{"id":"c0029992-f96c-4ea6-aef0-9d6144e0e97c","arxiv_id":"2501.18558","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A log-Gaussian Cox process driven by Whittle-Matérn fields is defined on any compact metric graph, with scalable quadrature-based likelihood inference and proven Hellinger convergence rates.","lead":"This paper introduces a way to model event locations, such as road accidents, directly on a road network treated as a connected metric graph, using a flexible Gaussian random field as the log-intensity. It also proves that a fast numerical approximation of the model's likelihood converges to the exact posterior as the computational mesh is refined, making city-scale Bayesian inference practical.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of Theorem C.1 does not yet establish the claimed O(p^{-γ}) rate for the actual midpoint rule: the quadrature weights are defined inconsistently, and the Assumption A.1(ii) verification uses a false upper bound.","rationale":"The reader's verdict is CONDITIONAL, and the same printed-proof defects were flagged in the reader's rationale: a false bound in the verification of Assumption A.1(ii) and an inconsistent definition of quadrature weights. My stress-test confirms these are the most load-bearing issues for the central claim that the midpoint-rule likelihood approximation yields posterior convergence rates O(p^{-γ}) and O(p^{-1}). They are genuine gaps in the proof as written, but they are local and repairable: the potential upper bound can be fixed by replacing ||u||_C with exp(||u||_C), and the quadrature weights can be set to the interval lengths, which is evidently the intended midpoint rule. The regularity assumptions (Assumption B.3 and Propositions B.4/B.5) are the other input to the theorem; the reader identified these as the weakest assumption, but I judge them less problematic than the proof defects because they are clearly stated and plausibly inherited from prior work. I also note a smaller ambiguity: Theorem 3.1 phrases the rate in terms of the total number of integration points p, whereas Theorem C.1 states it in terms of ||p|| = min_e p_e; these are comparable if the partition is uniform across edges, but the connection should be stated explicitly. The application and the model construction are valuable, and the independent implementation in MetricGraph is a real strength. Overall, the central claim is likely correct, but the convergence theorem needs correction before the rates can be regarded as fully proven; this matches the reader's CONDITIONAL verdict, so no change is needed.","tokens_in":18178,"tokens_out":12760,"duration_ms":140274,"concrete_test":"Re-derive the quadrature error bound in Theorem C.1 with weights w_{i,e} = |q_{i+1,e} - q_{i,e}| (interval lengths) and verify whether the Hellinger bound O(||p||^{-γ}) still follows. Then, with the printed weights w_i = |s~_{i+1,e} - s~_{i,e}|, compute the additional error term for a single edge with f(t) = exp(t) on a two-interval partition; if this term does not decay as p^{-γ}, the theorem's stated scheme does not achieve the claimed rate. Separately, re-derive Assumption A.1(ii) using the bound ∫_e exp(u_e)dt ≤ |Γ| exp(||u||_{C(Γ)}) and confirm the uniformity constant is independent of p.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Theorem C.1, the quadrature weights are defined as w_i = |s~_{i+1,e} - s~_{i,e}|, i.e., distances between consecutive evaluation points, but the error bound that follows applies Lemma B.1/B.2 to intervals [q_{i,e}, q_{i+1,e}] with evaluation point s~_{i,e} inside. For the midpoint rule described in Section 3 and used in the implementation, the weights must be the interval lengths |q_{i+1,e} - q_{i,e}|. With the printed weights, the telescoping identity Σ_i ∫_{q_i}^{q_{i+1}} |f(t)-f(s~_i)|dt ≥ |∫_e f(t)dt - Σ_i w_i f(s~_i)| is not valid: an extra term Σ_i (|q_{i+1}-q_i| - w_i) f(s~_i) appears and is not controlled. Thus the theorem's conclusion, as stated, applies to a different quadrature scheme than the one whose convergence is claimed. A second, independent defect appears in the verification of Assumption A.1(ii): the proof bounds ∫_e exp(u_e(t))dt by |Γ| ||u||_{C(Γ)}, which is false for positive u; the correct bound is |Γ| exp(||u||_{C(Γ)}). Since this upper bound is needed for the uniformity constant L(r) in Theorem A.2, the printed proof is incomplete. Both defects are repairable, but until corrected the stated rates are not established for the implemented midpoint-rule likelihood approximation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a class of log-Gaussian Cox processes on compact metric graphs, built on Whittle–Matérn fields defined as solutions to fractional-order SPDEs. The main contributions are: (i) a proof that these processes are well defined on any compact metric graph; (ii) a likelihood approximation based on a midpoint quadrature rule that avoids