{"id":"46f66033-e67b-4e9e-9631-ae742a11831d","arxiv_id":"2507.23724","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A new space-time Markov chain approximation for general diffusions on metric graphs is shown to converge in p-Wasserstein distance at explicit rates governed by a thinness quantifier of the subdivision.","lead":"The paper introduces a random-walk approximation, called STMCA, for diffusion processes on metric graphs and proves explicit bounds on how fast the approximation converges in Wasserstein distance as the grid is refined. It gives the first general numerical scheme with convergence rates for diffusions on networks, which matter for models of nerve fibers, porous media, and electrical networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 4.1 mixes F_{τ_{k-1}} and skeleton conditionals; the L2 bound in (4.2) is not justified and Theorem 2.10 relies on it.","rationale":"The reader identifies the weakest assumption exactly: the proof of Proposition 4.1 uses inconsistent conditional expectations. Independent inspection confirms that the process M_n in Lemma 4.2 is a martingale with skeleton-conditional increments, while the expression in (4.1) involves F_{τ_{k-1}}-conditional increments. The equality in (4.2) therefore does not follow, and (4.3) does not connect to the definition of K(t). This is load-bearing because Proposition 4.1 supplies the quantitative control |τ_{K(t)}−t| that enters the Wasserstein bound in Theorem 2.10. The issue appears fixable by replacing F_{τ_{k-1}} with B throughout the proof, since the skeleton conditional expectations are the natural quantities for K(t) and for Lemmas 4.2–4.3; hence the verdict should remain CONDITIONAL, not REJECT. The paper otherwise contains a coherent construction: explicit transition formulas (Section 3), a plausible embedding argument (Proposition 2.13), and detailed regularity estimates (Section 5). The numerical section is illustrative and does not verify rates, but that is a secondary weakness. The reader's CONDITIONAL verdict is appropriate; no change is needed.","tokens_in":32709,"tokens_out":12126,"duration_ms":113410,"concrete_test":"Rewrite the proof of Proposition 4.1 replacing every occurrence of E[· | F_{τ_{k−1}}] in (4.1)–(4.3) with the skeleton conditional expectation E[· | B], where B = σ(X_{τ_k}; k≥0). Check whether Lemma 4.2, Lemma 4.3, and (4.3) then apply verbatim; if they do, the flaw is a fixable typo and the theorem may still hold. To confirm the written equality in (4.2) is false, take standard Brownian motion on [0,1] with subdivision {0,1/2,1} and compute along a sample path the two cumulative sums Σ_{k≤K(t)} E[τ_k−τ_{k−1} | F_{τ_{k−1}}] and Σ_{k≤K(t)} E[τ_k−τ_{k−1} | X_{τ_{k−1}},X_{τ_k}]; if they differ with positive probability, the asserted equality in (4.2) fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Proposition 4.1 splits the error in (4.1) using conditional expectations E[τ_k−τ_{k−1} | F_{τ_{k−1}}] with respect to the natural filtration at the stopping time. By the strong Markov property, this equals the unconditional expected exit time from the current cell, i.e. Σ_y p_{x,y} t_{x,y}, a weighted average over the next skeleton state. However, the counter K(t) in Section 2.5 and the martingale M_n in Lemma 4.2 are defined with skeleton conditional expectations E[τ_k−τ_{k−1} | X_{τ_{k−1}}, X_{τ_k}] = t_{x,y}, which depend on the actual next state y and generally differ from the F-conditional mean. Equation (4.2) then asserts that the L2 norm of the F_{τ_{k-1}}-centered cumulative sum equals the sum of skeleton conditional variances Var[τ_k−τ_{k−1} | X_{τ_k},X_{τ_{k−1}}]; this equality would hold only for the skeleton-centered martingale of Lemma 4.2, not for the F-centered expression appearing in (4.1). Likewise, (4.3) bounds |t − Σ E[τ_k−τ_{k−1} | F_{τ_{k−1}}]| by sup v_j^1/v_j^0, but K(t) is defined via the skeleton sum; the discrepancy between the two sums is not controlled anywhere. Since Proposition 4.1 provides the quantitative embedding-time bound used to prove Theorem 2.10, the central convergence-rate claim is not established by the written proof.