Pith. sign in

REVIEW 3 major objections 4 minor 22 references

Nested Optimal Transport Distances

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper proposes a parallel, statistically consistent algorithm for computing nested optimal transport distances on quantized financial time series.

desk verdict Plausible parallel AW2 algorithm, but the headline speedup is asserted, not measured; the degenerate-Gaussian validation rests on an unproved extension. read the letter →

arxiv 2509.06702 v1 pith:I35ZHRAA submitted 2025-09-08 cs.LG q-fin.CP

Ruben Bontorno , Songyan Hou This is my paper
classification cs.LGq-fin.CP MSC 49Q2260B1060G07
keywords adaptedWassersteindistancenestedoptimaltransportfinancialtimeseriesgenerativemodelevaluationquantizationparalleldynamicprogrammingstatisticalconsistencyGaussianclosedform
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that nested (adapted) Wasserstein-2 distance is the right metric for judging whether synthetic financial time series can stand in for real data in decision tasks, because optimal stopping, hedging, and related dynamic problems are Lipschitz in this metric but not in ordinary Wasserstein or MMD distances. Its main contribution is an algorithm that makes the metric computationally usable: quantize samples onto a lattice, build a tree of shared histories, and run the dynamic programming principle backward with all branches processed in parallel. The paper proves the quantized distance converges almost surely to the true distance as sample size grows, and reports order-of-magnitude speedups over existing implementations. Numerical tests on Ornstein-Uhlenbeck paths and on a constructed fake Brownian motion show the algorithm matching closed-form Gaussian values, and only the nested distance separating the fake process from real Brownian motion.

What carries the argument

Adapted empirical measures with lattice quantization: sample paths are rounded to cube centers with grid size ΔN = N^(-1/(dT)), turning repeated prefixes into a tree. The backward dynamic programming recursion of Proposition 1 computes the nested distance recursively from conditional couplings of these quantized empirical measures; the key source of parallelism is that the value at every node pair depends only on descendant node pairs, so sibling branches are independent. The Markovian improvement replaces conditioning on the entire past by conditioning on the last value, which improves the convergence rate and reduces runtime.

What would settle it

Take the degenerate Gaussian law of L_{δ,t}Z from the fake Brownian motion example with δ=0.1 and t=0.5, and compute AW2 to the corresponding Brownian marginals by an independent method that does not rely on the Cholesky formula, such as direct finite-dimensional dynamic programming on a fine discretization of the three time points. If the values disagree with the closed form used in the figure, or if different choices of Cholesky factor change the answer, then the paper's central numerical claim that AW2 distinguishes the fake process is not established.

Watch

Extended reading notes

Core claim

The central discovery is that the adapted Wasserstein distance, defined by couplings that respect the flow of time, has a computational path that is both naturally parallel and statistically consistent. The path is quantize, build a tree, and backtrack: after mapping each sample path to a lattice point, coinciding prefixes give a tree structure, so the conditional laws required by the dynamic programming principle can be estimated locally along branches. The same dynamic programming recursion then runs backward over time, and since each branch's subproblem depends only on its descendants, all branches can be computed in parallel. The paper states that AW2 between the underlying law and its a

Load-bearing premise

The fake Brownian motion validation assumes the non-degenerate Gaussian closed form for AW2 also applies to the degenerate three-dimensional Gaussian in the example; the paper itself notes degenerate Gaussians lack a unique Cholesky decomposition and that the needed extension rests on an unproved assertion about filtered processes.

