{"id":"56c1d74c-efb6-4e12-935d-f3368260c63f","arxiv_id":"2511.02552","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A primal-dual-active-point algorithm with Radon-norm regularization identifies sparse contaminant sources in advection-diffusion problems from scarce sensor data, beating L2-regularized baselines in synthetic 2D/3D benchmarks.","lead":"This paper adapts a sparse-optimization algorithm to locate airborne contaminant sources from a few concentration sensors. For emergency response, the approach could give faster source estimates than smooth L2-regularized methods, but it assumes the wind field and source shapes are known.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Convergence theory and reported accuracy rely on a heuristic cluster-merging post-processing step not covered by the PDAP convergence results.","rationale":"The reader's weakest assumption (known wind field) is a legitimate modeling limitation, but it is explicitly scoped in Section 2.1 and does not directly threaten the internal correctness of the method. The discretization gap is more load-bearing: the paper's own Section 4 admits that continuum minimizers are not grid-supported and that the algorithm outputs clusters which are post-processed heuristically. The PDAP convergence result cited from [34] applies to exact global maximization over Ω, not to the nodal search actually implemented; the exception in Remark 2 covers only Dirac shape functions with linear FEM, not the RBF/elliptic cases used in most experiments. Therefore the reported 'handful of PDE solves' and localization errors rest on an unanalyzed heuristic. This does not invalidate the empirical results, but it strengthens the case for conditional acceptance pending a mesh-refinement/convergence study. Hence the reader's CONDITIONAL verdict remains appropriate, with this more specific technical condition.","tokens_in":26175,"tokens_out":6054,"duration_ms":71308,"concrete_test":"Run the 2D two-building benchmark with a true RBF source (φ1) centered exactly midway between two FE nodes; execute the discretized PDAP on a coarse mesh and on a uniformly refined mesh, and compare the post-processed location/amplitude to the continuum minimizer (computed on a much finer mesh or via a continuous-optimization solver). If the post-processed location error does not shrink with mesh refinement, or if the nodal cluster's barycenter is biased by more than one mesh cell, the heuristic is not validated. Repeat with the true source centered on a node as a control. This would settle whether the reported accuracy is due to the heuristic rather than the PDAP theory.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 adapts the PDAP convergence theory of [34] to the continuum problem (P), requiring exact global maximization of the dual variables. The actual implementation (Section 4) restricts admissible measures to grid-node Diracs and replaces Step 3's global maximization with a search over FE nodes. The paper acknowledges that continuum minimizers are generally off-grid and that this causes clustering, then merges clusters 'using barycentric coordinates' in a heuristic post-processing step. All quantitative localization results in Tables 1 and 3 are reported for the post-processed single source. No error analysis or mesh-refinement study links the nodal cluster to the continuum minimizer, and Remark 2's variational-discretization argument only covers piecewise-linear FEM with Dirac shape functions, not the RBF (φ1) or elliptic (φ2) cases used in most experiments. Thus the central claim that Algorithm 1 'efficiently approximates sparse minimizers' is supported only by the continuous PDAP theory, while the actual numerical algorithm has no proven convergence guarantee; the reported accuracy could be an artifact of the cluster-merging heuristic. Also, the assertion that the method 'requires only a handful of PDE solves per iteration' is undercut if a finer mesh is needed to make the heuristic reliable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops a variational regularization approach for identifying sparse contaminant sources in transient linear advection-diffusion equations from pointwise-in-time concentration measurements. The unknown source is represented as a positive Radon measure; the objective combines a quadratic data misfit with a Radon-norm/total-variation penalty. Theorem 1 states existence of sparse minimizers and a dual optimality condition; Corollary 1 gives an intensity subproblem. Algorithm 1 is a PDAP strategy that alternates between inserting locations where the dual variable exceeds α and re-optimizing intensities, with a claimed O(1/k) objective convergence in the continuum setting. Section 4 discretizes the forward/adjoint PDEs with stabilized finite elements and restricts source measures to grid-node Diracs, adding a heuristic barycentric cluster-merging step. Numerical experiments on 2D/3D benchmark geometries and two realistic domains compare favorably with an L2-regularized approach in accuracy and PDE-solve counts.","tokens_in":26450,"tokens_out":9133,"duration_ms":101605,"significance":"The paper addresses a practically important problem and contains useful theory: the convex relaxation is clean, Theorem 1 and Corollary 