{"id":"9e1b6805-6f8f-4c90-bed4-4e18aa1163c6","arxiv_id":"2411.14332","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A gradient-flow particle algorithm designs continuous measurement locations for nonlinear inverse problems and improves parameter reconstruction on Lorenz-63 and Schrödinger tests.","lead":"This paper builds a numerical method that chooses where and when to take measurements in nonlinear scientific experiments, treating time and location as continuous choices instead of a fixed list. It shows better parameter recovery on the Lorenz weather model and a Schrödinger equation, which could make data collection cheaper and more informative.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The outer-level update omits the implicit derivative of σ*[ρ] with respect to ρ; as written, Algorithms 1–2 are not gradient flows for the stated bilevel objective (3), only parameter-design alternations.","rationale":"The paper's central claim is about solving the continuous nonlinear OED problem (3), which requires the ρ-update to be a descent step for the reduced objective. The velocity fields (16) are correct for F at frozen σ, but the objective in (14) is F[ρ;σ*[ρ]], and the derivative through the inner solution is omitted. Algorithm 2's implicit-function update for σ* tracks the inner solution along a prescribed ρ trajectory but does not feed the chain-rule term back into the outer velocity. The paper's own limitation statement in Section 6 confirms that no convergence theory is provided, so numerical experiments are the only support. The experiments show that fixed-σ benchmarks and adaptive runs produce similar spike locations, which is encouraging but does not test whether the reduced objective is actually decreased along the algorithm trajectory. The sign inconsistency between (17) and (18) and the uncontrolled Schrödinger approximation add further reasons not to accept the central claim as written. The proposed scalar test would settle whether the omitted term is numerically significant; if it is not significant for these test problems, the method might still be a useful alternating heuristic, but the stated gradient-flow justification would need to be revised.","tokens_in":20731,"tokens_out":10373,"duration_ms":104819,"concrete_test":"Use a scalar model M(θ;σ)=θσ+σ^2, θ∈[0,1], with data generated from σtrue and a fixed nonzero noise realization. For a fixed two-atom design ρλ=(1−λ)δ_{θ1}+λδ_{θ2}, solve the inner problem accurately at several λ values and compute the exact reduced objective G(λ)=F_A[ρλ;σ*(ρλ)]. Compare dG/dλ by finite differences with the direction induced by Algorithm 1's Euler step using velocity (16a) evaluated at σ*(ρλ), and with Algorithm 2's coupled step using (17)–(22). If there is a λ interval where the algorithm's update increases G while dG/dλ says the exact objective decreases, the omitted chain-rule term is non-negligible and the central claim fails as stated. A stronger pass criterion: run the same comparison at λ=0.3, θ1=0.2, θ2=0.8; report the signed inner product ⟨ρ∇δG/δρ, v_alg⟩ — it must be negative for descent, and if it is positive the concern is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 states the outer flow as ∂tρ = ∇·(ρ∇ δF[ρ;σ*[ρ]]/δρ), i.e. a Wasserstein gradient flow for the reduced objective G(ρ)=F[ρ;σ*[ρ]]. The implemented velocity (16a)–(16b) is the derivative of F at a frozen σ, evaluated at σ*[ρ]. It omits the chain-rule term (∂F/∂σ)^T Dρσ*[ρ]. Algorithm 2 derives a σ* sensitivity (21)–(22) from the implicit function theorem, but this only moves the inner parameter along the already-chosen ρ trajectory; it does not correct the outer velocity with that sensitivity. Therefore the coupled update is not, in general, a descent direction for G. This is not a missing convergence proof: the update itself differs from the gradient of the objective being optimized. The sign inconsistency between (17) and (18) (minus vs plus in the particle velocity) reinforces that the printed formulas are unreliable. Section 5's Schrödinger runs additionally discard HessσM and ∇θ∇σM terms in (23)–(24) without an error estimate. Section 6 concedes that no convergence theory is provided; numerical experiments are the only support, and they compare against a fixed-σ benchmark rather than testing whether the reduced objective decreases along the algorithm trajectory.