Editorial extensions

If this is right

  • AW2 between a generative model's output and the real distribution can be estimated from samples alone, with the estimation error provably going to zero as N grows.
  • The metric becomes practical for long or sample-heavy financial time series, especially in Markovian models where the convergence rate is O(N^(-1/(2d))) and independent of T.
  • Generative financial models can be compared with a metric that is Lipschitz-continuous for dynamic decision problems, so small metric distance implies small differences in optimal stopping, hedging, or risk values.
  • The quantization-plus-tree DP scheme is naturally parallel across branches, so wall-clock time scales with available cores, enabling larger experiments.
  • Because W2 failed to separate the fake Brownian motion while AW2 succeeded, metric choice changes model-comparison conclusions: AW2 can detect time-causal structure, such as predictability, that W2 ignores.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension the paper does not develop is to use the parallel AW2 estimator as a training objective for generative financial models, not merely an evaluation metric.
  • The fake Brownian motion example suggests a practical screening device: when a candidate path model has large AW2 distance from a reference process but small W2 distance, the discrepancy is likely caused by exploitable time-causal structure, which may signal arbitrage-like behavior.
  • Because the algorithm's behavior depends on the quantization grid and the resulting tree collapse, adaptive or data-dependent grid choice is a plausible next step that could improve finite-sample accuracy beyond the worst-case rate.
  • A direct proof of the AW2 closed form for degenerate discrete-time Gaussian laws, rather than delegation to filtered processes, would place the numerical validation on firmer ground; a concrete check is whether different Cholesky factors of the same degenerate covariance produce different AW2 values, as the paper suggests.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper argues that the adapted (nested) Wasserstein distance AW2 is a suitable metric for evaluating generative models of financial time series, and proposes an algorithm for computing it. The algorithm first quantizes i.i.d. samples from the two distributions onto a lattice (adapted empirical measures), then computes AW2 exactly for the quantized measures by backward dynamic programming along the tree structure induced by shared quantized prefixes. A Markovian variant replaces the full conditional distribution given the past by the conditional distribution given the current state. The paper claims statistical consistency via Theorem 2 (citing Backhoff et al. and Acciaio–Hou), natural parallelizability, and substantial/orders-of-magnitude speedups over existing AW2 implementations. Numerical experiments compare the algorithm against Gaussian closed forms for an Ornstein–Uhlenbeck example and a 'fake Brownian motion' example, reporting convergence of the computed values and runtimes of the proposed method. The conclusion reiterates the consistency and speedup claims.

Significance. If the central claims are correct, the paper provides a useful algorithmic perspective: the quantization-plus-backward-DP construction is natural, and the consistency of the estimator follows from existing work rather than requiring new theory. The Markovian variant has the potential to mitigate the curse of dimensionality for a practically important class of models. However, the main advertised contribution—'substantial speedups over existing approaches'—is not measured against any AW2 baseline, and one of the two numerical validations rests on an unproved extension of a Gaussian closed form to a degenerate case that the paper itself flags as non-unique. The paper currently contains no code, so the runtime claims are not reproducible. For these reasons the significance of the contribution cannot be fully assessed until the advertised speedup and the degenerate-Gaussian ground truth are substantiated.

major comments (3)
  1. [§1 and §3] The abstract and introduction claim 'substantial speedups' and 'orders-of-magnitude speedups over existing implementations' for computing AW2, but the experimental section reports no comparison with any existing AW2 solver. Figures 1 and 3 show only wall-clock times for the proposed algorithm (and POT is used for W2, not for AW2). There is also no complexity bound for the backward DP / tree traversal, and no code released. Since the speedup is a headline claim, this is load-bearing. Please provide a benchmark against the cited implementations [8]-[10] on identical hardware with matched accuracy, or substantially temper the speedup claims.
  2. [§3, Figure 3, and Appendix A.4] The fake-Brownian experiment validates the algorithm against the theoretical AW2 value obtained from Theorem 4, which is stated for non-degenerate Gaussians. The distribution µX_{δ,t,1} in this experiment is degenerate: in L_{δ,t}, row 1 is √δ times row 3. The paper itself acknowledges in Appendix A.4 that degenerate Gaussians have non-unique Cholesky decompositions and that the closed form only extends through the filtered-process construction of [22], an extension that is asserted but not proved. As written, the numerical validation that AW2 distinguishes the fake Brownian motion from Brownian motion is unsupported. Please either replace the test with a non-degenerate Gaussian example for which Theorem 4 directly applies, or provide a rigorous derivation of the degenerate closed form and specify exactly which Cholesky factor/filtration is used. In addition, the displayed L_{δ,t} appear
  3. [§2, Markovian improvement] The statistical consistency of the Markovian variant is not formally stated. Theorem 2 applies to the adapted empirical measure Rhatµ_N of Definition 3, with conditional distributions conditioned on the full quantized past q_{1:t}. The Markovian implementation instead replaces these by conditionals given q_t. The paper points to [14, Theorem 6.1] for the improved rate, but does not state the precise assumptions or prove that the quantized Markovian estimator satisfies them. Since the OU experiment uses the Markovian implementation, this is not purely cosmetic. Please state the relevant convergence theorem for the Markovian variant explicitly and verify its hypotheses, or limit the consistency claim to the non-Markovian algorithm.
minor comments (4)
  1. [§3, first paragraph] The text says 'Let N=5, Δt=1/N', but N is also used for the number of samples in Definition 3 and in the figures. It appears the time horizon is being fixed at T=5 while the sample size varies. Please use separate symbols for the time horizon and the sample size.
  2. [§3, Figure 3 caption] The sentence 'The AW2 is calculated with our algorithm with non-Markovian implementation and the W2 is calculated by ot.lp.emd(POT)' is ambiguous: the preceding text also mentions 'PNOT: Python Nested Optimal Transport'. Is PNOT the name of the proposed implementation or an external package? Please clarify the relation between PNOT, the proposed algorithm, and POT.
  3. [Definition 3] Definition 3 states µ ∈ P1(R^{dT}), while Theorem 2 assumes µ ∈ P2(R^{dT}). Please align the moment assumptions.
  4. [Appendix A.4, Theorem 5] Theorem 5 is cited to [21], an arXiv preprint, and is used for the d>1 claim. Please give a proof or a more complete citation to a peer-reviewed version if one exists.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the algorithm's derivation is self-contained; convergence and closed-form results rely on independent published theorems.

