{"id":"0dadd1b4-28d8-4e99-bd26-55426f4b9847","arxiv_id":"2501.14135","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"In continuous time, the Aldous, Hoover-Keisler, Hellwig, and optimal-stopping topologies on naturally filtered processes coincide and are metrized by an adapted Wasserstein distance, whose completion is the space of general filtered processes.","lead":"A mathematics paper proves that several competing ways to measure convergence of stochastic processes with information flow, coming from different communities, all define the same topology in continuous time. It also constructs the complete metric space of all filtered processes, the adapted Wasserstein space, and shows Donsker-style random walk limits and Euler schemes converge in this metric.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Density proof in Prop 5.6 constructs a non-natural Y for càdlàg paths: future grid information is revealed before the path encodes it, so Theorem 1.5's completion claim lacks a valid proof as written.","rationale":"The reader's conditional verdict was driven mainly by reliance on the canonical Hoover-Keisler representation from [27] and on [18, Theorem 5.4]. My stress-test found a more concrete, internal gap: the density proof in Proposition 5.6 does not actually produce a naturally filtered process in the càdlàg case. The construction reveals Z at the next grid point through the filtration before the piecewise constant path carries that information, so the process is not Hoover-Keisler equivalent to its natural filtration version. This invalidates the proof of density of NFP_p in (FP_p, AW_p), which is one half of the central completion theorem. The flaw is localized and appears repairable: replacing the future-pointing filtration F^Z_{ceil(t)_T} by the past-pointing F^Z_{floor(t)_T} on each cell, or modifying the path so that the next grid value is encoded before it is used, should restore naturalness while preserving the epsilon-bicausal estimates. However, because the argument as written is wrong at a load-bearing step, the paper cannot be accepted as is; it needs a corrected proof of Proposition 5.6. The rest of the metric theory, including the AW_p metric and its equivalence with the Hoover-Keisler topology, is carefully developed and I found no comparable flaw there.","tokens_in":48498,"tokens_out":22251,"duration_ms":211241,"concrete_test":"Check the two-point counterexample. Let T = {1/2,1}, let Z_{1/2} be a fair +/-1 coin flip, Z_1 = Z_{1/2}, and define Y as in Prop 5.6 with piecewise constant càdlàg paths. For s = 0.4, t = 0.45, compute pp_s(Y) and compare it with the corresponding prediction process of S^{L(Y)}, the standard naturally filtered process with the same path law; they differ, so Lemma 5.2(ii) fails. Then test the proposed repair F^Y_t = F^Z_{floor(t)_T}: verify that the coupling constructed in Prop 3.10 remains epsilon-bicausal and that Y becomes naturally filtered. If the repair works, the main theorem is likely salvageable but the current proof must be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Proposition 5.6, which underpins the Theorem 1.5 completion claim, constructs Y from a discrete-time plain process Z by setting F^Y_t := F^Z_{ceil(t)_T} and Y := iota_T(Z). For X = D([0,1]; R^d), iota_T is piecewise constant: Y_t = Z_{t_{i-1}} on [t_{i-1}, t_i). The proof asserts that for 0 < s < t with ceil(s)_T = t_i one has t > t_i, hence pp_s(Y) is a Borel function of Y restricted to [0,t]. But t > t_i is false when s and t lie in the same grid cell, e.g. T = {1/2,1}, s = 0.4, t = 0.45. Then F^Y_s = F^Z_{t_i} reveals Z_{t_i}, while the piecewise constant path Y restricted to [0,t] does not determine Z_{t_i}. Thus the constructed Y fails Lemma 5.2(ii) and is not naturally filtered: if Z_{1/2} is a fair +/-1 coin flip with Z_1 = Z_{1/2}, then for s < 1/2 the prediction process pp_s(Y) knows the coin flip, whereas the natural-filtration process with the same path law does not. Consequently, NFP_p is not shown to be dense, and the completion theorem is not established by the given argument. A repair may exist, e.g. using F^Y_t = F^Z_{floor(t)_T} on each cell or encoding the future grid value in the path, but as written this is a load-bearing gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a continuous-time adapted Wasserstein distance AW_p on filtered processes, defined through relaxed (ε-)bicausal couplings, and claims that it metrizes a canonical adapted weak topology. The main theorems state that on naturally filtered continuous processes the Aldous MZ, Hoover–Keisler, Hellwig, and optimal stopping