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a computational method for nonlinear optimal experimental design (OED) over a continuously indexed design space. It formulates the problem as a bilevel optimization (3) in which the outer level minimizes a design criterion F over a probability measure ρ on the design space, and the inner level computes the maximum-likelihood-type parameter estimate σ*[ρ] from the weighted loss (1). The proposed solvers, Algorithm 1 (brute-force inner loop) and Algorithm 2 (one-step streamlined inner update), move particles by a Wasserstein gradient-flow velocity and update σ* either by full inner optimization or by an implicit-function-theorem step. Numerical experiments on the Lorenz 63 system and a steady-state Schrödinger model show that the adaptive designs concentrate measurements at physically plausible times or locations and improve parameter reconstruction relative to uniform sampling. The paper claims in Section 1.2 to be the first to address nonlinear OED by optimizing a continuous design distribution in Pr2(Ω).","tokens_in":21059,"tokens_out":4280,"duration_ms":39968,"significance":"If the method is correct, it is a useful contribution to OED: it addresses a genuinely nonlinear, continuously indexed setting that most existing discrete-design or linear-model methods do not cover, and the particle/gradient-flow formulation is a natural way to avoid finite-dimensional design grids. The experimental comparisons on Lorenz 63 and the Schrödinger equation are valuable demonstrations of the type of design patterns the method can discover, and the paper is honest about the lack of convergence theory in Section 6. However, the paper's central mathematical claim—that Algorithms 1 and 2 are Wasserstein gradient flows for the stated bilevel objective—is not supported by the printed formulas, because the outer velocity omits the implicit derivative of σ* with respect to ρ. The numerical results may still support the method as a heuristic, but the gradient-flow interpretation and the associated optimality claims need repair.","major_comments":[{"comment":"The paper labels (15) as the gradient flow of the reduced objective F[ρ;σ*[ρ]], but the velocity fields (16a)-(16b) are computed by differentiating F at a frozen σ and then evaluating at σ=σ*[ρ]. They omit the chain-rule term (∂F/∂σ)·Dρσ*[ρ] that arises from the implicit dependence of σ* on ρ. Therefore Algorithms 1 and 2 are not a Wasserstein gradient flow for the bilevel problem (3); they are an alternating design/parameter update heuristic. This is load-bearing for the central claim in Section 1.2. Please either compute the missing term and correct the velocities, or reframe the method as a heuristic and adjust the statements accordingly.","section":"Section 3.2, Eqs. (15)-(16)"},{"comment":"The signs in the particle update are inconsistent. Eq. (17) states θ̇_i = -∇_θ δF/δρ, while the forward-Euler update in (18) is θ_i^{t+1} = θ_i^t + Δt ∇_θ δF/δρ, with the minus sign omitted. This is not a harmless typo: the direction of particle motion and hence the entire algorithm depend on the sign. Please state the correct update and ensure that the implementation and the displayed formulas use the same sign.","section":"Section 3.2, Eqs. (17)-(18)"},{"comment":"The streamlined update in the Schrödinger experiments drops the second lines of both (23) and (24), i.e. the Hess_σ M and ∇_θ∇_σ M terms, with the rationale that the discrepancy M(θ;σ*)-data(θ) is small. No error estimate or numerical ablation is provided for this approximation. Since these terms carry second-order information about the nonlinear PtO map, the Schrödinger experiments demonstrate a simplified variant of Algorithm 2, not the full formula as written. Please provide a bound or a comparison showing that the dropped terms are negligible in the tested regime.","section":"Section 5, Eqs. (23)-(24)"},{"comment":"The authors explicitly concede that convergence properties of Algorithm 2 are open. This alone would be acceptable for a numerical methods paper, but in combination with the omission of the chain-rule term in Major Comment 1, the lack of a descent guarantee is more serious: the empirical figures show that plotted objectives improve on the tested problems, yet there is no evidence that the iterates descend the reduced objective G(ρ)=F[ρ;σ*[ρ]]. The paper should either provide a monotonicity/descent statement for a corrected update or clearly characterize Algorithm 2 as a heuristic without a gradient interpretation.","section":"Section 6, limitation 1"}],"minor_comments":[{"comment":"The set Γ(µ,ν) is used but not defined; please identify it as the set of couplings of µ and ν with the given marginals.","section":"Section 