1 provide a solid foundation, and the numerical evidence suggests that PDAP can work on realistic geometries with few PDE solves. If the discretization/post-processing gap is closed, the method would be a valuable addition. At present, however, the quantitative claims are not fully supported: the analyzed algorithm is the continuum PDAP, while the implemented algorithm restricts candidates to grid nodes and relies on an unanalyzed post-processing step; the numerical validation is synthetic and uses the same source generators for ground truth and inversion, with per-case tuned α. These issues are fixable but require additional analysis or experiments.","major_comments":[{"comment":"The implementation replaces M_+(Ω) by the cone of node Diracs and replaces Step 3's global maximization by a search over finite-element nodes. The convergence statements in Section 3—including the finite-termination bound (13)—are for the continuum PDAP with exact global maximizers and exact subproblem solves. No theorem links the grid-restricted minimizers to minimizers of (P) as h→0, and Remark 2 establishes exactness of the node maximum only for piecewise-linear elements with φ3=δ; the experiments in §§5.1, 5.3.1, 5.3.2 and 5.3.4 use φ1 or φ2. The clustering caused by off-grid sources is repaired with a heuristic barycentric merge, and Tables 1 and 3 report the post-processed locations. The reported accuracy is therefore not a consequence of the PDAP theory; it can depend on the merging heuristic. Please provide a mesh-refinement study of the reconstructed source and either an error a","section":"§4 (after Eq. (18))"},{"comment":"Because candidate insertion is based on the maximum over the node set, the algorithm can stop even when the continuous dual maximum exceeds α+tol in an element interior. For φ1 and φ2 the dual φ^k is not piecewise-linear or is not represented in the finite-element space, so such interior maxima can be missed; this invalidates the stopping certificate and the bound (13) for the implemented method. A local refinement or quadrature check of the maximum, or a proof that the discrete maximum suffices for the relevant generators, is required.","section":"§3, Algorithm 1 Step 5 / §4"},{"comment":"The numerical validation is self-referential in two respects: all ground truths are generated from the same shape functions φ1, φ2, φ3 that are used in the inversion, and α is selected per test case (e.g., 'based on empirical evaluation' in §5.4). The claimed superiority over L2-regularization is therefore demonstrated only in an idealized setting. Please include model-mismatch experiments (e.g., reconstruct a φ1 source with the φ2 generator, or a source whose shape is not in the dictionary), a sensitivity analysis or systematic rule for choosing α, and multiple noise realizations or error bars.","section":"§5, Tables 1–3"}],"minor_comments":[{"comment":"In the five-sensor case (cases c/d), the L2 method has 16 online PDE solves versus PDAP's 20; the statement that PDAP 'clearly outperforms' is based on total online+offline cost. Please state the comparison metric explicitly and also discuss the online-only view, which is relevant for real-time deployment.","section":"§5.3.1, Table 2"},{"comment":"Typo: 'PDAD algorithm' should be 'PDAP algorithm'.","section":"§5.2"},{"comment":"The known, divergence-free wind field assumption is stated but not discussed as a limitation for emergency-response applications. A brief paragraph on sensitivity to wind-field errors would help contextualize the numerical results.","section":"§2.1"},{"comment":"The clustering radius of 40 m is an additional free parameter; its influence on the quantitative results in Table 3 should be reported.","section":"§5.4"},{"comment":"The convergence rate and finite-termination bound are adapted from prior work without proof. A short proof sketch or a precise statement of the assumptions needed for Eq. (13) in this setting would make the paper more self-contained.","section":"§3, Eq. (13)"},{"comment":"The mesh-independence study is performed only for the forward/adjoint discretization, not for the reconstructed source locations. A companion study for the inverse reconstructions would directly address the discretization gap noted in the major comments.","section":"§5, Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The discretization gap between the continuum PDAP theory and the grid-node implementation is the key issue. A reasonable revision path is to add a mesh-refinement study for the inverse problem, a model-mismatch experiment, and a systematic α-selection or sensitivity analysis, and to temper the 'outperforms L2' claim accordingly. The paper has solid theoretical foundations and a useful application-oriented contribution, so I do not see grounds for rejection if these points are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the paper is a competent engineering-style adaptation of known sparse-measure tools, and the actual new contribution is the combination: Radon-norm regularization and PDAP applied to simultaneous recovery of an initial condition and a continuous-in-time source in a transient advection-diffusion equation, with 2D/3D building-resolved benchmarks and a chemistry-plant case tied to ERIMap. That is a useful within-field step, not a new research direction. Second, the central claim that Algorithm 1 efficiently approximates sparse minimizers is only proven for the continuum problem; the actual implementation discretizes the measure to grid nodes and then merges clusters heuristically. All localization errors are reported after that post-processing, so the convergence theory and the numerical results are not connected. The paper is honest about the clustering, but it does not resolve the gap.