{"id":"6c8f1ea1-bab0-4e00-9f3b-59e37168201e","arxiv_id":"2508.17641","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"An entropic formulation of martingale optimal transport is solved by Sinkhorn-type algorithms with sparse Newton iterations, yielding approximate constraint satisfaction and fast practical convergence.","lead":"The authors propose Sinkhorn-type algorithms for entropic optimal transport with martingale-type constraints, using sparse Newton iterations on the dual objective. The paper reports super-exponential convergence in practice with controllable thresholds for total constraint violations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim of controllable approximate feasibility is unsupported: the abstract does not show a quantitative trade-off between entropic regularization level, constraint violation, and sparse-Newton convergence, so the regime where the algorithm is both fast and accurate may not exist.","rationale":"The paper addresses a real problem, and the abstract's idea of Sinkhorn-type iterations with sparse Newton steps on the dual is plausible, especially because martingale conditions do become row-wise linear constraints on the coupling matrix. There is no indication of bad faith, and the abstract is candid that exact martingale conditions are treated as infeasible and an approximate solution is sought. That candidness is exactly what exposes the gap: 'controllable thresholds' requires a quantitative trade-off among regularization bias, Hessian sparsity, ill-conditioning, and convergence speed. The reader identified Hessian sparsity and approximate constraint satisfaction as the weakest assumptions; I agree with that identification, but I would sharpen it by emphasizing the interaction between them. The failure mode is not merely that either assumption could be wrong; it is that small regularization needed for small violations can make the Hessian denser or worse conditioned, slowing or destroying the Newton convergence that the method exploits. This interaction is the load-bearing point. A single reproduction experiment sweeping the regularization parameter and recording both constraint violation and convergence order would decide the question. Since neither a proof nor an experiment is inspectable from the supplied material, the honest verdict remains UNVERDICTED; my concern does not move that verdict.","tokens_in":13877,"tokens_out":4634,"duration_ms":55103,"concrete_test":"Obtain the uncorrupted manuscript at arXiv:2508.17641 and locate the theorem or experimental section that quantifies the relation between the entropic regularization parameter and the constraint violation norm; then run the published or reconstructed algorithm on a discrete MOT instance with n at least 200, scanning the regularization parameter downward until a stated violation threshold (e.g., 1e-6) is met, while recording Newton iteration counts and Hessian sparsity at each value. If the constraint violation does not decrease as the regularization shrinks, or if super-linear convergence is lost before the threshold is met, the abstract's central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract makes two coupled assertions: (i) sparse-Newton iterations converge super-exponentially, and (ii) total constraint violations stay below a controllable threshold. The second is the more load-bearing because the paper itself concedes that exact martingale constraints are 'typically infeasible' and that only an approximate constraint-satisfied coupling is produced. For the central claim to hold, the algorithm must have a regime in which the regularization parameter is small enough that the violation of the martingale constraints is below the advertised tolerance while the Hessian remains sparse enough and well-conditioned enough for super-exponential Newton convergence. The abstract states that this behavior occurs 'in practice' but provides no theorem or experiment. In particular, no bound is stated connecting the regularization parameter to the constraint violation, and nothing is said about how Hessian sparsity behaves as the regularization parameter decreases. If the sparse-Hessian approximation degrades precisely when the regularization is made small enough to meet a strict MOT tolerance, the algorithm's efficiency and the central claim fail in the target regime. The supplied body text is unreadable, so the essential trade-off cannot be checked from the available evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes entropic regularization for discrete martingale optimal transport and for the broader class of optimal transport problems with row-wise linear (super-/sub-)martingale-type constraints. It presents Sinkhorn-type algorithms with sparse Newton iterations that exploit the approximate sparsity of the Hessian of the dual objective, and it claims that these algorithms achieve super-exponential convergence in practice while keeping total constraint violations below controllable thresholds. The abstract is readable, but the supplied full text is not: the body is character-corrupted and effectively unreadable, so no derivations, algorithm pseudocode, convergence statements, or experimental tables can be verified. My report is therefore necessarily based on the abstract and on the structure of the claims.","tokens_in":14113,"tokens_out":2828,"duration_ms":32046,"significance":"If the claims could be verified, the paper would address a real computational need: discrete martingale optimal transport is a structurally constrained OT problem with applications in quantitative finance, and an efficient solver with explicit control over constraint violations would be valuable. The idea of combining entropic regularization with sparse Newton iterations is plausible and potentially novel. However, the significance cannot be assessed from the submitted text: the central claims are stated as empirical observations ('in practice') with no supporting theorem, no quantitative bound linking the regularization parameter to constraint violation, and no readable numerical evidence. No machine-checked proofs, reproducible code, or falsifiable predictions are visible. The contribution is therefore currently unsubstantiated.","major_comments":[{"comment":"The central claim that the proposed algorithms 'in practice' enjoy both super-exponential convergence and robustness with controllable thresholds for total constraint violations is not supported by any equation, theorem, or