2.2, Eq. (8)"},{"comment":"The text sets '∆t = 50 1', which appears to be a typographical artifact; please clarify the intended step size.","section":"Section 4.2, warm-start A-optimal"},{"comment":"The sentence 'the top-right panel shows that the design score starts from a value much close to optimality' appears to refer to the wrong panel of Fig. 24; the top-right panel shows reconstruction error, not the design score. Please correct the cross-reference.","section":"Section 5.2, last paragraph"},{"comment":"The hyperparameters T, Δt, Δt', and T' are reported separately for each experiment but not collected in one place; a summary table would make the experiments easier to reproduce.","section":"Section 6, limitation 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's novelty claim in Section 1.2 is stated strongly ('first to address nonlinear OED by optimizing a continuous design distribution'). The relationship to the authors' own linear-case paper [37] and to [34] should be checked carefully, and the novelty statement may need to be softened. The main technical issue is the missing chain-rule term in the outer velocity; if the authors can supply that term or convincingly argue it is negligible, I would be willing to reconsider."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this paper has a good problem, a reasonable heuristic, and a central derivation that doesn't hold up. The authors claim to run a Wasserstein gradient flow for the bilevel objective F[ρ;σ*[ρ]], but the velocity they implement is the gradient of F at fixed σ, evaluated at σ*[ρ]. They never compute the chain-rule term from the implicit dependence of σ* on ρ. So Algorithms 1 and 2 are not actually gradient flows for the stated problem (3); they are alternating heuristics. This isn't a missing convergence proof—the update itself is different from the gradient of the objective being optimized.\n\nThere is real value here too. The paper is, as far as I can tell, the first in its reference list to optimize a continuous design measure for nonlinear OED. The particle method is natural, and the numerical experiments are informative: the Lorenz test recovers many of the benchmark-important observation times, and the Schrödinger runs correctly push designs toward the diagonal and toward the potential's support. The strategic-versus-uniform sampling comparison in Fig. 6 shows a practical advantage. If the method is repositioned as an alternating heuristic, the empirical story is reasonably convincing.\n\nThe soft spots are proportional to a central claim. Besides the omitted chain-rule term, there is a sign inconsistency between (17) and (18): (17) has θ̇ = −∇F, while the Euler step (18) has θ^{t+1} = θ + Δt∇F. For D-optimality you actually want the plus sign (it's a maximization), so one of these equations is wrong for each criterion. The Schrödinger section drops the Hess_σ M and ∇θ∇_σ M terms in (23)–(24) without an error estimate, relying on σ* being close to σtrue. And the step sizes vary wildly (10^-7 to 0.5) across experiments, suggesting the method is not robust. The authors admit there is no convergence theory.\n\nNet: the problem is worth work, the numerics are suggestive, but the paper as written does not support the \"gradient flow\" claim. A serious referee could help the authors fix the derivation or reframe the method. I'd send it to review, but with clear instructions that the central claim needs retraction or major revision.\n\nBest.","headline":"The paper's central derivation is flawed—the 'gradient flow' omits the chain-rule term from σ*[ρ]—but the empirical design results are promising and worth refereeing.","tokens_in":21555,"tokens_out":7299,"would_cite":false,"duration_ms":65822,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62K05","49Q22","65K10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that continuous nonlinear optimal experimental design can be solved efficiently by evolving a design distribution with Wasserstein gradient flow, and demonstrates improved parameter reconstruction on the Lorenz 63 and…","keywords":["optimal experimental design","nonlinear inverse problems","Wasserstein gradient flow","particle method","bilevel optimization","A-optimality","D-optimality","continuous design space"],"falsifier":"On a small nonlinear model where the reconstruction is highly sensitive to the design, compute both the paper's particle velocity and the full derivative that includes the change of $\\sigma^*$ with $\\rho$; if they differ substantially, the algorithm is not