approximating the Gaussian field, together with claimed Hellinger convergence rates for the resulting posterior as the number of integration points grows (O(p^{-γ}) for α=1 and O(p^{-1}) for α=2); (iii) an implementation in the MetricGraph R package with R-INLA; and (iv) an application to traffic accident data from Al-Ahsa, Saudi Arabia. The central theoretical result is Theorem 3.1, whose proof is deferred to Appendix C and relies on an abstract posterior-approximation theorem of Cotter–Dashti–Stuart.","tokens_in":34,"tokens_out":7688,"duration_ms":138978,"significance":"If the theoretical claims are established, the paper makes a substantial contribution: it provides the first class of log-Gaussian Cox processes on arbitrary compact metric graphs, offers a scalable likelihood-based inference method that preserves the exactness of finite-dimensional distributions for integer smoothness, gives explicit posterior convergence rates, and demonstrates the methodology on a large real network. The open-source implementation and the excursion-set analysis are useful practical additions. However, the proofs as printed contain load-bearing gaps in the verification of the conditions needed for the convergence theorem; these gaps are local and appear repairable, but until they are fixed the main theoretical guarantee is not established.","major_comments":[{"comment":"The quadrature weights are defined inconsistently with the claimed error bound. The theorem sets w_i = |s~_{i+1,e} - s~_{i,e}|, the distance between consecutive evaluation points, whereas the proof bounds the quadrature error by ∫_{q_i}^{q_{i+1}} |f(t) - f(s~_{i,e})| dt, which is only valid when the weights are the interval lengths |q_{i+1,e} - q_{i,e}|. With the printed weights, the identity |∫_e f(t) dt - Σ_i w_i f(s~_{i,e})| ≤ Σ_i ∫_{q_i}^{q_{i+1}} |f(t) - f(s~_{i,e})| dt does not hold: an additional term Σ_i (|q_{i+1,e} - q_{i,e}| - w_i) f(s~_{i,e}) appears and is not controlled. Since Section 3 and the implementation use the midpoint rule with interval-length weights, Theorem C.1 as stated does not prove the claimed convergence rate for the implemented method. The weight definition and the corresponding step in the proof must be corrected.","section":"Appendix C, Theorem C.1"},{"comment":"The proof of the upper bound for the potential uses the inequality ∫_e exp{u_e(t)} dt ≤ |E||Γ| ||u||_{C(Γ)}, which is false; for example, if u ≡ 0 the left-hand side is |E||Γ| while the right-hand side is 0. The correct bound is |E||Γ| exp(||u||_{C(Γ)}), which still yields a finite constant L(r) for ||u|| < r. As printed, condition (ii) of Assumption A.1 is not established, so the application of Theorem A.2 is not justified. This is repairable, but the inequality must be corrected.","section":"Appendix C, verification of Assumption A.1(ii)"},{"comment":"Assumption B.3(ii) is vacuous for the α=2 case used in Theorem C.1. Since α~ = min{α−1/2, 1/2} = 1/2 when α=2, the condition γ∈(0, α~−1) reads γ∈(0, −1/2), which is empty. Thus the assumed regularity τ_e∈C^{1,γ}(e) cannot hold for any γ, and the proof's assertion that μ0(eC^1(Γ)) = 1 for α=2 does not follow from Assumption B.3 as stated. The intended condition is presumably γ∈(0, α−3/2), consistent with Proposition B.5(ii); please correct it and check the consistency of all statements that refer to this assumption.","section":"Appendix B, Assumption B.3(ii)"}],"minor_comments":[{"comment":"The phrase 'with s~_{0,e} = 0' is confusing: s~_{0,e} is not used elsewhere and the partition starts at q_{1,e}. Either delete it or state explicitly what it refers to.","section":"Appendix C, Theorem C.1"},{"comment":"The sentence 'there exists a constant K, independent of p_e, such that there exist K1, K2 > 0 such that ...' introduces a constant K that is never used; it should be removed or used in the bounds.","section":"Appendix C, Theorem C.1"},{"comment":"The notation e~a_i (and e~a) is visually awkward and could be confused with an exponential. A clearer notation such as h_i or w_i for quadrature weights would improve readability, especially since Theorem C.1 uses w_i.","section":"Section 3, equation (7)"},{"comment":"It would be helpful to state explicitly that, for the midpoint rule, the weights in (7) are the lengths of the mesh intervals, linking the implementation to the corrected weight definition in Theorem C.1.","section":"Section 3, paragraph on likelihood approximation"},{"comment":"The paper relies on results from the preprint [11] for sample-path regularity. Since [11] is not yet peer-reviewed, and since Assumption B.3 currently contains an inconsistency, the authors should ensure that the statements imported from [11] are reproduced correctly and that the dependence on [11] is clearly flagged.","section":"Appendix B, Proposition B.4 and B.5"}],"recommendation":"major_revision","confidential_remarks":"The core construction is plausible and the application