\n\nWhat the paper does well: the theoretical framework is standard but correctly assembled—representer theorem, optimality conditions, and convergence rate adapted from the authors' prior work. The numerical experiments are extensive and cover challenging geometries, including OpenStreetMap building imprints and a real-world-inspired chlorine-release scenario. The SUPG-stabilized adjoint discretization and the discussion of dual-of-stabilized-primal versus stabilized-adjoint equations are genuinely useful implementation material. The authors also clearly state limitations, including the known-wind-field assumption and the heuristic nature of cluster merging.\n\nWhere the soft spots are:\n- The stress-test concern holds up. Remark 2's variational-discretization argument covers only piecewise-linear FEM with Dirac shape functions, but most experiments use RBF or elliptic shape functions. No error analysis or mesh-refinement study links the nodal cluster to the continuum minimizer, and no code or data is provided. This makes the 'outperforms L2 regularization' claim empirical, not proven.\n- The validation is self-referential: ground truth sources come from the same shape-function families used in inversion, alpha is hand-tuned per case, and the L2 baseline is not systematically tuned. That weakens the superiority claim to 'outperforms a particular L2 implementation on synthetic cases.'\n- The known-wind-field assumption is load-bearing for the stated emergency-response application. It is stated clearly in Section 2.1, so it is not a hidden flaw, but it does limit the practical conclusions.\n\nWho it is for: people working on measure-valued inverse problems, sparse source identification in transport, or greedy/conditional-gradient methods for PDE-constrained optimization. It deserves a serious referee. I would send it to review and ask for code, a baseline sensitivity study, and either a mesh-refinement argument for the clustering step or a clear discussion of when the heuristic fails. Accept conditional on those revisions; do not desk reject.","headline":"Solid engineering adaptation of PDAP/Radon-norm sparse inversion to contaminant source identification; the numerics are broad and plausible, but the reported accuracy is not covered by the convergence theory because of a heuristic cluster-merging post-processing step.","tokens_in":26898,"tokens_out":2746,"would_cite":true,"duration_ms":33663,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35R30","49J52","65K10","65M32"],"pacs":[],"model":"deepseek-v4-flash","headline":"A convex sparse-optimization reformulation turns contaminant source identification into a tractable problem whose minimizers are finitely many point sources, and a primal-dual-active-point algorithm finds them with only a handful of forward","keywords":["source identification","advection-diffusion equation","Radon measure","sparse regularization","primal-dual-active-point","PDE-constrained optimization","inverse problems","sensor data"],"falsifier":"Generate synthetic sensor data with the two-building benchmark using the true wind field, then run Algorithm 1 with a wind field perturbed by a 10% rotation or with a time-varying wind during the observation window; if the reconstructed source center moves by more than the benchmark's ~0.001 m error, the known-wind assumption is the load-bearing limitation.","tokens_in":26084,"feed_emoji":"💨","tokens_out":5426,"duration_ms":60688,"temperature":0.7,"pith_summary":"The paper tries to establish that locating an unknown number of airborne contaminant sources from scarce noisy sensor readings can be solved reliably by lifting the search from 'where are the sources?' to 'which positive measure explains the data?'. It shows that penalizing the total mass of that measure (a Radon-norm penalty) produces a convex problem that nevertheless has sparse minimizers: sums of finitely many Dirac deltas, with no more atoms than there are observations. The paper then adapts a primal-dual-active-point (PDAP) strategy that alternates between inserting candidate source locations where an adjoint field exceeds a threshold and re-optimizing intensities, pruning empty entries. On 2D and 3D benchmarks, including building geometries imported from an open map database, the method separates overlapping sources that L2-regularized reconstruction smears into one blob, and it does so with far fewer PDE solves. A sympathetic reader would care because this is a route to near-real-time source estimates for evacuation and emergency response.","feed_headline":"Sparse solver recovers contaminant sources from few sensors","feed_subtitle":"Recovers hidden contaminant sources from sparse sensor data in a few dozen PDE solves.","key_machinery":"The load-bearing object is