following the gradient of the stated objective. A simpler observational check is to run Algorithm 2 and record $F[\\rho;\\sigma^*[\\rho]]$: if this score rises over iterations on a fixed problem, the update is not descending problem (3).","tokens_in":20526,"feed_emoji":"📡","tokens_out":8506,"duration_ms":67692,"temperature":0.7,"pith_summary":"This paper addresses optimal experimental design when the possible measurements form a continuous set, such as any observation time in an interval or any detector location in space. It proposes to optimize an entire probability distribution over the design space rather than choosing from finitely many experiments. The method alternates between moving design particles along a Wasserstein gradient flow and updating the unknown parameters, so that the design and the reconstruction improve together. If the claim holds, experimenters can automatically discover which measurement times and locations carry the most information for nonlinear inverse problems.","feed_headline":"Gradient flow picks the best measurements in inverse problems","feed_subtitle":"A bilevel particle method treats design as a continuous probability distribution and improves parameter recovery.","key_machinery":"The central object is the design measure $\\rho$ in the probability space $\\mathrm{Pr}_2(\\Omega)$, represented by particles $\\{\\theta_i\\}$. The engine is the Wasserstein-2 gradient flow (9), whose McKean-Vlasov particle form (11) moves each particle along $-\\nabla_\\theta \\frac{\\delta F}{\\delta \\rho}(\\theta_i)$. For nonlinear models the criterion $F$ is built from the Fisher information matrix $I[\\rho;\\sigma] = \\int_\\Omega \\nabla_\\sigma M(\\theta;\\sigma) \\nabla_\\sigma M(\\theta;\\sigma)^\\top \\rho(\\theta)\\,d\\theta$, with A-optimality $F^A = \\mathrm{Tr}(I^{-1})$ and D-optimality $F^D = \\log \\det(I)$. The explicit velocities (16a)-(16b) are the derivatives of these criteria with respect to $\\rho$, evaluated at the current $\\sigma^*$.","core_discovery":"The central discovery is that the nonlinear optimal design problem (3) can be attacked by a bilevel gradient-flow scheme. The outer level evolves the design measure $\\rho \\in \\mathrm{Pr}_2(\\Omega)$ by the Wasserstein gradient flow $\\partial_t \\rho = \\nabla_\\theta \\cdot (\\rho \\nabla_\\theta \\frac{\\delta F[\\rho;\\sigma^*[\\rho]]}{\\delta \\rho})$, with explicit velocity fields for A- and D-optimality given in (16); the inner level supplies the reconstruction $\\sigma^*[\\rho]$ from (1). Algorithm 1 solves this inner problem accurately at each outer step, while Algorithm 2 replaces it with a single implicit-function-theorem update and reduces the per-iteration cost. In the Lorenz 63 and Schrödinger experiments the optimized design concentrates measurements at informative times and locations, and strategic sampling at those points gives faster and better parameter reconstruction than uniform sampling.","pith_inferences":["A natural next test is to compare this bilevel descent against the full gradient that includes the change of $\\sigma^*$ with $\\rho$; if the omitted term matters, the method may still succeed as an alternating heuristic rather than a true gradient method.","The diagonal concentration found in the Schrödinger example suggests a transferable sensor-placement principle: when sources and detectors are local, coincident pairs dominate information, which could be probed in other inverse problems.","The open convergence questions in the paper could be attacked empirically by fixing the time step, increasing $N$ and $T$, and checking whether the design measure approaches the benchmark ground-truth design computed with $\\sigma = \\sigma_{\\mathrm{true}}$.","The flat D-optimal density along the diagonal despite a varying potential hints that D-optimality is insensitive to the magnitude of the parameter field; verifying this on synthetic potentials with different contrast would clarify when A- and D-optimal designs differ."],"forward_implications":["The solver produces a full continuous design distribution, so experimental resources can be allocated without discretizing the design space into a fixed candidate set.","Because the design and the parameter estimate are updated together, the method applies to nonlinear models where the optimal design depends on the unknown parameters themselves.","In the Lorenz 63 tests, sampling at the algorithm-identified observation times converges faster and to lower reconstruction error than uniform sampling.","In the Schrödinger tests, the optimized