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Space-Time Markov Chain Approximation (STMCA) for general diffusions on finite metric graphs: a random walk on subdivisions whose transition probabilities and conditional holding times match those of the target diffusion. The main result (Theorem 2.10) asserts a quantitative Wasserstein convergence rate O(|Delta|_X^alpha) for alpha < 1/4 wedge 1/p under a regularity condition, with improved rates for adapted subdivisions. The proof proceeds by an embedding property (Proposition 2.13), embedding-time estimates (Proposition 4.1), regularity estimates (Section 5), and a final Wasserstein estimate (Section 6). The paper also gives explicit Green-function formulas for the transition quantities and numerical experiments on star graphs.","tokens_in":33055,"tokens_out":7755,"duration_ms":76921,"significance":"If Theorem 2.10 holds as stated, it is a meaningful contribution: it appears to give the first explicit convergence rate for random-walk approximations of general diffusions on metric graphs, and the explicit transition formulas make the scheme implementable. The embedding construction and the extension of the one-dimensional STMCA framework to metric graphs are natural and potentially useful. The numerical experiments illustrate sticky and boundary effects, although they do not verify the rate. The main theorem is not a definitional circularity: it extends earlier work and contains no fitted parameters. However, the written proof of the central embedding-time estimate has a genuine gap, and the regularity section relies heavily on the author's unpublished preprint [2]; both need to be addressed.","major_comments":[{"comment":"The proof of Proposition 4.1 conflates two different conditional centerings. Equation (4.1) writes the error as tau_{K(t)} - sum_{k=1}^{K(t)} E[tau_k - tau_{k-1} | F_{tau_{k-1}}], but the martingale M_n of Lemma 4.2 is centered at E[tau_k - tau_{k-1} | X_{tau_k}, X_{tau_{k-1}}]. The equality in (4.2) between the L^2 norm of the F_{tau_{k-1}}-centered sum and the sum of skeleton conditional variances is therefore not justified; in fact the two centered sums differ by sum_k (E[cdot | F_{tau_{k-1}}] - E[cdot | skeleton]), which is not a martingale increment and is not controlled anywhere. Since Proposition 4.1 supplies the quantitative embedding-time bound used in Theorem 2.10, this is a load-bearing gap. The proof can likely be repaired by rewriting (4.1) with the skeleton-conditional sums throughout and bounding the second term by the maximal one-step skeleton holding time, but as written the central estimate does not follow.","section":"Section 4.1, Eqs. (4.1)-(4.2)"},{"comment":"The second term in (4.1) is bounded in (4.3) by sup_{y,j} v_j^1(y)/v_j^0(y) <= K_Delta |Delta|_X, but K(t) is defined in Section 2.5 using cumulative sums of E[tau_k - tau_{k-1} | X_{tau_k}, X_{tau_{k-1}}], not of the F_{tau_{k-1}}-conditional expectations appearing in (4.3). The difference between the two cumulative sums is a sum of K(t) terms, and K(t) is typically of order T/|Delta|_X, so the accumulated discrepancy is not bounded by the one-step maximum. Without an additional estimate comparing the two sums, the bound on |t - sum E[tau | F]| does not follow from the definition of K(t).","section":"Section 4.1, Eq. (4.3) and Section 2.5"},{"comment":"The regularity estimates that feed Theorem 2.10 depend on the time-change representation in [2, Theorem 2.14] and on [6, Theorem 3.1], and the proof of Proposition 5.5 contains several substantial steps that are only indicated, such as 'by excursion flipping', 'by the time-change characterization', and the claim that (5.2) is immediate. In particular, the construction of the process Z with speed measure m_Z and the local-time change-of-variables identities following (5.3) would need to be written out, and the stochastic domination in Lemma 5.6, P_v(T^v_{l*} < T) <= P_v(T^Z_{l*} < T), is asserted rather than proved. Since Condition 2.9 implying Condition 2.8 is one of the two sufficient conditions for Theorem 2.10, this chain of unstated dependencies makes the main theorem conditional on [2].","section":"Section 5, Proposition 5.5 and Lemma 5.6"}],"minor_comments":[{"comment":"The notation |U|_X is defined separately for vertex neighborhoods and edge-segments; consider unifying the two cases in one displayed definition to avoid ambiguity.","section":"Section 2.2"},{"comment":"The statement says 'for every x >= 0', but x is an element of the metric graph Gamma; it should read 'for every x in Gamma'.","section":"Theorem 2.10"},{"comment":"There are numerous typos and small grammatical errors, including 'elge-lengths' (Section 2.1), 'poiting' (Section 2.1), 'spate space' (Proposition 2.13), 'vells' (proof of Proposition 3.7), 'Conditon' (Section 5.1), and 'the the maximum' in the abstract.","section":"Throughout"},{"comment":"The definition of W_p^T has a formatting issue: the supremum and the L^p norm should be displayed unambiguously so that the metric on path space is