full rationale

The paper's derivation chain for the algorithm is: define AW2 via bicausal couplings (Def. 2), state the dynamic programming principle (Prop. 1, proved by citing [17]), define adapted empirical measures (Def. 3), and cite a published convergence theorem (Theorem 2, proved by citing [18, Thm 2.7]) to guarantee that the quantized empirical measures converge in AW2. The algorithm then exactly evaluates AW2 between the quantized measures by backward DP; no fitted parameter enters and the quantization scale Δ_N is prescribed as N^{-1/(dT)} by Theorem 2, not tuned to match the target. The Gaussian closed forms used as benchmarks are external results ([19] for W2, [20] for AW2 in one dimension, [21] for general d); although [21] is co-authored by Hou, it is a published theorem independent of the present paper's algorithm. The only notable weakness is the Appendix A.4 assertion that the closed-form extends to degenerate Gaussians 'by following the same proof in [21]'; this is an omitted proof and a correctness risk for the fake-Brownian-motion validation, but it does not feed back into the algorithm's derivation, so it is not circular. Similarly, the promised 'substantial speedups' are not benchmarked against existing AW2 solvers, but that is a missing comparison, not a circular reduction. No prediction in the paper reduces to its inputs by construction.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new postulates or entities; all mathematical content is imported from cited literature, and the fake Brownian motion is an example, not an invented physical or theoretical entity.

free parameters (2)
  • Quantization cube edge length Δ_N = N^{-1/(dT)}
    Chosen from the theoretical rate in [14, 18], not fitted to data; it controls approximation error and the tree structure.
  • Fake Brownian motion parameters δ, t = δ = 0.1, t = 0.5
    Fixed for the demonstration example; chosen by hand to make the plain Wasserstein distance fail to distinguish the processes, not fitted to data.
assumptions (4)
  • standard math Dynamic programming principle for AW2 (Proposition 1)
    Invoked to compute AW2 via backward induction; proof deferred to [17, Prop 5.2].
  • standard math Almost-sure convergence of adapted empirical measures for Δ_N = N^{-1/(dT)} (Theorem 2)
    Cornerstone of statistical consistency; proof deferred to [18, Thm 2.7].
  • standard math Markovian improvement rate O(N^{-1/(2d)}) independent of T
    Used to justify the Markovian implementation for the OU example; taken from [14, Thm 6.1].
  • domain assumption Closed-form AW2 for degenerate Gaussians via chosen Cholesky or filtered process representation
    Used in the fake Brownian motion validation, even though Theorem 4 assumes non-degenerate Gaussians; Appendix A.4 appeals to [22] without proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Nested Optimal Transport Distances." pith.science (2026). https://pith.science/paper/I35ZHRAA

@misc{pith2026250906702,
  author       = {Pith},
  title        = {Pith review of: Nested Optimal Transport Distances},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I35ZHRAA}},
  note         = {Machine review of arXiv:2509.06702}
}
read the original abstract

Simulating realistic financial time series is essential for stress testing, scenario generation, and decision-making under uncertainty. Despite advances in deep generative models, there is no consensus metric for their evaluation. We focus on generative AI for financial time series in decision-making applications and employ the nested optimal transport distance, a time-causal variant of optimal transport distance, which is robust to tasks such as hedging, optimal stopping, and reinforcement learning. Moreover, we propose a statistically consistent, naturally parallelizable algorithm for its computation, achieving substantial speedups over existing approaches.