is compelling, but the proof of the central convergence theorem is not sound as printed: the weight mismatch in Theorem C.1 means the stated rates are for a different quadrature than the one implemented, and the verification of Assumption A.1(ii) uses a false bound. Both are repairable. The Assumption B.3(ii) issue appears to be a typo but renders the α=2 case vacuous as stated. I recommend major revision rather than rejection because the fixes are local and the overall framework is valuable. The paper's novelty relative to the authors' prior work on Whittle–Matérn fields is mostly the LGCP extension and the convergence analysis, which is acceptable; the discussion could do more to position this relative to the existing linear-network LGCP literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's what you should know about arXiv:2501.18558. It's a solid, useful paper with a genuine new method and a repairable proof hole. The model itself is a composition of existing pieces—Whittle–Matérn fields on metric graphs plus the standard LGCP construction—so the novelty is not in the model class. What's new is the inference scheme: they approximate the likelihood integral with a midpoint rule but keep the latent field exact, using the Markov property for integer smoothness to get sparse precision matrices. That's a real improvement over Simpson et al.'s finite-element approximation, and it gives the first posterior convergence rates for this kind of LGCP on general compact metric graphs. The Al-Ahsa accident application is large (178k segments) and runs in about a minute, which makes the scalability claim concrete.\n\nThe soft spots are exactly where the stress-test note lands. In Theorem C.1, the quadrature weights are defined as distances between consecutive evaluation points, w_i = |s~_{i+1} - s~_i|, but the error analysis uses intervals [q_i, q_{i+1}] with evaluation point inside. The telescoping identity only works if the weights are the interval lengths. So the theorem, as stated, proves convergence for a different quadrature rule than the midpoint rule used in the implementation. That's an inconsistency in the statement, not a fatal flaw in the idea—the fix is to set w_i = |q_{i+1} - q_i| and the same proof goes through.\n\nThe second defect is in verifying Assumption A.1(ii): the proof bounds ∫_e exp(u_e(t)) dt by |Γ| ||u||_{C(Γ)}, which is false for positive u. The correct bound is |Γ| exp(||u||_{C(Γ)}). This is a one-line repair and doesn't change the rates, but as printed the uniform bound L(r) is not established.\n\nBoth issues are repairable, and I would expect them to be fixed in revision. The application is not fully reproducible: the accident data are proprietary, no scripts are shipped, and the study area was manually delineated. That's common in applied work, but it does temper the practical claims. The citation pattern is fine; the authors cite their own prior work on the fields, which is appropriate since they rely on it.\n\nWho is this for? Spatial statisticians working on network point patterns, and anyone doing large-scale LGCP inference with INLA. They'll get a useful method and a clear statement of rates, once the proof is cleaned up. I'd send it out for peer review. The defects are real but localized; a competent referee can verify the repairs quickly. I'd read it closely myself.","headline":"A genuinely useful inference scheme for LGCPs on metric graphs, with two repairable proof bugs in the main convergence theorem.","tokens_in":19037,"tokens_out":2675,"would_cite":true,"duration_ms":25645,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M30","60G60","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Log-Gaussian Cox processes are well-defined on every compact metric graph, and a midpoint-rule likelihood approximation drives the Hellinger distance between true and approximate posteriors to zero at a known rate.","keywords":["Gaussian random field","Linear network","Log-Gaussian Cox process","Metric graph","Spatial point process","Stochastic partial differential equation","Bayesian inference","Excursion sets"],"falsifier":"Recompute the posterior for a Whittle–Matérn log-Gaussian Cox process on a compact metric graph with $\\alpha=1$ or $\\alpha=2$ using increasing numbers of integration points $p$, and measure the Hellinger distance to a reference posterior from a very fine quadrature or an independent exact sampler. If the distance plateaus instead of decaying like $p^{-\\gamma}$ or $p^{-1}$, the assumptions are violated. A sharper check is to build a graph and a non-constant $\\tau$ satisfying only part (i) of Assumption B.3 but not part (ii), and test whether $\\alpha=2$ solutions lose the $C^1$ modification the proof requires.","tokens_in":17924,"feed_emoji":"🛣️","tokens_out":9253,"duration_ms":97642,"temperature":0.7,"pith_summary":"The paper extends log-Gaussian Cox processes from Euclidean domains and linear