clear.","section":"Section 2.4"},{"comment":"The proof references 'Lemmata C.1 and 4.4' but the displayed bound after conditioning on B applies Lemma 4.4 to each increment; please clarify the exact role of Lemma C.1 in that step.","section":"Proposition 4.5"}],"recommendation":"major_revision","confidential_remarks":"From the editor's viewpoint, the manuscript's dependence on the author's unpublished preprint [2] should be resolved before acceptance. I would ask the author either to state and prove the needed time-change theorem in an appendix or to replace the reference with a published version. The self-citation pattern is not itself problematic, but the proof currently inherits the main analytical input of [2] without a self-contained statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a real extension—the first STMCA approximation for general diffusions on finite metric graphs with explicit Wasserstein rates, including sticky vertices and natural boundaries—but the main rate theorem has a proof gap in Proposition 4.1. Worth a serious referee, but I wouldn't cite the rate yet.\n\nWhat is new: the framework genuinely goes beyond the 1D STMCA of [3] and the Walsh-Brownian-motion-only invariance principle of [33]. The explicit transition formulas in Propositions 3.1–3.4 are useful, the vertex asymptotics in Proposition 3.6 are clean, and the idea of adapting subdivisions to the speed measure to double the rate is conceptually nice. The regularity section extends 1D moment bounds to star graphs and then to finite graphs via time changes; I didn't spot an obvious flaw there, though it leans on the author's own unpublished preprint [2] and on [3]. That is a citation burden, but not by itself a problem.\n\nThe soft spot is Proposition 4.1, and the reader's stress-test is on target. The counter K(t) is defined with skeleton-conditional expectations E[τ_k−τ_{k−1} | X_{τ_{k−1}}, X_{τ_k}], but the proof's split in (4.1) subtracts F_{τ_{k−1}}-conditional expectations. Those two quantities differ: the F-conditional one is a weighted average over the next skeleton state, while the skeleton sum keeps the realized t_{x,y}. Equation (4.2) then asserts that the L2 norm of the F-centered cumulative sum equals the sum of skeleton conditional variances. That equality only holds for the skeleton-centered martingale of Lemma 4.2, not for the F-centered expression appearing in (4.1). Similarly, (4.3) bounds |t − Σ E[·|F_{τ_{k−1}}]| by one cell's mean exit time, but t is near the skeleton sum; the accumulated difference between the skeleton and F sums is never controlled. Since Proposition 4.1 supplies the embedding-time estimate used to prove Theorem 2.10, the advertised convergence rate is not established by the written proof. This may be fixable—perhaps by redefining K(t) with F-conditional sums, or by proving the two sums are close—but as written it is a load-bearing gap.\n\nSmaller points: the numerical section illustrates sample paths and densities but does not verify the claimed rates, and no code is provided. Several lemmas are deferred to prior work by the same group, which slows verification but is not fatal.\n\nFor anyone working on graph diffusion approximations, the framework and explicit transition formulas are worth a look. The rate theorem, though, needs a revision. Yes, send to a serious referee: the contribution is significant and the gap is specific and addressable.","headline":"Genuine extension with useful explicit transition formulas, but the central rate theorem rests on an unjustified conditional-expectation swap in Proposition 4.1.","tokens_in":33565,"tokens_out":2448,"would_cite":false,"duration_ms":25777,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J60","60J55","35J08","60J50"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that a space-time Markov chain on a subdivision of a metric graph approximates any general diffusion on the graph in p-Wasserstein distance with rate $|\\Delta|_X^\\alpha$ for every $\\alpha < \\tfrac14 \\wedge \\tfrac1p$, and…","keywords":["diffusion on network","Markov chain approximation","random walk approximation","invariance principle","generalized second order differential operators","path regularity","metric graph","Wasserstein distance"],"falsifier":"Check on a two-edged star graph with asymmetric spinning probabilities and positive stickiness $\\rho$ whether the two quantities $E_x[\\tau_k-\\tau_{k-1}|\\mathcal{F}_{\\tau_{k-1}}]$ and $E_x[\\tau_k-\\tau_{k-1}|X_{\\tau_k},X_{\\tau_{k-1}}]$ have equal cumulative sums along a skeleton path; if their sums differ by more than the allowed $O(|\\Delta|_X)$ error, Proposition 4.1 fails and Theorem 2.10's main estimate does not follow.","tokens_in":32473,"feed_emoji":"🎲","tokens_out":7679,"duration_ms":72981,"temperature":0.7,"pith_summary":"This paper introduces a random-walk approximation, the Space-Time Markov Chain Approximation (STMCA), for a general diffusion on a finite metric graph, and proves that the approximation converges to the diffusion in p-Wasserstein distance at a controlled rate. The rate is expressed through a thinness quantifier of the subdivision: for any exponent $\\alpha$ below $\\tfrac14 \\wedge \\tfrac1p$, the distance is bounded by a constant times $|\\Delta|_X^\\alpha$, and with adapted subdivisions the same mechanism doubles the rate in the maximal cell size. The proof couples the diffusion and the chain through the embedded skeleton of the diffusion, so that the chain's positions and jump times match the diffusion's skeleton in distribution. The paper also gives explicit formulas for the transition probabilities and conditional transition times, so the scheme can be run numerically.","feed_headline":"New walk scheme converges to graph diffusions at rate 1/4","feed_subtitle":"The STMCA chain matches the diffusion's skeleton in law, enabling explicit formulas and adapted mesh speed-ups.","key_machinery":"The Space-Time Markov Chain Approximation (STMCA) is a $V_\\Delta$-valued random walk whose transitions are $(p_{x,y}, t_{x,y})$, where $p_{x,y}=P_x(T_{U_x}=T_y)$ is the probability that the diffusion exits the cell centered at $x$ through $y$, and $t_{x,y}=E_x(T_{U_x}|T_{U_x}=T_y)$ is the conditional expected exit time. The embedding property of Proposition 2.13 identifies the STMCA path with the diffusion sampled at the random times $\\tau^\\Delta_{K(t)}$, reducing Wasserstein distance to bounds on $|\\tau_{K(t)}-t|$ and on the path regularity of $X$. The thinness quantifier $|\\Delta|_X$, defined via re-oriented scale and speed measures on vertex neighborhoods and edge segments, controls the time-change error; explicit Green-function and Dirichlet-problem formulas (Propositions 3.1–3.6) supply the transition data. Moment bounds for the embedding times (Proposition 4.1) and Kolmogorov-type regularity estimates (Section 5) are the remaining ingredients.","core_discovery":"Theorem 2.10 states that for an NSE (natural-scale-on-edges) general diffusion on a finite metric graph, the STMCA on any covering subdivision $\\Delta$ satisfies $W_p^T(X,\\tilde X^\\Delta) \\le C |\\Delta|_X^\\alpha$ for every $\\alpha \\in (0, \\tfrac14 \\wedge \\tfrac1p)$ and $\\varepsilon>0$, uniformly over $(\\varepsilon,V)$-symmetric covering subdivisions, provided the diffusion satisfies a H\\\"older-regularity condition (Condition 2.8) or a speed-measure lower bound (Condition 2.9). This is claimed to be the first explicit quantitative convergence rate for random-walk approximations of general diffusions on metric graphs, and it implies convergence in law as the thinness quantifier goes to zero. With subdivisions adapted so that $|\\Delta|_X \\lesssim |\\Delta|^2$, the bound becomes $O(|\\Delta|^{2\\alpha})$, i.e. a rate arbitrarily close to $\\tfrac12 \\wedge \\tfrac2p$ in the maximal cell size. The central identity making the result possible is the embedding property: if $\\tau_k$ are the successive hitting times of the subdivision vertices and $K(t)$ is the random counter built from skeleton-conditional expected jump times, then $X_{\\tau_{K(t)}}$ has the same law as the STMCA at time $t$.","pith_inferences":["Editorial inference: If the embedding-time identification holds, the same skeleton-coupling argument should yield explicit constants in the rate, making the STMCA usable as a certified Monte Carlo method for diffusions on networks.","Editorial inference: The thinness quantifier suggests a natural adaptive mesh criterion: refine cells where the speed measure is large or where boundaries are sticky, which the numerics indicate matters for reproducing boundary repulsion.","Editorial inference: The restriction to NSE diffusions may be largely removable; the author notes that many non-NSE cases reduce to NSE by a transformation, so the scheme likely extends to skew diffusions on graphs after a deterministic change of coordinates.","Editorial inference: A direct testable extension is to replace the p-Wasserstein bound with a path-space strong approximation result, which would follow if the embedding times $\\tau_{K(t)}$ could be shown to be close to $t$ in a stronger sense than $L^2$."],"forward_implications":["For every $T>0$, the STMCA processes converge