experiment in the readable portion of the manuscript. Since the full text is unreadable, there is no way to check whether a convergence proof or numerical verification exists. This is load-bearing: the paper's stated contribution rests on these two properties, and neither is presently evidenced.","section":"Abstract"},{"comment":"The manuscript concedes that exact martingale conditions are 'typically infeasible' and that an entropically regularized solution only approximately satisfies the constraints, but it provides no quantitative trade-off between the regularization parameter, the amplitude of total constraint violations, and the sparsity or conditioning of the Hessian. Without such a bound, the claim of 'controllable thresholds' is not established, and it remains possible that the algorithm is fast only in regimes where the constraint violation exceeds any meaningful tolerance. The authors should state an explicit relationship between the regularization parameter and the constraint-violation norm, and should demonstrate empirically that the Hessian remains sufficiently sparse and well conditioned as the regularization parameter is decreased.","section":"Abstract"},{"comment":"The body of the manuscript is not readable; it appears as a character-encoding corruption rather than as coherent text. Consequently, none of the algorithm definitions, assumptions, derivations, or numerical results can be checked. This is not a typographical or stylistic issue but a fundamental obstacle to review. A resubmission must contain a readable full text, including clear statements of the algorithm, the convergence result (with proof or precise reference), and experimental protocols and results.","section":"Full text"}],"minor_comments":[{"comment":"The text contains a reference to 'arXiv:2508.17638v1 [cs.CV]', which appears to belong to a different paper and should be removed or corrected.","section":"Full text"},{"comment":"The term 'super-exponential convergence' is used without a formal definition; the authors should specify whether they claim local quadratic convergence, an iteration-complexity bound, or an empirical rate.","section":"Abstract"},{"comment":"The phrase 'controllable thresholds for total constraint violations' should be made precise: the norm in which the violation is measured and the sense in which the threshold is controllable (e.g., via the regularization parameter or via a posteriori bounds) are not stated.","section":"Abstract"}],"recommendation":"uncertain","confidential_remarks":"The manuscript as submitted is not reviewable because the full text is unreadable. I recommend returning the paper to the authors to provide a properly encoded PDF or source text before any technical evaluation. The accessible abstract alone does not provide enough evidence for the central claims, but there is no indication of an irreparable error; the issue is lack of verifiable content."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the abstract describes a legitimate algorithmic idea—entropic regularization of discrete MOT, with Sinkhorn-type iterations accelerated by sparse Newton steps on the dual Hessian. The martingale conditions as row-wise equality constraints on the coupling matrix is a clean observation that naturally covers a broader class of structural OT problems. The authors are also candid that exact martingale conditions are typically infeasible, so they aim for approximately constrained solutions. That is honest and reasonable.\n\nThe soft spot is exactly where the stress-test note lands. The abstract says the algorithms enjoy super-exponential convergence and controllable constraint violations 'in practice.' No theorem is claimed, but no experiments or quantitative bounds are visible either. The load-bearing question is whether there is a regime where the entropic regularization parameter is small enough that the constraint violations fall under the advertised tolerance, while the Hessian remains sparse and well-conditioned enough for quadratic Newton convergence to hold. The abstract doesn't address this trade-off. I cannot say the approach fails—the logic is plausible—but the evidence presented here does not demonstrate the central claim. The phrase 'in practice' is doing a lot of work, and with the full text corrupted on my end, I could not check the actual tables, plots, or proofs.\n\nI would not flag this as circular or fitting; it is a forward optimization method and the convergence claim is empirical rather than derived from the desired output. The novelty is incremental rather than paradigm-shifting, but the combination for MOT with sparse Hessians appears to be a genuine extension of existing Sinkhorn/Newton work, and the authors don't oversell it in the abstract.\n\nMy recommendation: send it to peer review. The claims are substantive, the problem is timely, and a referee with a readable manuscript can test the central trade-off—either by checking the numerical experiments or trying to construct a counterexample. A desk rejection based on the abstract alone would be premature, especially since the corrupted text prevents anyone from fairly evaluating the full contribution. In a reading group, I'd bring it up for discussion of the regularization-versus-sparsity trade-off, but I wouldn't rely on the results until they're independently verifiable.","headline":"Plausible algorithmic extension for discrete MOT (Sinkhorn + sparse Newton), but the central convergence/accuracy trade-off is only claimed 'in practice' and the body was unreadable; still worth sending to a serious referee.","tokens_in":14528,"tokens_out":2248,"would_cite":false,"duration_ms":27093,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","90C25"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that entropic optimal transport under martingale-type constraints, including discrete martingale optimal transport, is solved efficiently by Sinkhorn-type iterations with sparse Newton steps, achieving super-exponential…","keywords":["entropic optimal transport","martingale optimal transport","Sinkhorn algorithm","sparse Newton method","dual objective Hessian","structural constraints","approximate constraint satisfaction","super-exponential convergence"],"falsifier":"Take a discrete martingale OT instance with a dense, nonlocal cost matrix and measure the density of the dual Hessian at iteration one; if the Hessian is not approximately sparse, or if a truncated sparse Newton step fails to decrease