topologies coincide with the AW_p topology (Theorem 1.1); that the space FP_p of filtered processes modulo Hoover–Keisler equivalence is the AW_p-completion of the space of naturally filtered processes (Theorem 1.5); that AW_p is a complete metric inducing the Hoover–Keisler topology (Theorem 3.13); and that martingales are closed, optimal stopping values are continuous, Prohorov-type compactness holds, and Donsker and Euler approximation results extend to AW_p. The proofs rely substantially on the canonical representation theory of Beiglböck–Pammer–Schrott–Zhang and on discrete-time results of Bartl–Beiglböck–Pammer.","tokens_in":48792,"tokens_out":8447,"duration_ms":78478,"significance":"If the results are correct, the paper provides a unified and complete metric framework for adapted weak convergence of continuous-time stochastic processes, with zero-distance classes equal to Hoover–Keisler equivalence classes. This would be a substantial contribution to adapted transport theory, mathematical finance, and the theory of weak convergence of stochastic processes. The paper is also commendably explicit about where it relies on prior work: the canonical representation from [27] and the discrete-time density and metrization results from [18] are clearly cited, and the technically involved invariance of AW_p under canonical representatives is relegated to Appendix A. The distinction between the relaxed distance AW_p and the strict adapted Wasserstein distance is carefully discussed, including the failure of the strict version to satisfy Donsker-type approximation and separability. However, the density of naturally filtered processes, which is the load-bearing step for the completion theorem, is not established by the argument given in Proposition 5.6.","major_comments":[{"comment":"The proof that the constructed continuous-time process Y is naturally filtered contains a false inequality. The text fixes i with ⌈s⌉_T = t_i and asserts 'As t > t_i' before the measurability argument; this is not true when s and t lie in the same grid interval. For example, with T = {1/2, 1}, s = 0.4, and t = 0.45, one has ⌈s⌉_T = 1/2 but t < 1/2. In that case F^Y_s = F^Z_{1/2} contains Z_{1/2}, while the piecewise constant path ι_T(Z) restricted to [0, t] does not determine Z_{1/2}; for D([0,1]; R^d), the definition of ι_T even sets Y_{t_i} = Z_{t_{i-1}} for t_i < 1. Taking Z_{1/2} = ±1 as a fair coin flip and Z_1 = Z_{1/2}, the prediction process pp_s(Y) is the Dirac measure at that coin flip, whereas the natural filtration of the same path law has trivial conditional information at s. Hence Y fails Lemma 5.2(ii), and Proposition 5.6 does not establish that NFP_p is dense in FP_p. Since Theorem 1.5, the completion statement, relies on this density, its proof is incomplete as written. A repair may exist—for instance by choosing F^Y_t = F^Z_{t_{i-1}} on each grid cell or by encoding grid values in the path—but the argument given in the manuscript is not valid.","section":"Section 5, Proposition 5.6"}],"minor_comments":[{"comment":"There is a typo: 'Proportion 3.19' should read 'Proposition 3.19'.","section":"Section 3.3"},{"comment":"The proof of Proposition 5.5 is only a sketch, and its Gδ characterization of NFP is deferred to a cited remark in [27]. If this Polishness statement is part of the advertised results, the proof should either be completed in the text or explicitly marked as a consequence of [27].","section":"Section 5, Proposition 5.5"},{"comment":"The definition of ι_T for D([0,1]; R^d) makes Y left-continuous at grid points, while the filtration uses ⌈t⌉_T. A short remark explaining this choice and its consequences for the measurability of prediction processes would help prevent the kind of confusion that arises in the proof.","section":"Section 5, Proposition 5.6"},{"comment":"In the proof of Proposition 3.10, the statement that e_T is not continuous for D([0,1]; R^d) is acknowledged, but the approximation argument still uses convergence of d_X(f, ι_T(e_T(f))) for every f; it would be useful to spell out why this convergence is uniform enough for the dominated convergence step in the D case.