{"id":"69a9c420-acec-4992-aa8b-04bbe1d371ce","arxiv_id":"2507.09808","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A fully-corrective Frank-Wolfe method on measure spaces solves the infinite-dimensional volunteer-allocation problem for cardiac arrest response, with convergence under approximate subproblem solves and a finite-grid support result for Manhattan travel.","lead":"This paper develops a Frank-Wolfe optimization algorithm that works directly with probability-like measures, avoiding the need to break a city into a grid, to allocate volunteer responders for cardiac arrest emergencies. It proves convergence even when the inner optimization is only approximate and tests the method on synthetic cases and on Auckland, New Zealand.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (18) drops the 'x∈S' condition in \\bar B(y,t)=B(y,t)∩S, so the influence function is wrong on cvx(S)\\S and Section 5.2's negative centroid value does not follow from problem (P).","rationale":"The reader's rejection is grounded in a concrete mismatch between the stated variational problem and the derivative used by the algorithm. I checked the derivation: in (2) and in problem (P), the inner measure is μ(\\bar B(y,t)) = μ(B(y,t)∩S). The correct first-order term for ν=bδ_x is therefore b·1{x∈S, ||y-x||≤t}. Eq. (18) uses b·1{||y-x||≤t}, and on cvx(S)\\setminus S the two differ by a positive term. Section 5.2 evaluates h exactly at such a point, so the negative value at the centroid is an artifact of the dropped condition, not a property of the stated objective. Since Algorithm 1's support-selection step and the L1 finite-support theorem both use this h, the paper's central computational and structural claims are unsupported for the stated problem. This is a correctness defect, not a stylistic choice. The Theorem 2 summability assertion (Σ_k |h_{μ_k}(x*)|<∞) is also not implied by the displayed sum-of-squares argument, but it is secondary to the influence-function issue. If the authors intended the objective without the intersection with S, they need to state that objective and reconcile it with (2) and Theorem 3; as written, the two formulations conflict. The proposed test—recomputing h at the centroid with the x∈S indicator—would settle whether the Section 5.2 counterexample survives and hence whether the influence function is the derivative of the stated problem.","tokens_in":22286,"tokens_out":8869,"duration_ms":111730,"concrete_test":"Recompute the influence value at the centroid in Section 5.2 using the corrected indicator 1{x∈S and ||y_i-x||≤t}, which is identically zero for x=o because o∉S. If the resulting h(o) is nonnegative while Eq. (28) gives the negative value -0.0129·(b/3)e^{-b/3}, then Eq. (18) is not the von Mises derivative of problem (P) and the counterexample's conclusion is an artifact of the dropped x∈S condition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing object is the influence function h_mu in Theorem 7 / Eq. (18). The stated objective in (2) and in problem (P) uses \\bar B(y,t)=B(y,t)∩S. For a perturbation by bδ_x, the correct first-order term is (bδ_x-μ)(\\bar B(y,t)) = b·1{x∈S, ||y-x||≤t} - μ(\\bar B(y,t)). Eq. (18) instead writes b·1_{[||y-x||,∞)}(t), i.e., b·1{||y-x||≤t}, omitting the requirement x∈S. For every x∈cvx(S)\\setminus S, this makes the influence function too large by the nonnegative quantity b∫_S η(dy)∫_0^{||y-x||} exp(-μ(\\bar B(y,t))) dβ(t). The error is not cosmetic: Algorithm 1 selects support points by minimizing h_mu, so points outside S appear spuriously attractive. Section 5.2 is exactly this case: S is the three vertices, the centroid o lies in cvx(S)\\setminus S, and Eq. (28) uses the outside-S indicator. Under the stated objective, δ_o contributes to no \\bar B(y_i,t), so the correct influence at o is ∫η(dy)∫ μ(\\bar B(y,t)) exp(-μ(\\bar B(y,t))) dβ(t)>0, not a negative value. The claimed conclusion that the optimal measure is not supported on the demand points therefore does not follow, and the L1 finite-support theorem in Section 7 inherits the same mismatch. The fc-FW convergence results in Lemma 1 and Theorem 2 concern an algorithm whose subproblem is defined by the wrong variational object for the stated problem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper treats the problem of allocating a fixed total mass b of volunteers, modeled as the intensity measure of a spatial Poisson process, to minimize the expected probability of death from out-of-hospital cardiac arrest (OHCA). The authors formulate an infinite-dimensional convex optimization problem over measures supported on the convex hull of the incident region, derive an influence function as the first-order variational object, and analyze a fully-corrective Frank-Wolfe (fc-FW) algorithm that iteratively adds Dirac atoms. They prove structural properties (compactness of the feasible set, convexity of the