the dual objective, the central convergence claim fails. Alternatively, run the algorithm to its stopping threshold and compute the total martingale constraint violation on a fine grid of regularization strengths; if the violation cannot be made to track the prescribed threshold, the controllability claim fails.","tokens_in":13739,"feed_emoji":"🧮","tokens_out":6594,"duration_ms":68449,"temperature":0.7,"pith_summary":"This paper is trying to establish that entropic optimal transport under martingale-type constraints, including discrete martingale optimal transport, can be solved efficiently by Sinkhorn-type iterations accelerated with sparse Newton steps. The key claim is that these iterations converge super-exponentially in practice while keeping the total violation of the martingale constraints below a user-controllable threshold. Because exact martingale conditions are typically infeasible, the paper deliberately solves an entropically regularized problem that returns an approximately constraint-satisfied coupling. If correct, this gives a practical, scalable solver for a broad class of structurally constrained transport problems, not only the martingale case.","feed_headline":"Sparse Newton iterations solve martingale transport fast","feed_subtitle":"A Sinkhorn-type algorithm keeps martingale constraints within a preset tolerance while converging quickly.","key_machinery":"The central object is the entropically regularized dual problem for martingale-type optimal transport, and the mechanism is the sparse Newton iteration applied to that dual. At each step, instead of building and solving a dense Newton system for the dual variables, the algorithm uses the approximate sparsity of the Hessian $H$ of the dual objective, so the linear solve is restricted to a sparse system. The Sinkhorn-type (Bregman) updates handle the entropic and marginal structure, while the Newton corrections handle the row-wise (in)equality constraints. The approximate sparsity is what converts an otherwise costly second-order method into one whose per-iteration cost stays manageable.","core_discovery":"On its own terms, the paper's central discovery is that the dual objective of the entropically regularized martingale optimal transport problem has an approximately sparse Hessian, and that this sparsity can be exploited inside Sinkhorn-type iterations to obtain a fast algorithm. The authors formulate discrete martingale optimal transport as an entropic OT problem whose (super-)martingale conditions become row-wise equality or inequality constraints on the coupling matrix, then derive Sinkhorn-type updates combined with sparse Newton steps on the dual. They report that the resulting algorithms show super-exponential convergence and robustness, with the total constraint violation controllable through the chosen threshold. The claim is that the approach extends to the prevalent class of OT problems with structural row-wise constraints.","pith_inferences":["The practical payoff depends on how often the dual Hessian is genuinely sparse in realistic instances; for low-temperature (small $\\epsilon$) entropic OT, the kernel becomes nearly rank-one and the Hessian may concentrate, which could help or hurt depending on the cost structure.","Because the method returns an approximate solution with controlled constraint violation, applications that need exact martingale pricing may need a post-projection step; the paper does not claim exact constraint satisfaction.","A natural testable extension is to compare the sparse-Newton Sinkhorn method against a primal-dual first-order solver on high-dimensional marginals, measuring both total violation and wall-clock time; this would separate the benefit of second-order acceleration from the benefit of entropic smoothing."],"forward_implications":["Discrete martingale optimal transport becomes solvable by an iterative entropic method with second-order acceleration, so larger instances than those reachable by dense Newton or plain Sinkhorn become practical.","Any optimal transport problem whose structural constraints are row-wise equalities or inequalities on the coupling matrix falls under the same algorithm, broadening the class of solvable structured OT problems.","The entropic formulation gives the user a controllable trade-off: smaller regularization improves constraint satisfaction but makes the problem harder, and the threshold on total constraint violations can be set in advance.","If the observed super-exponential convergence holds beyond the tested regimes, high-accuracy solutions require only a few outer iterations, making the method competitive with specialized MOT solvers."],"supporting_citations":[],"fun_headline_variants":["Sparse Newton Sinkhorn for fast martingale transport","Entropic martingale transport speed up via sparse Newton","Sinkhorn with sparse Newton handles martingale OT","Fast approximate martingale transport with sparsity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method works only if the Hessian of the dual objective is approximately sparse in the regimes that matter, and only if an entropically regularized solution that approximately satisfies the martingale constraints is acceptable.","fun_headline_variants_meta":{"raw":{"variants":["Sparse Newton Sinkhorn for fast martingale transport","Entropic martingale transport speed up via sparse Newton","Sinkhorn with sparse Newton handles martingale OT","Fast approximate martingale transport with sparsity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000235,"raw_usage":{"total_tokens":1444,"prompt_tokens":834,"completion_tokens":610,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":548}},"tokens_in":450,"tokens_out":610,"duration_ms":6401,"temperature":1.0,"reasoning_tokens":548,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:00:57.316962+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a discrete martingale OT instance with a dense, nonlocal cost matrix and measure the density of the dual Hessian at iteration one; if the Hessian is not approximately sparse, or if a truncated sparse Newton step fails to decrease the dual objective, the central convergence claim fails. Alternatively, run the algorithm to its stopping threshold and compute the total martingale constraint violation on a fine grid of regularization strengths; if the violation cannot be made to track the prescribed threshold, the controllability claim fails.","supporting_citations":[],"review_version":2}