networks to any compact metric graph, such as an arbitrary road network. The random intensity is driven by a Gaussian Whittle–Matérn field, defined as the solution of a fractional stochastic differential equation on the graph, which makes the point process well-defined and can produce differentiable intensities when the smoothness parameter is large enough. Inference uses a midpoint rule for the integral in the likelihood while evaluating the latent field exactly, so the Gaussian process is never approximated. The central theoretical result is that the posterior from this approximate likelihood converges to the true posterior in Hellinger distance at rate $O(p^{-\\gamma})$ for $\\alpha=1$ and at rate $O(p^{-1})$ for $\\alpha=2$, where $p$ is the number of integration points. The paper demonstrates the method on a road-accident data set with over 150,000 road segments and identifies high-risk segments through excursion probabilities.","feed_headline":"Point processes on road networks get a provable error rate","feed_subtitle":"Approximating only the integral in the likelihood gives provable convergence and scales to 150,000 road segments.","key_machinery":"The central object is the Gaussian Whittle–Matérn field on a metric graph, defined as the solution to $(\\kappa^2-\\Delta_\\Gamma)^{\\alpha/2}(\\tau u)=\\mathcal{W}$, where $\\Delta_\\Gamma$ is the Kirchhoff–Laplacian and $\\mathcal{W}$ is Gaussian white noise. This spectral definition gives a well-defined Gaussian process on arbitrary compact metric graphs, with sample-path regularity controlled by $\\alpha$. The likelihood approximation is a midpoint rule: the integral over $\\Gamma$ in the point-process likelihood is replaced by a weighted sum at integration points, while the latent field is evaluated exactly at those points. For integer $\\alpha$, the finite-dimensional distributions of the field are computed exactly and sparsely by adding observation and integration locations as degree-2 vertices and exploiting the field's Markov property. Theorem A.2, imported from the Bayesian inverse-problem literature, converts a bound on the quadrature error of the potential into the stated Hellinger convergence rates.","core_discovery":"The paper's central claim is that a log-Gaussian Cox process with intensity $\\Lambda(s)=\\exp(m(s)+u(s))$, where $u$ is a Gaussian Whittle–Matérn field on a compact metric graph $\\Gamma$, is a valid point process model on that graph, and that likelihood-based Bayesian inference for it can be both exact and scalable. Exactness comes from evaluating the finite-dimensional distributions of $u$ at the observation and integration points, which is possible in closed form when the smoothness $\\alpha$ is a positive integer because the fields are Markov. Scalability comes from replacing the intractable integral of $\\exp(u)$ over $\\Gamma$ by a midpoint quadrature with $p$ integration points, leaving the field itself untouched. Theorem 3.1 then states that the Hellinger distance between the true posterior and the approximate posterior is $O(p^{-\\gamma})$ for any $0<\\gamma<1/2$ when $\\alpha=1$, and $O(p^{-1})$ when $\\alpha=2$. The paper calls this the first complete convergence proof for a likelihood approximation in an SPDE-driven log-Gaussian Cox process.","pith_inferences":["Extending the convergence rates to non-integer $\\alpha$ will likely require combining the midpoint likelihood with a finite-element or rational approximation of the fractional Laplacian; the paper lists this as future work, and the resulting rates would have to account for the additional approximation error.","The same proof template—exact finite-dimensional field evaluations plus midpoint quadrature plus a potential bound—could transfer to other SPDE-driven point process models, such as space-time log-Gaussian Cox processes, whenever the required sample-path regularity is available.","The weighted-Kirchhoff variant mentioned in the discussion would change vertex conditions to reflect directional traffic flow; if implemented, the expected suppression of intensity near high-degree intersections would disappear, so hotspot maps on road networks might shift away from junctions.","Readers should treat the reported hotspot maps as conditional on the chosen threshold and mesh spacing; the paper does not report sensitivity of excursion sets to those tuning choices."],"forward_implications":["Any compact metric graph, not only linear networks or graphs with Euclidean edges, can now support a log-Gaussian Cox process with a valid stochastic intensity.","For integer smoothness, Bayesian inference no longer requires approximating the latent field: only the likelihood integral is discretized, and the posterior error is controlled by an explicit rate.","Setting $\\alpha=2$ yields differentiable intensities, which isotropic covariance models on networks cannot provide, and the posterior approximation converges