the Radon-norm-regularized objective over the cone of positive measures, and the identity that makes it algorithmic is the pre-dual relation y^T F(mu) = ∫ phi^I d mu^I + ∫ phi^C d mu^C, where (phi^I, phi^C) = -F* y is the solution of the adjoint advection-diffusion problem with reversed wind. This identity converts the infinite-dimensional optimality condition into a finite certificate: a candidate measure is optimal iff the convolved adjoint fields stay below alpha everywhere and touch alpha at every support point. The Primal-Dual-Active-Point (PDAP) algorithm exploits the certificate by greedily inserting the global maximizer of the dual field, then solving a con","core_discovery":"The paper's central claim is that the conventionally hard, nonconvex problem of finding an unknown number of point-like contaminant sources can be relaxed without loss: model the source as a positive Radon measure mu and minimize J(mu)=1/(2σ^2)||F(mu)-d||^2 + alpha(mu_I(Ω)+mu_C(Ω)). Theorem 1 states that this convex problem has a minimizer supported on at most N_obs points, so the relaxed solution is exactly sparse, and a pair (mu_I, mu_C) is optimal precisely when the dual variables phi^I, phi^C (adjoint fields convolved with source shape functions) never exceed alpha and equal alpha on the support. Algorithm 1 realizes that characterization: compute the misfit, solve one adjoint problem to","pith_inferences":["If the wind field is uncertain, the dual field—and hence every candidate insertion—is computed under the wrong transport direction; a natural extension the paper does not develop is to make the wind an unknown or to average over an ensemble of wind fields, which would test how strongly the method depends on this input.","The discretization to mesh nodes induces clustering of near-optimal sources around true off-grid locations; the paper's heuristic barycentric merge is a post-process, not part of the optimization, so a sliding/refinement variant could remove this artifact and sharpen location estimates.","The method's reliance on only forward/adjoint solves and its insensitivity to added sensors suggest it could transfer to other linear wave-type inverse problems, such as acoustic or seismic source localization, where the same convex measure-lifting arguments apply.","Because the objective is convex and the certificate is explicit, the dual variables also give a principled criterion for placing new sensors: insert them where the dual field is still near its upper bound, connecting this work to optimal experimental design."],"forward_implications":["The relaxation is exact in the sense that at least one optimal reconstruction is a sum of finitely many Dirac delta sources, so sparsity is a theorem, not a heuristic.","Because the number of atoms never exceeds the number of observations, the method can in principle identify multiple distinct release points from few sensors; the two-source benchmark shows separation where L2 regularization yields a single smooth lump.","Computational cost scales mildly with data: each iteration requires at most two forward and one adjoint PDE solve, and in the reported tests adding sensors reduces the number of iterations.","The same framework handles initial-condition releases (instantaneous) and continuous source terms, with source shapes ranging from point Diracs to elliptic-PDE-defined profiles.","In the large plant-site test, all eight sources were located with a maximum error below 11 m on a roughly 500 m square domain using 27 iterations, suggesting practical emergency-response feasibility."],"fun_headline_variants":["Sparse solver finds contaminant sources from few sensors","PDAP: few sensors, exact source recovery in seconds","Rapid contaminant source ID from sparse measurements","Sparse source pinpointing with just a handful of detectors","Few sensor reads, full source map: PDAP algorithm"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole pipeline assumes the wind vector field is known, fixed, smooth, bounded, and divergence-free, so if the actual wind differs or changes over the measurement window, the adjoint-based candidate locations and final source estimates are unreliable.","fun_headline_variants_meta":{"raw":{"variants":["Sparse solver finds contaminant sources from few sensors","PDAP: few sensors, exact source recovery in seconds","Rapid contaminant source ID from sparse measurements","Sparse source pinpointing with just a handful of detectors","Few sensor reads, full source map: PDAP algorithm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1286,"prompt_tokens":752,"completion_tokens":534,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":456}},"tokens_in":496,"tokens_out":534,"duration_ms":7363,"temperature":1.0,"reasoning_tokens":456,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T00:07:54.937559+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic sensor data with the two-building benchmark using the true wind field, then run Algorithm 1 with a wind field perturbed by a 10% rotation or with a time-varying wind during the observation window; if the reconstructed source center moves by more than the benchmark's ~0.001 m error, the known-wind assumption is the load-bearing limitation.","supporting_citations":[],"review_version":1}