Figures

Figures reproduced from arXiv: 2509.06702 by the authors.

Figure 1
Figure 1. Numerical optimal values, absolute errors, and runtimes for [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualization of empirical measures and adapted empirical measures with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Numerical optimal values, absolute errors, and runtimes for [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 18 canonical work pages

  1. [17]

    Causal transport in discrete time and applications.SIAM Journal on Optimization, 27(4):2528–2562, 2017

    Julio Backhoff-Veraguas, Mathias Beiglbock, Yiqing Lin, and Anastasiia Zalashko. Causal transport in discrete time and applications.SIAM Journal on Optimization, 27(4):2528–2562, 2017

    work page 2017

  2. [18]

    Convergence of adapted empirical measures on Rd.The Annals of Applied Probability, 34(5):4799–4835, 2024

    Beatrice Acciaio and Songyan Hou. Convergence of adapted empirical measures on Rd.The Annals of Applied Probability, 34(5):4799–4835, 2024

    work page 2024

  3. [14]

    Estimating processes in adapted wasserstein distance.The Annals of Applied Probability, 32(1):529–550, 2022

    Julio Backhoff, Daniel Bartl, Mathias Beiglböck, and Johannes Wiesel. Estimating processes in adapted wasserstein distance.The Annals of Applied Probability, 32(1):529–550, 2022

    work page 2022

  4. [22]

    The wasserstein space of stochastic processes.Journal of the European Mathematical Society, 2024

    Daniel Bartl, Mathias Beiglböck, and Gudmund Pammer. The wasserstein space of stochastic processes.Journal of the European Mathematical Society, 2024. A Technical Appendices and Supplementary Material A.1 Notations We regardRdT as the space ofd-dimensional discrete-time paths withT time steps, x= (x 1, . . . , xT ), equipped with the Euclidean norm ∥ · ∥....

    work page 2024

  5. [8]

    The nested sinkhorn divergence to learn the nested distance.Computational Management Science, 19(2):269–293, 2022

    Alois Pichler and Michael Weinhardt. The nested sinkhorn divergence to learn the nested distance.Computational Management Science, 19(2):269–293, 2022

    work page 2022

  6. [10]

    Fitted value iteration methods for bicausal optimal transport

    Erhan Bayraktar and Bingyan Han. Fitted value iteration methods for bicausal optimal transport. arXiv preprint arXiv:2306.12658, 2023

    work page arXiv 2023

  7. [1]

    Generating synthetic data in finance: opportunities, challenges and pitfalls

    Samuel A Assefa, Danial Dervovic, Mahmoud Mahfouz, Robert E Tillman, Prashant Reddy, and Manuela Veloso. Generating synthetic data in finance: opportunities, challenges and pitfalls. InProceedings of the First ACM International Conference on AI in Finance, pages 1–8, 2020

    work page 2020

  8. [2]

    Deep generative modeling for financial time series with application in var: A comparative review

    Lars Ericson, Xuejun Zhu, Xusi Han, Rao Fu, Shuang Li, Steve Guo, and Ping Hu. Deep generative modeling for financial time series with application in var: A comparative review. arXiv preprint arXiv:2401.10370, 2024

    arXiv 2024

Show all 22 references
  1. [3]

    Synthetic data for portfolios: A throw of the dice will never abolish chance.arXiv e-prints, pages arXiv–2501, 2025

    Adil Rengim Cetingoz and Charles-Albert Lehalle. Synthetic data for portfolios: A throw of the dice will never abolish chance.arXiv e-prints, pages arXiv–2501, 2025

    2025

  2. [4]