objective, existence of optima), establish a sufficient-decrease lemma and a sublinear convergence rate under a weak subproblem condition, present numerical experiments on synthetic and Auckland data, and prove an L1-norm support-reduction result. The central claim is that the method solves the infinite-dimensional problem without discretization and that optimal measures have nontrivial support even for a few demand points.","tokens_in":22615,"tokens_out":26437,"duration_ms":296117,"significance":"If the results were correct, the paper would offer a useful extension of Frank-Wolfe methods to measure-valued optimization in an applied setting, with the attractive feature that the subproblem needs only a point of non-positive influence rather than a global minimizer. The city-scale numerical study and the L1-norm structural result would support practical relevance. However, the core variational object, the influence function, is incorrect for the stated problem, and the main analytic and numerical conclusions rely on that object. Consequently, the convergence results describe an algorithm that minimizes a different functional than the one in problem (P), and the claimed advances are not established for the stated problem.","major_comments":[{"comment":"Equation (18) defines h_µ(x) with the term -b I_{[∥y-x∥,∞)}(t) for every x∈cvx(S). Since \\bar B(y,t)=B(y,t)∩S in (2) and (P), the correct first-order term under the perturbation bδ_x is -b 1{x∈S} I_{[∥y-x∥,∞)}(t); for x∈cvx(S)\\S, the atom δ_x contributes zero to every µ(\\bar B(y,t)). Thus (18) is valid only on S, and the von Mises derivative identity (19)/(23) fails for ν with mass outside S. Because Algorithm 1 minimizes h_µk over all of cvx(S), the subproblem in Step 3 is not the variational problem for (P); consequently Lemma 1 and Theorem 2 concern a different algorithm, and the claimed O(1/√k) convergence is not a statement about the stated problem.","section":"Theorem 7, Eq. (18)"},{"comment":"Section 5.2 evaluates the influence at the centroid o∈cvx(S)\\S using the incorrect Eq. (18). The correct influence at o is ∫_S η(dy)∫_0^∞ \\tilde µ(\\bar B(y,t)) exp(-\\tilde µ(\\bar B(y,t))) dβ(t) > 0, since the -b indicator term is absent. The negative value in (28) is an artifact of the missing x∈S condition. Moreover, for the objective (2), any mass placed at x∉S contributes to no \\bar B(y,t); moving it to a point of S with positive λ_i strictly decreases J, so an optimal measure for (P) is supported on S. The conclusion that supp(µ*) is not contained in {y1,y2,y3} is therefore not just unproved but incompatible with the stated model.","section":"§5.2, Eq. (28)"},{"comment":"Section 7 repeats the same error: the influence function (34) is only valid for x∈S. For x in the interior of a rectangle Sj,k that does not meet the finite demand set S, the correct h_µ(x) is constant in x. Therefore Lemma 3's strict concavity inequality (36) is false (both sides are equal), Lemma 4's claim that minimizers lie at rectangle vertices is false (a constant is minimized everywhere), and Theorem 10's proof collapses. The support statement for discrete η may be true for the trivial reason that mass off S is wasted, but the derivation via strict concavity and the claimed vertex look-up / support-reduction procedure are not valid for problem (P).","section":"§7, Lemma 3 and Theorem 10"},{"comment":"All numerical experiments in Section 6 use the influence function in the form (29), which is derived from the incorrect (18). The reported optimal measures with mass in the interior of cvx(S) cannot be optimal for (P), since such mass is absent from every \\bar B(y,t); the contour plots of h in Figures 2(b) and 2(d) certify optimality with respect to the wrong functional. The implementation therefore does not solve the problem introduced in (P), and the claims of 'complex solution structure' and 'scalability' are not supported by the experiments.","section":"§6, Algorithm 1 implementation"}],"minor_comments":[{"comment":"The proof says 'J is convex by Theorem 3'; the correct reference for convexity is Theorem 5.","section":"Theorem 6 proof"},{"comment":"The text says 'Theorem 5 guarantees that h_µk(x*) → 0'; the intended reference is Theorem 2.","section":"§3.2, item (b)"},{"comment":"The phrase 'Theorem 4 directly implies' should read 'Lemma 4'.","section":"Theorem 10 proof"},{"comment":"The notation δ_x := IA(x) is ambiguous; it should be the set function δ_x(A) = 1{x∈A}.","section":"Definition 3"},{"comment":"The text refers to 'Figure 2' before any figure is introduced; the intended reference appears to be Figure 1, the triangle diagram.","section":"§5.2"},{"comment":"The paper relies on [33] for Theorem 1, Theorem 8, Lemma 2, and Lemma 6 but does not clearly identify these as imported results; a remark at the start of Section 3.2 would remove ambiguity.","section":"§3.2"}],"recommendation":"reject","confidential_remarks":"The