the PDE solution map (which determines the appropriate r in the Lipschitz condition (8)) and the integrability order appearing in the stability condition (9) may differ; for parabolic equations the forward map is typically smoothing, while the inverse map may have a different order. Without an argument that r = r1 holds for at least one of the complex models in the headline claim, the rates in Theorems 1-2 and the efficiency result in Theorem 3 are not connected to the simulations.","section":"§3.1, paragraph after Eq. (9), r = r1 restriction"},{"comment":"The proofs of all theorems are in a supplementary file that is not included in the arXiv version under review, so the derivation cannot be independently checked. Additionally, the paper's final discussion (Section 7) lists 'verifying the critical though challenging assumptions, such as Lipschitz and stability conditions' as future work, which is a direct admission that the assumptions are not currently established for the models that the paper claims to cover. For a methodological paper whose main contribution is theoretical, the assumptions should be verified at least for a nontrivial PDE example—for instance, a reaction-diffusion equation with a known reaction function—or the claims should be restated as conditional on Assumption 3 without applying them to the simulated cases.","section":"Theorems 1-4 and Section 7 (Discussion)"},{"comment":"The efficiency claim in Theorem 3 requires Assumption 6, which imposes Gaussian noise, a specific form h0 = -B_θ f(V; θ0), and the identification of f(V;θ) as the population projection. This is a strong assumption that is not verified for the SemiPDE models in the simulations. While the paper notes that Theorem 3 remains value as a root-n consistency result without Assumption 6, the abstract's statement that 'our estimator for θ attains parametric efficiency for complex PDE models' is not supported unless Assumption 6 is shown to hold for such models. Please provide at least a characterization of which PDE families satisfy Assumption 6.","section":"§4, Assumption 6 and Theorem 3"}],"minor_comments":[{"comment":"Equation (5) uses 'Fpuq' as a differential operator but later the argument is 'Fpu(x))'; the notation should be made uniform, and the distinction between the operator F and its function representation f(V) should be clarified at first use.","section":"Throughout"},{"comment":"The definition of the FNN space uses L layers but the formula writes L_0 through L_{L+1}; the indexing is confusing and should be reconciled with the width vector dimensions m_0,...,m_{L+1}.","section":"§2.3, Eq. (7)"},{"comment":"Table 1 would benefit from a column header indicating the quantity reported (e.g., L2 error versus parameter error) and from a statement of the Monte Carlo standard error, since the comparison between SemiPDE and Baseline 3 is sometimes close (e.g., Case 1 at n=800).","section":"§5, Table 1"},{"comment":"The stopping time T in Section 3.2 is defined via the gradient flow, but Algorithm 1 uses a validation-loss-based early stopping; the connection between the two stopping rules should be explained, otherwise it is unclear that the algorithmic estimator analyzed in the theory is the one implemented.","section":"Algorithm 1 and §3.2"},{"comment":"The estimates of θ for SemiPDE 1 (1232) and Benchmark 1 (853) are listed without standard errors or confidence intervals; given the paper's emphasis on inference, reporting the proposed confidence intervals for θ in the real-data example would strengthen the practical message.","section":"§6.1, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a serious attempt to bring semiparametric efficiency to PDE-constrained M-estimation with neural networks, and I believe the framework is worth publishing after major revision. The reader's skepticism about Assumption 3 is well-founded and lands: the central rates and efficiency claims are not tied to any concrete PDE example, and the only worked example is an ODE. In addition, the supplementary proofs are absent from the arXiv version; for a JRSS-style paper this is common, but the main text should summarize the proof strategy in sufficient detail. I would advise the editor to request the supplementary material before final acceptance and to require verification of Assumption 3 for at least one genuinely PDE-based example (e.g., the reaction-diffusion case) or a clear statement limiting the theoretical claims to models satisfying Assumption 3 without claiming application to Navier-Stokes or Nernst-Planck."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a serious semiparametric theory paper that deserves a referee, but the load-bearing assumption is exactly the part the authors admit they cannot verify. Read it for the profiling idea and the rate analysis; don't take the efficiency claims at face value until the stability conditions are checked for real PDEs.\n\nWhat is genuinely new: they propose a semiparametric PDE model where the known physical operator L(theta) is parametric and the unknown mechanism F(u) is a neural network, then estimate theta and F by profiling out the PDE solver. The decoupling is smart—it avoids joint M-estimation instability. They also analyze gradient flow rather than just the global minimizer, which is more honest about what training actually does. The convergence rate for the nonparametric part matches the minimax derivative-estimation rate n^{-2s/(2(s+r)+d)} log n, and they show root-n normality and efficiency for theta under an orthogonality condition. Those are real contributions over the earlier M-estimation framework.