    Adapted wasserstein distances and stability in mathematical finance.Finance and Stochastics, 24:601–632, 2020

    Julio Backhoff-Veraguas, Daniel Bartl, Mathias Beiglböck, and Manu Eder. Adapted wasserstein distances and stability in mathematical finance.Finance and Stochastics, 24:601–632, 2020

    2020

  3. [5]

    Time consistent dynamic risk processes.Stochastic Processes and their Applications, 119(2):633–654, 2009

    Jocelyne Bion-Nadal. Time consistent dynamic risk processes.Stochastic Processes and their Applications, 119(2):633–654, 2009

    2009

  4. [6]

    Incorporating statistical model error into the calculation of acceptability prices of contingent claims.Mathematical Programming, 174 (1):499–524, 2019

    Martin Glanzer, Georg Ch Pflug, and Alois Pichler. Incorporating statistical model error into the calculation of acceptability prices of contingent claims.Mathematical Programming, 174 (1):499–524, 2019

    2019

  5. [7]

    Sensitivity of multiperiod optimization problems with respect to the adapted wasserstein distance.SIAM Journal on Financial Mathematics, 14(2):704–720, 2023

    Daniel Bartl and Johannes Wiesel. Sensitivity of multiperiod optimization problems with respect to the adapted wasserstein distance.SIAM Journal on Financial Mathematics, 14(2):704–720, 2023

    2023

  6. [9]

    Computational methods for adapted optimal transport

    Stephan Eckstein and Gudmund Pammer. Computational methods for adapted optimal transport. The Annals of Applied Probability, 34(1A):675–713, 2024

    2024

  7. [11]

    Springer, 2014

    Georg Ch Pflug and Alois Pichler.Multistage stochastic optimization, volume 1104. Springer, 2014

    2014

  8. [12]

    Causal optimal transport and its links to enlargement of filtrations and continuous-time stochastic optimization.Stochastic Processes and their Applications, 130(5):2918–2953, 2020

    Beatrice Acciaio, Julio Backhoff-Veraguas, and Anastasiia Zalashko. Causal optimal transport and its links to enlargement of filtrations and continuous-time stochastic optimization.Stochastic Processes and their Applications, 130(5):2918–2953, 2020

    2020

  9. [13]

    Time-causal vae: Robust financial time series generator.arXiv preprint arXiv:2411.02947, 2024

    Beatrice Acciaio, Stephan Eckstein, and Songyan Hou. Time-causal vae: Robust financial time series generator.arXiv preprint arXiv:2411.02947, 2024

    2024 arXiv

  10. [15]

    On the rate of convergence in wasserstein distance of the empirical measure.Probability Theory and Related Fields, 162(3):707–738, 2015

    Nicolas Fournier and Arnaud Guillin. On the rate of convergence in wasserstein distance of the empirical measure.Probability Theory and Related Fields, 162(3):707–738, 2015

    2015

  11. [16]

    Pot: Python optimal transport.Journal of Machine Learning Research, 22(78):1–8, 2021

    Rémi Flamary, Nicolas Courty, Alexandre Gramfort, Mokhtar Z Alaya, Aurélie Boisbunon, Stanislas Chambon, Laetitia Chapel, Adrien Corenflos, Kilian Fatras, Nemo Fournier, et al. Pot: Python optimal transport.Journal of Machine Learning Research, 22(78):1–8, 2021

    2021

  12. [19]

    On wasserstein geometry of gaussian measures.Probabilistic approach to geometry, 57:463–472, 2010

    Asuka Takatsu. On wasserstein geometry of gaussian measures.Probabilistic approach to geometry, 57:463–472, 2010. 5

    2010

  13. [20]

    Adapted optimal transport between gaussian processes in discrete time.Electronic Communications in Probability, 30:1–14, 2025

    Madhu Gunasingam and Ting-Kam Leonard Wong. Adapted optimal transport between gaussian processes in discrete time.Electronic Communications in Probability, 30:1–14, 2025

    2025

  14. [21]

    Entropic adapted wasserstein distance on gaussians.arXiv preprint arXiv:2412.18794, 2024

    Beatrice Acciaio, Songyan Hou, and Gudmund Pammer. Entropic adapted wasserstein distance on gaussians.arXiv preprint arXiv:2412.18794, 2024

    2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.