at the faster $O(p^{-1})$ rate.","The method scales to city-sized networks: the accident application fits the model on a graph with more than 150,000 road segments in roughly one minute per model.","Excursion sets computed from the fitted posterior can map road segments where the latent risk is significantly elevated, turning the point-process fit into hotspot localization."],"supporting_citations":[{"why":"Defines Gaussian Whittle–Matérn fields on metric graphs and establishes existence, uniqueness, and the sample-path regularity used throughout the paper.","marker":"[9]"},{"why":"Supplies the exact computation of finite-dimensional distributions and precision matrices for Whittle–Matérn fields with integer smoothness, making the likelihood evaluation exact.","marker":"[7]"},{"why":"Provides the Markov properties of Gaussian random fields on compact metric graphs that give sparse precision matrices for integer $\\alpha$.","marker":"[10]"},{"why":"Gives the regularity results and the variance-stationary construction for non-constant $\\tau$, which underpin Assumption B.3 and Propositions B.4 and B.5.","marker":"[11]"},{"why":"Provides the general theorem on approximation of Bayesian inverse problems that the proof uses to turn potential error into Hellinger-distance error.","marker":"[15]"},{"why":"Introduced the midpoint-rule likelihood approximation for log-Gaussian Cox processes that this paper adapts to metric graphs without approximating the Gaussian process.","marker":"[33]"},{"why":"Earlier Cox-process models on linear networks; the comparison shows why a general metric-graph formulation is needed and why isotropy limits scalability.","marker":"[24]"}],"fun_headline_variants":["Point processes on road networks get provable accuracy","Scalable Bayesian point processes on 150k road segments","Exact inference for log-Gaussian Cox processes on graphs","Provable error bounds for likelihood approximation on metric graphs","Road-accident hotspots found with guaranteed convergence rates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof needs the latent field to have smooth enough sample paths on the graph—Hölder continuous when $\\alpha=1$ and continuously differentiable when $\\alpha=2$—which the paper guarantees only under Assumption B.3 on $\\tau$ (constant or variance-stationary); if that regularity fails on some compact graph, the stated rates do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Point processes on road networks get provable accuracy","Scalable Bayesian point processes on 150k road segments","Exact inference for log-Gaussian Cox processes on graphs","Provable error bounds for likelihood approximation on metric graphs","Road-accident hotspots found with guaranteed convergence rates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1355,"prompt_tokens":978,"completion_tokens":377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":300}},"tokens_in":594,"tokens_out":377,"duration_ms":4048,"temperature":1.0,"reasoning_tokens":300,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T23:00:36.749290+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the posterior for a Whittle–Matérn log-Gaussian Cox process on a compact metric graph with $\\alpha=1$ or $\\alpha=2$ using increasing numbers of integration points $p$, and measure the Hellinger distance to a reference posterior from a very fine quadrature or an independent exact sampler. If the distance plateaus instead of decaying like $p^{-\\gamma}$ or $p^{-1}$, the assumptions are violated. A sharper check is to build a graph and a non-constant $\\tau$ satisfying only part (i) of Assumption B.3 but not part (ii), and test whether $\\alpha=2$ solutions lose the $C^1$ modification the proof requires.","supporting_citations":[{"cited_title":"Bolin, A","cited_arxiv_id":null,"evidence_quote":"Defines Gaussian Whittle–Matérn fields on metric graphs and establishes existence, uniqueness, and the sample-path regularity used throughout the paper."},{"cited_title":"Markov properties of Gaussian random fields on compact metric graphs","cited_arxiv_id":"2304.03190","evidence_quote":"Provides the Markov properties of Gaussian random fields on compact metric graphs that give sparse precision matrices for integer $\\alpha$."},{"cited_title":"Bolin, L","cited_arxiv_id":null,"evidence_quote":"Gives the regularity results and the variance-stationary construction for non-constant $\\tau$, which underpin Assumption B.3 and Propositions B.4 and B.5."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the general theorem on approximation of Bayesian inverse problems that the proof uses to turn potential error into Hellinger-distance error."},{"cited_title":"Møller and J","cited_arxiv_id":null,"evidence_quote":"Earlier Cox-process models on linear networks; the comparison shows why a general metric-graph formulation is needed and why isotropy limits scalability."}],"review_version":1}