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The main technical concern is localized to Proposition 5.6, and I do not see evidence of internal inconsistency in the rest of the framework. If the authors can repair the density proof, the paper would be a strong candidate for acceptance. The reliance on [18] and [27] is substantial but explicitly acknowledged, and the Appendix A treatment of canonical representatives is a positive feature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nHere is my take on Bartl et al. The continuous-time equivalence results (Theorem 1.1) and the AW_p metric structure are real, and the paper is mostly careful. But there is a load-bearing gap in Proposition 5.6, and the stress-test note correctly identifies it. The density proof constructs Y from discrete Z with F^Y_t = F^Z_{ceil(t)_T}. To show Y is naturally filtered, the proof claims that for 0<s<t with ceil(s)_T=t_i one has t>t_i. That is false when s and t lie in the same grid cell. Then F^Y_s already contains Z_{t_i}, while the piecewise constant path Y on [0,t] does not determine Z_{t_i}. So Lemma 5.2(ii) fails; the constructed Y need not have a natural filtration. Since Proposition 5.6 is the density step behind Theorem 1.5, the completion theorem is not established by the given argument.\n\nThis is the main soft spot. I do not see a wrong central idea; a fix is plausible, e.g. choose a left-continuous grid filtration or encode the future grid value in the path, but that needs to be written and checked. Proposition 5.5 is also only sketched; that is minor. The paper's heavy reliance on [18] and [27] is legitimate—those are published external results, not restatements—though it does mean the referee must verify that [27]'s canonical representation is being used correctly.\n\nWhat the paper does well: it gives a genuine unification, not just a survey; the epsilon-bicausal relaxation is a sensible way to handle cadlag paths; the optimal stopping continuity section is clean; and the quantitative Donsker/Euler rates are useful. The proofs for Theorem 3.13 and the equivalence of topologies are detailed and appear sound.\n\nFor whom: people working in adapted transport, mathematical finance, and stochastic analysis. It deserves a serious referee, but the referee should take Proposition 5.6 seriously. My recommendation: send it to peer review, and in the report ask for a corrected density argument. As it stands, I would not rely on Theorem 1.5.","headline":"Important continuous-time unification, but the density proof behind the completion theorem has a real gap.","tokens_in":49374,"tokens_out":2572,"would_cite":false,"duration_ms":25005,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B10","60G07","60G40","60G44","60F17","49Q22"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes one canonical adapted weak topology for continuous-time stochastic processes, metrized by a relaxed adapted Wasserstein distance, and shows its completion is the space of all filtered processes modulo Hoover–Keisler…","keywords":["adapted weak topology","adapted Wasserstein distance","causal transport","Hoover–Keisler equivalence","optimal stopping","Donsker's theorem","filtered stochastic processes","completion of stochastic processes"],"falsifier":"Compute $\\mathcal{AW}_1$ between the scaled random walk $B_n$ and Brownian motion $B$: the paper predicts a bound of order $\\log(n)/n^{1/3}$ tending to zero, so any positive lower bound along a subsequence, or any pair of naturally filtered continuous processes converging in all four listed topologies but not in $\\mathcal{AW}_p$, would falsify the main equivalence theorems.","tokens_in":48254,"feed_emoji":"🎲","tokens_out":10718,"duration_ms":90709,"temperature":0.7,"pith_summary":"On the space of laws of continuous-time processes with natural filtrations, the paper aims to establish that there is one canonical adapted weak topology, not several rival ones. It proves that the Aldous MZ topology, the Hoover–Keisler topology, Hellwig's information topology, and the optimal-stopping topology all coincide, and that this common topology is metrized by a relaxed adapted Wasserstein distance $\\mathcal{AW}_p$. A second theorem identifies the completion of the incomplete space of naturally filtered processes: it is exactly the space $(\\mathrm{FP}_p,\\mathcal{AW}_p)$ of all filtered processes modulo Hoover–Keisler equivalence, which is Polish, has martingales as a closed subset, and supports Donsker-type and Euler-scheme approximations. The upshot for a general reader is a single language in which \"the process and its flow of information converge\" has one precise meaning, with optimal stopping values continuous whenever the limit has continuous paths.","feed_headline":"One topology unifies all adapted weak convergence notions","feed_subtitle":"Aldous, Hoover–Keisler, Hellwig and optimal stopping agree, with a complete Wasserstein metric.","key_machinery":"The load-bearing object is the relaxed adapted Wasserstein distance $\\mathcal{AW}_p$, defined on filtered processes by infimizing $\\mathbb{E}_\\pi[d_{\\mathcal{X}}^p(X,Y)]^{1/p}+\\varepsilon$ over $\\varepsilon$-bicausal couplings: couplings whose conditional-independence structure is causal in both directions up to a time shift of $\\varepsilon$. The $\\varepsilon$ relaxation is what turns the too-strong strict adapted Wasserstein distance into a genuine weak topology. The second pillar is the canonical Hoover–Keisler representative from the representation theory of filtered processes: every equivalence class has a canonical filtered process on a standard Borel space whose path map is continuous on its support, which yields compactness of $\\varepsilon$-bicausal couplings, well-definedness of $\\mathcal{AW}_p$ on equivalence classes, and ultimately completeness. Discretization arguments import the discrete-time adapted Wasserstein theory to bridge between continuous-time processes and finite-grid approximations.","core_discovery":"The central claim is Theorem 1.1 and Theorem 1.5: on continuous-time processes with natural filtrations, Aldous MZ, Hoover–Keisler, Hellwig, and optimal-stopping convergence agree with convergence in $\\mathcal{AW}_p$; and $(\\mathrm{FP}_p,\\mathcal{AW}_p)$, the quotient of all filtered processes by $\\mathcal{AW}_p$-equivalence, is the completion of the naturally filtered processes. The metric identifies exactly the Hoover–Keisler equivalence classes, so $\\mathcal{AW}_p(X,Y)=0$ means the two processes carry the same probabilistic information in the strongest iterated-prediction sense. The paper also proves that the completed space is Polish, that martingales form a closed subset, and that scaled random walks and Euler schemes converge in $\\mathcal{AW}_p$ to Brownian motion and SDE solutions.","pith_inferences":["Editorial extension: because $\\mathcal{AW}_p$-convergence implies continuity of optimal stopping values for continuous limits, the Donsker and Euler results provide a ready-made stability guarantee for numerical schemes in robust finance, although the paper does not draw that financial conclusion.","Editorial extension: the completion theorem suggests that in model uncertainty, the closure of a class of models should be taken not by enriching path spaces but by allowing arbitrary filtrations while penalizing information gaps only up to $\\varepsilon$ time shifts; one could test this interpretation by constructing insider-information models whose $\\mathcal{AW}_p$-limits differ from their weak l","Editorial extension: Remark 3.4 notes alternative penalties such as $\\sqrt{\\varepsilon}$ for continuous martingales; one could check whether those penalties define the same topology with different quantitative rates, which would give a family of adapted metrics tailored to different classes of processes.","Editorial extension: since optimal-stopping continuity fails for discontinuous limits (Example D.1), a natural open problem is to characterize the largest class of payoff functions or limiting processes that restore continuity; the paper explicitly defers this question."],"forward_implications":["All of the main convergence notions—Aldous MZ, Hoover–Keisler, Hellwig, and optimal stopping—define the same topology on continuous naturally filtered processes, so a convergence result proved in one framework transfers immediately to the others.","The completion of naturally filtered processes is exactly the space of all filtered processes modulo Hoover–Keisler equivalence, equipped with the complete Polish metric $\\mathcal{AW}_p$; every sequence of natural-filtration models therefore has a limit in this larger space.","$\\mathcal{AW}_p$-equivalence coincides with Hoover–Keisler equivalence, so functionals such as optimal stopping values are constant on equivalence classes and continuous along $\\mathcal{AW}_p$-convergent sequences whenever the limit has continuous paths.","Martingales form a closed subset of $(\\mathrm{FP}_p,\\mathcal{AW}_p)$, hence limits of martingale approximations—random walks, Euler schemes, empirical processes—are automatically martingales in the limit.","Donsker's theorem and Euler approximations hold in the adapted weak sense with quantitative rates, such as $\\mathcal{AW}_1(B,B_n)=O(\\log n/n^{1/3})$ and $\\mathcal{AW}_1(X,X^n)=O(\\sqrt{\\log n/n})$, making the discretization error in optimal stopping problems controllable."],"supporting_citations":[{"why":"supplies canonical Hoover–Keisler representatives, the Polish-space and Prohorov-type results, and continuity-point machinery that $\\mathcal{AW}_p$ relies on.","marker":"[27]"},{"why":"establishes the discrete-time adapted Wasserstein space and its completion, used in time-discretization