\n\nThe soft spots. First and most importantly, Assumption 3 (Lipschitz/stability/sensitivity) is the foundation for every theorem, and the authors state in Section 3.1 that verifying it for specific SemiPDE models remains open. The only concrete verification, Example 3, is a first-order linear ODE—not a PDE. For the reaction-diffusion, Nernst-Planck, and Navier-Stokes examples used in the simulations, no verification is provided. Parabolic smoothing can kill the lower bound in (9), which collapses the root-n efficiency. This is not a hidden flaw—they flag it themselves—but it means the main claims are conditional in a way that limits practical applicability. Second, the proofs are all in the supplementary, which the arXiv version lacks, so I could not check the derivation of the efficient variance or the training dynamics. Third, the real-data sections report point estimates but no confidence intervals, despite the paper promising valid inference; Table 4's theta estimates come without any uncertainty quantification. That is a gap between theory and demonstration.\n\nWho should read this: statisticians working on semiparametric inference in PDE-governed data and people building theory for neural-network nuisance estimation. The paper is honest about its limitations, and the central idea holds up modulo Assumption 3. I would send it to a serious referee, but I would ask for verification of the stability condition on at least one genuinely parabolic example, and for the supplementary material to be made available.\n\nCandidly, I would not cite it in my next paper until the assumption is checked, because the efficiency claim depends on it.","headline":"A serious semiparametric theory paper whose efficiency claims rest on an unverified stability condition; worth a referee, not yet a cite.","tokens_in":21239,"tokens_out":2053,"would_cite":false,"duration_ms":78148,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G08","62G20","62F12","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that a profiling M-estimator for semiparametric PDEs estimates unknown mechanisms at the optimal rate and physical parameters at parametric efficiency.","keywords":["semiparametric PDE","profiling M-estimation","neural networks","semiparametric efficiency","nonparametric convergence rate","PDE discovery","root-n consistency"],"falsifier":"Simulate from a SemiPDE whose stability margin shrinks with $n$, for example by placing the unknown mechanism under an integral operator that nearly annihilates its effect on the solution, so that the constant $C_3$ in Assumption 3(b) tends to zero; under the paper's theory the $L^2$ error of $\\hat f$ should still track $n^{-2s/(2(s+r)+d)}\\log n$ and the Theorem 4 intervals should hold coverage, so a visible departure from either would falsify the central claim.","tokens_in":20203,"feed_emoji":"🧮","tokens_out":8883,"duration_ms":83319,"temperature":0.7,"pith_summary":"This paper introduces SemiPDE, a model class in which observed data follow $Y = u(X) + \\varepsilon$ and $u$ solves a PDE made of two parts: a known physical law $L(u;\\theta)$ with unknown parameters $\\theta$, and a completely unknown mechanism $F(u)$. The paper proposes a penalized profiling M-estimator that replaces $u$ by the numerical solution $\\tilde u(x;\\theta,F)$, so statistical estimation is decoupled from PDE solving, and estimates $F$ in an overparametrized neural network space. The paper claims this is the first semiparametric inference theory for deep M-estimation in PDE models, with the unknown mechanism estimated at the optimal nonparametric rate $n^{-2s/(2(s+r)+d)}\\log n$, the solution at a faster rate, and $\\theta$ at root-$n$ asymptotic normality and parametric efficiency. The payoff is a way to identify interpretable physical parameters from noisy data while allowing unknown terms to be learned rather than assumed away. The cost is a load-bearing stability condition that the authors note remains open to verify for general SemiPDE models.","feed_headline":"Profiling M-estimation learns PDE mechanisms at optimal rates","feed_subtitle":"A semiparametric PDE model pairs known physical laws with learned neural mechanisms, giving root-n inference on parameters.","key_machinery":"The machinery is the penalized profiling M-estimator in equation (6), where $u(x)$ is replaced by the numerical solution $\\tilde u(x;\\theta,F)$ of the SemiPDE and the unknown operator $F$ is searched inside an overparametrized feedforward neural network space $S_n$ with penalty $\\|\\phi - \\phi_0\\|^2$. This decoupling keeps statistical error confined to the neural-network parameters and lets the PDE solver contribute high accuracy without adding noise. The theoretical load is carried by Assumption 3 (Lipschitz, stability with a positive identifiability margin, and sensitivity conditions on the solution map $(\\theta,F)\\mapsto u$), by Assumptions 5 and 6 (existence of a sufficiently smooth orthogonal direction $h_0$ making the score $\\psi = \\partial_\\theta l - \\partial_F l[h_0]$ orthogonal to the nuisance tangent), and by the gradient-flow analysis of Algorithm 1 that yields the nonparametric rates.","core_discovery":"The central claim is that profiling a numerical PDE solver inside M-estimation turns the SemiPDE model into a tractable semiparametric problem: Theorem 1 gives $\\|\\hat f - f_0\\|_{L^2}^2 = O_p(n^{-2s/(2(s+r)+d)}\\log n)$ for the unknown mechanism, matching the minimax rate for estimating an $r$-th derivative up to a logarithm; Theorem 2 gives the solution rate $n^{-2(s+r)/(2(s+r)+d)}\\log n$; and Theorem 3 shows $n^{1/2}(\\hat\\theta - \\theta_0) \\to N(0,\\Sigma)$, with $\\Sigma = \\Sigma_{\\mathrm{eff}}$ when Assumption 6 holds, so $\\hat\\theta$ is semiparametrically efficient. This is achieved without constructing an efficient score function: only the existence of a smooth direction $h_0$ annihilating the nuisance gradient is assumed, and the efficient covariance is reformulated as the minimizer of a convex optimization problem and estimated by finite differences. The paper therefore claims the first systematic analysis of neural-network estimators for PDE models that covers both the training dynamics, analyzed as gradient flow, and the differential structure of the model.","pith_inferences":["Editorial inference: the decoupling idea is not specific to PDEs; the same profiling with a numerical forward map could extend semiparametric efficiency to other implicit models such as ODE systems and integral equations, provided stability and margin conditions analogous to Assumption 3 hold.","Editorial inference: the paper's rate for $\\hat f$ degrades as $r$ grows, suggesting that high-order PDEs are inherently harder for mechanism recovery; a testable extension is to check whether a spectral or multi-scale parameterization of $F$ recovers faster rates than the neural-network space.","Editorial inference: the reliance on Gaussian noise in Assumption 6 means the parametric efficiency claim may not extend to heavy-tailed errors; replacing Gaussianity with sub-exponential tails and a robustified loss would be a natural stress test."],"forward_implications":["If the theorems hold, practitioners can estimate the parametric part of a PDE at root-$n$ parametric efficiency without deriving an explicit efficient score function, which is unavailable for most PDE models.","The unknown mechanism $F$ is learned at the optimal nonparametric rate up to a log factor, so adding a data-driven term does not destroy the parametric convergence of $\\theta$.","The profiling strategy avoids the slow convergence of joint PINN-style M-estimation, so PDE discovery can exploit accurate numerical solvers rather than treating the solution as a trainable function.","The finite-difference variance estimator in equations (13)-(14) turns efficient covariance estimation into a stable convex program, yielding valid simultaneous confidence intervals for linear functionals of $\\theta$."],"supporting_citations":[{"why":"Supplies the profiling/generalized-smoothing idea that the SemiPDE estimation strategy builds on to separate nuisance estimation from parameter estimation.","marker":"Ramsay et al. (2007)"},{"why":"Establishes the parametric PDE parameter-estimation baseline, using numerical PDE solvers inside a least-squares criterion, which the profiling estimator generalizes to a nonparametric term.","marker":"Xun et al. (2013)"},{"why":"Provides the joint PINN-style M-estimation baseline whose slow convergence motivates the decoupled profiling approach.","marker":"Raissi et al. (2019)"},{"why":"Defines the orthogonal-score/debiased-machine-learning framework that the paper contrasts with its existence-only assumption on $h_0$.","marker":"Chernozhukov et al. (2018)"},{"why":"Supplies the neural-network nonparametric regression rate and the $\\log n$ factor used in the convergence-rate analysis.","marker":"Schmidt-Hieber (2020)"},{"why":"Gives the optimal nonparametric convergence rate against which Theorem 1's rate is benchmarked.","marker":"Stone (1980)"},{"why":"Provides the technique for approximating the orthogonal direction $h_0$ by neural-network tangent spaces in semiparametric M-estimation.","marker":"Yan et al. (2025)"},{"why":"Documents the Lipschitz and stability conditions used in forward and inverse PDE analysis that Assumption 3 mirrors.","marker":"De Ryck and Mishra (2024)"}],"fun_headline_variants":["Deep profiling yields root-n inference for semiparametric PDE models","Profiling PDE solvers achieves efficient semiparametric inference","Semiparametric PDEs with neural components: root-n parameters","First semiparametric inference for deep PDE estimation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands on Assumption 3: the mapping from the physical parameters and the unknown mechanism to the PDE solution must be Lipschitz, stable with a positive identifiability margin, and sensitive to $\\theta$, uniformly over the parameter set; the paper itself flags that verifying these conditions for specific SemiPDE models remains open.","fun_headline_variants_meta":{"raw":{"variants":["Deep profiling yields root-n inference for semiparametric PDE models","Profiling PDE solvers achieves efficient semiparametric inference","Semiparametric PDEs with neural components: root-n parameters","First semiparametric inference for deep PDE estimation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000835,"raw_usage":{"total_tokens":3692,"prompt_tokens":1042,"completion_tokens":2650,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":658,"completion_tokens_details":{"reasoning_tokens":2579}},"tokens_in":658,"tokens_out":2650,"duration_ms":20422,"temperature":1.0,"reasoning_tokens":2579,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:57:52.296646+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate from a SemiPDE whose stability margin shrinks with $n$, for example by placing the unknown mechanism under an integral operator that nearly annihilates its effect on the solution, so that the constant $C_3$ in Assumption 3(b) tends to zero; under the paper's theory the $L^2$ error of $\\hat f$ should still track $n^{-2s/(2(s+r)+d)}\\log n$ and the Theorem 4 intervals should hold coverage, so a visible departure from either would falsify the central claim.","supporting_citations":[],"review_version":1}