stress-test concern is real and load-bearing. The influence-function error cannot be patched without discarding the paper's main examples and the L1 structural result, which suggests rejection is appropriate rather than minor or major revision. I also note the unusually heavy reliance on the authors' own preprint [33] for the optimality conditions and smoothness inequality, though this is not by itself a reason to reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central claim of this paper doesn't hold up. The influence function in Theorem 7, Eq. (18), drops the x∈S condition that appears in \\bar B(y,t)=B(y,t)∩S. For a perturbation bδ_x, the correct first-order term is b·1{x∈S, ||y-x||≤t} - µ(\\bar B(y,t)); Eq. (18) writes b·1_{[||y-x||,∞)}(t), omitting x∈S. For any x in cvx(S)\\S, the influence function is too large by a nonnegative amount. The algorithm selects new support points by minimizing h_µ, so points outside S look spuriously attractive. Section 5.2's negative influence at the centroid is exactly this artifact: the centroid lies outside S, so the correct influence is positive. The claimed conclusion that the optimal measure is not supported on the demand points therefore doesn't follow from problem (P). The L1 finite-support theorem inherits the same mismatch.\n\nWhat's genuinely new and good: the fc-FW convergence analysis (Lemma 1, Theorem 2) is abstract and, as far as I can tell, correct. It shows a sufficient decrease from h(x*)≤0 alone, with an O(1/√k) rate on the influence values. That's a useful contribution independent of the specific application. The restriction-to-cvx(S) argument (Theorem 3) and the convexity/existence results are standard but fine, though the existence proof should verify lower semicontinuity carefully. The L1 piecewise-concavity observation is interesting, and the Auckland experiments suggest the method scales; but no code or data are included, and there is no baseline comparison.\n\nMinor issue: Theorem 2 claims ∑|h|<∞, but the proof gives only ∑h²<∞. That's an overclaim, though it doesn't affect the main rate.\n\nThis is a repairable paper. The fix is either to remove the ∩S from the objective (so volunteers everywhere count) or to add the x∈S indicator in the influence function. But as submitted, the derivation breaks the central claim, and the numerical results in Sections 5.2 and 6 are not evidence for problem (P).\n\nIf I were the editor, I'd send it to review because the fc-FW theorem deserves airing, and the authors may well fix the derivative. But I'd expect major revision at best. For a reading group, it's a good case study in how a dropped indicator can sink an infinite-dimensional optimization paper; otherwise it's not something I'd cite yet.","headline":"The paper's influence function drops the x∈S condition from the stated objective, so the central claim and the numerical evidence for it do not hold; the fc-FW convergence analysis is still a useful standalone contribution.","tokens_in":23224,"tokens_out":5044,"would_cite":false,"duration_ms":51102,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C48","90C25"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a fully-corrective Frank-Wolfe algorithm on the infinite-dimensional space of measures solves the OHCA volunteer-allocation problem, with $O(1/\\sqrt{k})$ decay of influence values at selected locations even when…","keywords":["measure optimization","Frank-Wolfe recursion","fully-corrective algorithm","influence function","von Mises derivative","out-of-hospital cardiac arrest","emergency response","P-means"],"falsifier":"Take the three-vertex equilateral example of Section 5.2 and compute the directional derivative at the centroid $o$ by finite differences, $(J(\\mu+t(b\\delta_o-\\mu))-J(\\mu))/t$ for small $t$, comparing it with Eq. (18); if the two disagree for points outside the demand set, the influence function used by the algorithm is not the derivative the paper derives, and the descent and convergence guarantees as stated do not govern those iterates.","tokens_in":22009,"feed_emoji":"🚑","tokens_out":15150,"duration_ms":161110,"temperature":0.7,"pith_summary":"The paper considers the problem of choosing a spatial probability distribution of volunteers over a city to minimize the probability of death from out-of-hospital cardiac arrest. It claims that this optimum can be found by a fully-corrective Frank-Wolfe algorithm that works directly in the infinite-dimensional space of measures, without first discretizing the city or imposing a parametric family. The central theoretical result is that even when the per-iteration task of minimizing the influence function—the location-wise first-order score—is solved only approximately, the influence values at the selected locations tend to zero at rate $O(1/\\sqrt{k})$, and if solved globally the rate improves to $O(1/k)$. The paper also proves the