and denseness of naturally filtered processes.","marker":"[18]"},{"why":"proves equality of all adapted topologies in discrete time, imported through discretization and used for causal coupling and optimal-stopping arguments.","marker":"[12]"},{"why":"defines Hoover–Keisler equivalence via iterated prediction processes, the relation that $\\mathcal{AW}_p$-equivalence is shown to match.","marker":"[43]"},{"why":"defines the Hoover–Keisler topology on general filtered processes and the compactness criterion that the completed space inherits.","marker":"[42]"},{"why":"defines Aldous' extended weak topology and the first-order prediction process, one of the four notions unified in Theorem 1.1.","marker":"[9]"},{"why":"provides the Meyer–Zheng compactness criterion for càdlàg martingales that justifies the topology on prediction-process spaces.","marker":"[57]"},{"why":"gives the explicit Brownian-motion/random-walk coupling underlying the quantitative Donsker bound in $\\mathcal{AW}_1$.","marker":"[54]"},{"why":"supplies the Euler-scheme error estimate used to prove $\\mathcal{AW}_p$-convergence of SDE discretizations.","marker":"[52]"}],"fun_headline_variants":["One adapted metric unifies all weak convergence views","Adapted Wasserstein distance completes process space","Hoover–Keisler, Aldous, Hellwig: now one topology","All convergence notions agree via adapted Wasserstein","Donsker holds in adapted Wasserstein metric"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes as given the Hoover–Keisler classification result that every stochastic process has a canonical representative on a standard Borel space whose path map is continuous on its support, and if that result failed the metric-completeness argument would collapse.","fun_headline_variants_meta":{"raw":{"variants":["One adapted metric unifies all weak convergence views","Adapted Wasserstein distance completes process space","Hoover–Keisler, Aldous, Hellwig: now one topology","All convergence notions agree via adapted Wasserstein","Donsker holds in adapted Wasserstein metric"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000351,"raw_usage":{"total_tokens":1895,"prompt_tokens":908,"completion_tokens":987,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":911}},"tokens_in":524,"tokens_out":987,"duration_ms":8521,"temperature":1.0,"reasoning_tokens":911,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:22:33.304046+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $\\mathcal{AW}_1$ between the scaled random walk $B_n$ and Brownian motion $B$: the paper predicts a bound of order $\\log(n)/n^{1/3}$ tending to zero, so any positive lower bound along a subsequence, or any pair of naturally filtered continuous processes converging in all four listed topologies but not in $\\mathcal{AW}_p$, would falsify the main equivalence theorems.","supporting_citations":[{"cited_title":"Beiglböck, G","cited_arxiv_id":null,"evidence_quote":"supplies canonical Hoover–Keisler representatives, the Polish-space and Prohorov-type results, and continuity-point machinery that $\\mathcal{AW}_p$ relies on."},{"cited_title":"Bartl, M","cited_arxiv_id":null,"evidence_quote":"establishes the discrete-time adapted Wasserstein space and its completion, used in time-discretization and denseness of naturally filtered processes."},{"cited_title":"Backhoff-Veraguas, D","cited_arxiv_id":null,"evidence_quote":"proves equality of all adapted topologies in discrete time, imported through discretization and used for causal coupling and optimal-stopping arguments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines Hoover–Keisler equivalence via iterated prediction processes, the relation that $\\mathcal{AW}_p$-equivalence is shown to match."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines the Hoover–Keisler topology on general filtered processes and the compactness criterion that the completed space inherits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines Aldous' extended weak topology and the first-order prediction process, one of the four notions unified in Theorem 1.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the Meyer–Zheng compactness criterion for càdlàg martingales that justifies the topology on prediction-process spaces."},{"cited_title":"Komlós, P","cited_arxiv_id":null,"evidence_quote":"gives the explicit Brownian-motion/random-walk coupling underlying the quantitative Donsker bound in $\\mathcal{AW}_1$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Euler-scheme error estimate used to prove $\\mathcal{AW}_p$-convergence of SDE discretizations."}],"review_version":1}