in law to the diffusion as the thinness quantifier $|\\Delta|_X$ tends to zero (Corollary 2.11).","The explicit formulas for transition probabilities and times make the scheme implementable, with asymptotics near vertices that simplify the sticky case.","Adapting the subdivision so that $|\\Delta|_X \\le |\\Delta|^2$ pushes the convergence rate in the maximal cell size up to any exponent below $\\tfrac12 \\wedge \\tfrac2p$.","The construction generalizes the one-dimensional STMCA, the classical invariance principle, sticky random walks, and oscillating random walks to general metric-graph diffusions.","The Wasserstein bound is uniform over $(\\varepsilon,V)$-symmetric covering subdivisions, so the rate statement holds for a whole family of discretizations."],"supporting_citations":[{"why":"Supplies the one-dimensional STMCA construction and its convergence proof, which the graph version extends.","marker":"[3]"},{"why":"Provides the time-change characterization of general diffusions on star graphs and the Dirichlet problem used for transition formulas.","marker":"[2]"},{"why":"Supplies Wasserstein convergence rates and moment bounds used to derive Condition 2.8 from Condition 2.9.","marker":"[6]"},{"why":"Gives the invariance principle for Walsh Brownian motion that the STMCA generalizes and whose limitations motivate the new scheme.","marker":"[33]"},{"why":"Serves as the standard reference for scale functions, Doob's inequality, and local time identities used in the regularity and embedding-time proofs.","marker":"[34]"}],"fun_headline_variants":["Random walk on meshes converges to graph diffusions at explicit rate","Space-time chain nails graph diffusion approximation with proven rate","For metric graphs, a simple walk matches diffusions to order 1/4","Explicit convergence rate for walk approximations on metric graphs","Asymmetric walk on graph subdivisions converges to diffusion law"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of the key embedding-time bound assumes that conditioning on the full past before a jump gives the same expected jump length as conditioning only on the two vertices the jump connects; this equality is asserted without proof, and the bound on the time error collapses if it fails.","fun_headline_variants_meta":{"raw":{"variants":["Random walk on meshes converges to graph diffusions at explicit rate","Space-time chain nails graph diffusion approximation with proven rate","For metric graphs, a simple walk matches diffusions to order 1/4","Explicit convergence rate for walk approximations on metric graphs","Asymmetric walk on graph subdivisions converges to diffusion law"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000325,"raw_usage":{"total_tokens":1850,"prompt_tokens":1001,"completion_tokens":849,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":764}},"tokens_in":617,"tokens_out":849,"duration_ms":8043,"temperature":1.0,"reasoning_tokens":764,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:27:52.347450+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check on a two-edged star graph with asymmetric spinning probabilities and positive stickiness $\\rho$ whether the two quantities $E_x[\\tau_k-\\tau_{k-1}|\\mathcal{F}_{\\tau_{k-1}}]$ and $E_x[\\tau_k-\\tau_{k-1}|X_{\\tau_k},X_{\\tau_{k-1}}]$ have equal cumulative sums along a skeleton path; if their sums differ by more than the allowed $O(|\\Delta|_X)$ error, Proposition 4.1 fails and Theorem 2.10's main estimate does not follow.","supporting_citations":[{"cited_title":"Anagnostakis, A","cited_arxiv_id":null,"evidence_quote":"Supplies the one-dimensional STMCA construction and its convergence proof, which the graph version extends."},{"cited_title":"General diffusions on the star graph as time-changed Walsh Brownian motion","cited_arxiv_id":"2502.19299","evidence_quote":"Provides the time-change characterization of general diffusions on star graphs and the Dirichlet problem used for transition formulas."},{"cited_title":"Ankirchner, T","cited_arxiv_id":null,"evidence_quote":"Supplies Wasserstein convergence rates and moment bounds used to derive Condition 2.8 from Condition 2.9."},{"cited_title":"Pavlyukevich and A","cited_arxiv_id":null,"evidence_quote":"Gives the invariance principle for Walsh Brownian motion that the STMCA generalizes and whose limitations motivate the new scheme."},{"cited_title":"Revuz and M","cited_arxiv_id":null,"evidence_quote":"Serves as the standard reference for scale functions, Doob's inequality, and local time identities used in the regularity and embedding-time proofs."}],"review_version":1}