objective is convex, the feasible set is compact, an optimum exists, and that under the $L_1$ travel metric with discrete demand the optimal measure is supported on a finite grid. If these claims are right, volunteer recruitment targets for cities can be computed directly on continuous maps, and the method transfers to the broader class of $P$-means problems.","feed_headline":"O(1/√k) Frank-Wolfe solves volunteer allocation without grids","feed_subtitle":"Converges at O(1/√k) even when influence-function searches are local; L1 travel pins support to a grid.","key_machinery":"The machinery is the influence function together with the fully-corrective Frank-Wolfe recursion it drives. For a measure $\\mu$, the influence function $h_\\mu(x)$ is the location-wise rate at which $J$ changes when a small point mass is moved to $x$, and Eq. (18) gives its closed form. The fc-FW algorithm keeps a growing set of scaled Dirac atoms, adds an atom at a point $x^*(\\mu_k)$ with $h_{\\mu_k}(x^*(\\mu_k))\\le 0$, and then solves the finite-dimensional convex problem of minimizing $J$ over the convex hull of all accumulated atoms. Lemma 1 shows that $L$-smoothness plus the non-positive-influence condition yields a sufficient decrease of the objective at every step, and Theorem 2 turns this into influence values decaying as $O(1/\\sqrt{k})$; when the influence subproblem is solved globally, the objective value converges at $O(1/k)$. For the $L_1$ metric, Lemma 3 shows the influence function is strictly concave on each rectangle of the demand-coordinate grid, which forces its minima, and hence the optimal support, onto the grid vertices.","core_discovery":"On the paper's own terms, the discovery is that the infinite-dimensional emergency-response problem (P), minimize $J(\\mu)=\\int_S \\eta(dy)\\int_0^\\infty \\exp(-\\mu(B(y,t)\\cap S))\\,d\\beta(t)$ over nonnegative Borel measures with total mass $b$, is tractable by adaptive particle accumulation rather than a priori discretization. The paper proves that $J$ is convex and $L$-smooth in total variation, that the feasible set is weakly compact, and that optimal measures exist; the first-order object is the influence function $h_\\mu(x)$, which Theorem 7 identifies as the von Mises derivative of $J$. The algorithmic claim is that a fully-corrective Frank-Wolfe recursion—add the atom $b\\delta_x$ at a point where $h_{\\mu_k}(x)\\le 0$, then re-optimize the weights over all accumulated atoms—produces a convergent sequence: $h_{\\mu_k}(x^*_k)\\to 0$ at rate $O(1/\\sqrt{k})$, with no need to globally minimize the nonconvex influence function. The numerical portion claims that even simple three- or four-point instances yield non-obvious optimal measures, that continuous and Auckland-scale cases run in hundreds to thousands of iterations, and that with the $L_1$ metric the influence function is piecewise strictly concave, so the optimal support lies on the grid built from demand coordinates.","pith_inferences":["An extension the paper leaves implicit: for candidate volunteer locations outside the incident region, the factor counting whether $x$ lies in $B(y,t)\\cap S$ should be zero; rerunning the three- and four-point experiments with that corrected derivative would test whether the reported off-demand-point support persists.","The proof machinery is not tied to cardiac arrest: any convex, $L$-smooth objective on measures with a computable influence function should admit the same fully-corrective Frank-Wolfe treatment, making the algorithm a candidate for other spatial resource-allocation and coverage problems.","The $L_1$ grid-support theorem suggests a practical speedup the paper does not test: restrict each influence-function search to the $O(n^2)$ grid vertices and benchmark it against the continuous Adam-based searches on the Auckland data."],"forward_implications":["City-scale volunteer allocation can be solved directly on continuous maps, so optimality no longer depends on a chosen area-unit discretization.","At each iteration any point with non-positive influence suffices; a global solve of the nonconvex influence subproblem is not needed for the $O(1/\\sqrt{k})$ guarantee.","If the influence subproblem is solved globally, the objective-value convergence improves to $O(1/k)$, recovering the standard Frank-Wolfe rate in the infinite-dimensional setting.","With $L_1$ travel and discrete incident locations, the optimal volunteer measure is concentrated on the $O(n^2)$ grid vertices determined by demand coordinates, enabling lookup-style algorithms.","The framework transfers to $P$-means problems, whose objectives share the same exponential-of-negative-measure form."],"supporting_citations":[{"why":"Supplies the deterministic Frank-Wolfe recursion, its $O(1/k)$ complexity theorem, and the optimality and support lemmas that the paper extends to fully-corrective updates.","marker":"[33]"},{"why":"Provides