design concentrates source-detector pairs along the diagonal $\\theta_1=\\theta_2$, indicating that collocated measurements are most informative.","Algorithm 2 lowers the per-iteration cost from $O(dNT')$ to $O(d^3+d^2N)$, making larger particle counts feasible."],"supporting_citations":[{"why":"Supplies the continuous linear OED gradient-flow formulation that this paper extends to nonlinear models.","marker":"[37]"},{"why":"Defines the A- and D-optimality criteria used as the design objectives.","marker":"[39]"},{"why":"Provides the Wasserstein gradient-flow theory behind the outer-level update.","marker":"[50]"},{"why":"Supplies the optimal transport background for Wasserstein distances and gradient flows.","marker":"[21]"},{"why":"Justifies the measure-valued interpretation of the gradient-flow PDE and the particle representation.","marker":"[5]"},{"why":"Underlies the implicit-function-theorem step used to update $\\sigma^*$ in Algorithm 2.","marker":"[42]"},{"why":"Provides the Lorenz 63 model used in the first set of numerical experiments.","marker":"[44]"},{"why":"Provides the Schrödinger equation model used in the second set of numerical experiments.","marker":"[56]"}],"fun_headline_variants":["Gradient flow optimizes continuous measurement design","Continuous OED solved by adaptive gradient flow","Bilevel gradient flow picks best measurement times","Wasserstein flow improves parameter reconstruction","Adaptive sampling beats uniform in inverse problems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that when the design distribution is nudged, the best-fit parameters change so slowly that their indirect effect on the design score can be ignored; if that term matters, the outer update need not be a descent direction for the stated problem.","fun_headline_variants_meta":{"raw":{"variants":["Gradient flow optimizes continuous measurement design","Continuous OED solved by adaptive gradient flow","Bilevel gradient flow picks best measurement times","Wasserstein flow improves parameter reconstruction","Adaptive sampling beats uniform in inverse problems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000157,"raw_usage":{"total_tokens":1171,"prompt_tokens":844,"completion_tokens":327,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":262}},"tokens_in":460,"tokens_out":327,"duration_ms":4117,"temperature":1.0,"reasoning_tokens":262,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:17:59.550157+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a small nonlinear model where the reconstruction is highly sensitive to the design, compute both the paper's particle velocity and the full derivative that includes the change of $\\sigma^*$ with $\\rho$; if they differ substantially, the algorithm is not following the gradient of the stated objective. A simpler observational check is to run Algorithm 2 and record $F[\\rho;\\sigma^*[\\rho]]$: if this score rises over iterations on a fixed problem, the update is not descending problem (3).","supporting_citations":[{"cited_title":"Op timal experimental design for linear models via gradient flow","cited_arxiv_id":null,"evidence_quote":"Supplies the continuous linear OED gradient-flow formulation that this paper extends to nonlinear models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the A- and D-optimality criteria used as the design objectives."},{"cited_title":"Optimal Transport for Applied Mathematicians: Calculus of V ariations, PDEs, and Modeling, volume 87","cited_arxiv_id":null,"evidence_quote":"Provides the Wasserstein gradient-flow theory behind the outer-level update."},{"cited_title":"Figalli and F","cited_arxiv_id":null,"evidence_quote":"Supplies the optimal transport background for Wasserstein distances and gradient flows."},{"cited_title":"Gradient flows: in metric spaces and in the space of probability measures","cited_arxiv_id":null,"evidence_quote":"Justifies the measure-valued interpretation of the gradient-flow PDE and the particle representation."},{"cited_title":"The implicit function theorem: history, theory, and applications","cited_arxiv_id":null,"evidence_quote":"Underlies the implicit-function-theorem step used to update $\\sigma^*$ in Algorithm 2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Lorenz 63 model used in the first set of numerical experiments."},{"cited_title":"Mastering quantum mechanics: essentials, theory, and appl i- cations","cited_arxiv_id":null,"evidence_quote":"Provides the Schrödinger equation model used in the second set of numerical experiments."}],"review_version":1}