the OHCA stochastic model, the Auckland area-unit incident rates, and the prior finite-dimensionalization that the paper aims to avoid.","marker":"[29]"},{"why":"Supplies the spatial Poisson process machinery and the void-probability formula used to derive the objective in Eq. (2).","marker":"[26]"},{"why":"Provides the death-probability function $\\beta$ used in the numerical experiments and in the three-point centroid calculation.","marker":"[8]"},{"why":"Grounds the definition of the influence function as a von Mises derivative of a statistical functional.","marker":"[12]"},{"why":"Supplies the dimension-reduction viewpoint from experimental design that motivates solving measure optimization through first-order variational objects.","marker":"[17]"},{"why":"Identifies the $P$-means problem class to which the paper claims the framework and analysis extend.","marker":"[20]"},{"why":"Provides the spatial-tessellation background for the $P$-means connection and the support structure of optimal measures.","marker":"[23]"}],"fun_headline_variants":["Frank-Wolfe on measure spaces optimizes cardiac arrest response","Infinite-dim Frank-Wolfe: no grid, O(1/√k) volunteer allocation","Adaptive particles beat fixed grids for emergency response","Measure-space optimizer allocates volunteers without discretization","O(1/√k) FW solves rescue placement without grids"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the influence function $h_\\mu(x)$ in Eq. (18) is the true first-order derivative of the death-probability objective for every candidate volunteer location in the convex hull of the incident region; the derivation counts a trial mass at $x$ toward every ball $B(y,t)$ when $\\|y-x\\|\\le t$, even when $x$ lies outside the incident region, where that mass cannot intersect $B(y,t)\\cap S$.","fun_headline_variants_meta":{"raw":{"variants":["Frank-Wolfe on measure spaces optimizes cardiac arrest response","Infinite-dim Frank-Wolfe: no grid, O(1/√k) volunteer allocation","Adaptive particles beat fixed grids for emergency response","Measure-space optimizer allocates volunteers without discretization","O(1/√k) FW solves rescue placement without grids"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001653,"raw_usage":{"total_tokens":6629,"prompt_tokens":1078,"completion_tokens":5551,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":5462}},"tokens_in":694,"tokens_out":5551,"duration_ms":47503,"temperature":1.0,"reasoning_tokens":5462,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:52:34.418939+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the three-vertex equilateral example of Section 5.2 and compute the directional derivative at the centroid $o$ by finite differences, $(J(\\mu+t(b\\delta_o-\\mu))-J(\\mu))/t$ for small $t$, comparing it with Eq. (18); if the two disagree for points outside the demand set, the influence function used by the algorithm is not the derivative the paper derives, and the descent and convergence guarantees as stated do not govern those iterates.","supporting_citations":[{"cited_title":"URL https://arxiv.org/abs/2407.00307","cited_arxiv_id":null,"evidence_quote":"Supplies the deterministic Frank-Wolfe recursion, its $O(1/k)$ complexity theorem, and the optimality and support lemmas that the paper extends to fully-corrective updates."},{"cited_title":"Management Science","cited_arxiv_id":null,"evidence_quote":"Provides the OHCA stochastic model, the Auckland area-unit incident rates, and the prior finite-dimensionalization that the paper aims to avoid."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the spatial Poisson process machinery and the void-probability formula used to derive the objective in Eq. (2)."},{"cited_title":"(2003) Optimal defibrillation response intervals for maximum out-of-hospital cardiac arrest survival rates","cited_arxiv_id":null,"evidence_quote":"Provides the death-probability function $\\beta$ used in the numerical experiments and in the three-point centroid calculation."},{"cited_title":"Von Mises Calculus for Statistical Functionals 19","cited_arxiv_id":null,"evidence_quote":"Grounds the definition of the influence function as a von Mises derivative of a statistical functional."},{"cited_title":"The Annals of Statistics 2(5):849–879","cited_arxiv_id":null,"evidence_quote":"Supplies the dimension-reduction viewpoint from experimental design that motivates solving measure optimization through first-order variational objects."},{"cited_title":"Statistics and Computing 12:115–123","cited_arxiv_id":null,"evidence_quote":"Identifies the $P$-means problem class to which the paper claims the framework and analysis extend."},{"cited_title":"Wiley series in probability and statistics","cited_arxiv_id":null,"evidence_quote":"Provides the spatial-tessellation background for the $P$-means connection and the support structure of optimal measures."}],"review_version":1}