Pith. sign in

REVIEW 3 major objections 5 minor 31 references

Convex local relaxations optimize high-dimensional Markov interpolations and recover Benamou–Brenier velocity and Glauber kernels from duals and local statistics.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 03:05 UTC pith:XUTN66YS

load-bearing objection Clean multistage extension of the authors' static cluster OT hierarchy, with exact Benamou–Brenier recovery before spatial relaxation and usable numerics. the 3 major comments →

arxiv 2607.09423 v1 pith:XUTN66YS submitted 2026-07-10 math.OC cs.NAmath.NA

Convex Relaxations for the Optimization of Markov Processes

Hongyi Zhang , Yuehaw Khoo , Tianyun Tang This is my paper
classification math.OC cs.NAmath.NA MSC 90C2249Q2260J1090C25
keywords Markov process optimizationsequential couplingsconvex relaxationcluster momentsdynamic optimal transportBenamou–BrenierGlauber dynamicshigh-dimensional probability
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper attacks the problem of finding a cost-minimizing Markov process that connects two prescribed distributions when the state space is high-dimensional. Full couplings or distributions become intractable, so the authors rewrite the problem as a chain of sequential couplings and replace each coupling by a sparse collection of local marginals or cluster moments. The resulting convex programs give certified lower bounds and recover low-order statistics of every intermediate law. In the unconstrained kinetic-cost case the formulation is exactly dynamic optimal transport: primal solutions recover the Benamou–Brenier measure curve on any prescribed time grid, and dual potentials recover the velocity field. The same local statistics can be fitted to parametric kernels such as single-spin Glauber dynamics, producing an implementable process between Ising models. A sympathetic reader cares because the method supplies both computable certificates and reconstructible dynamics without ever storing a full high-dimensional joint law.

Core claim

After time discretization with kinetic cost and no kernel constraints, the sequential-coupling problem has the same optimal value as continuous Benamou–Brenier dynamics; its primal solutions recover the geodesic measures at the grid times, its dual potentials recover the velocity under standard support assumptions, and the same local statistics can be fitted to general parametric Markov kernels.

What carries the argument

The sequential-coupling formulation of the multistage Markov problem, together with its marginal and cluster-moment relaxations that retain only low-order local information on a chosen cluster graph and link consecutive times by projected mass-conservation constraints.

Load-bearing premise

The method works only when the chosen clusters and cluster graph already capture enough of the local dependence that the projected constraints stay tight; if locality is lost along the path, the lower bounds and recovered dynamics can become loose.

What would settle it

On a Gaussian-to-Gaussian Benamou–Brenier instance whose exact geodesic and velocity are known in closed form, solve the moment relaxation on a path cluster graph of increasing radius and check whether recovered means, covariances and affine velocity coefficients converge to the analytic values; persistent large relative error after the graph becomes dense would falsify the recovery claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies optimization of time-inhomogeneous Markov processes that interpolate between prescribed endpoint laws while minimizing additive one-step costs, subject to admissible-kernel constraints. To mitigate the curse of dimensionality, it reformulates the problem as sequential couplings and develops convex outer approximations that retain only local cluster marginals or cluster moments linked by consistency, mass-conservation, and projected kernel constraints. Dynamic optimal transport is identified as the unconstrained kinetic-cost special case: after time discretization the sequential-coupling problem has the same optimal value as the continuous Benamou–Brenier problem and recovers grid-time measures (Proposition 3.1); dual potentials recover the velocity under standard connected-support hypotheses (Proposition 3.2), and duals of the moment relaxation yield approximate velocities via SOS certificates. For more general constrained processes the recovered local statistics are fitted to a parametric kernel family, illustrated by single-spin Glauber dynamics between Ising models. Numerical experiments on Gaussian–Gaussian and Gaussian–Ginzburg–Landau transport and on 1D/2D Ising processes support the recovery procedures.

Significance. If the claims hold, the work supplies a principled convex pathway for high-dimensional dynamic transport and constrained Markov-process optimization that yields certified lower bounds and low-order intermediate statistics without full joint laws or nonconvex neural training. The Benamou–Brenier equivalence (Propositions 3.1–3.2) is clean and free of time-discretization error in the value, which is a useful structural contribution relative to Eulerian grid methods and sample-based flow matching. Extending sparse cluster-marginal/moment ideas from static OT to multistage sequential couplings, and converting relaxed statistics into implementable kernels (Glauber fitting), is a natural and practically relevant step. Strengths include explicit dual recovery formulas, transparent outer-approximation constructions, and reproducible numerical illustrations against closed-form Gaussian geodesics. The main practical caveat—dependence of tightness on the user-chosen cluster graph—is acknowledged as open in Section 6 and does not undermine the exact pre-relaxation theory.

major comments (3)
  1. Section 5.1 / Figure 2: for the Gaussian benchmark the paper reports relative covariance and velocity-coefficient errors as the path-graph radius r varies, but does not report the corresponding gap between the moment-relaxation objective and the known closed-form W_2^2. Because the central selling point of the convex program is a computable lower bound, documenting how the objective gap closes with r (and with complete graph) is load-bearing for assessing practical tightness of (Mom2) even in the setting where the exact value is known.
  2. Section 4.2 and Section 5.3 (Figures 8, 10): the Glauber fitting procedure is validated by matching recovered local marginals, but the manuscript never compares the one-step cost (or total action) realized by the fitted kernels against the LP lower bound from (Mar1). Without that comparison it is unclear whether the recovered dynamics is near-optimal for problem (4.8) or merely reproduces selected marginals of some (possibly suboptimal) process consistent with the outer approximation.
  3. Section 3.3, equations (3.10)–(3.12): the velocity field extracted from the SOS dual of (Mom2) is presented as an approximation of the Benamou–Brenier velocity, yet no quantitative relation (e.g., consistency as moment degree or cluster graph is refined, or even empirical dual-gap / velocity error tables beyond the Gaussian affine case) is given. A short statement of what is rigorously guaranteed versus what is heuristic would clarify the scope of the recovery claim for non-Gaussian targets such as the Ginzburg–Landau experiment.
minor comments (5)
  1. Section 2.2, display (2.16): the global positivity PSD is written with Diag(\pi_A^k); a one-line remark that this is the finite-state realization of the L^2 test-function inequality would help continuous-state readers.
  2. Notation: the same symbols P_A, P_AB are used for projections on \Omega and on \Omega\times\Omega (Section 2.1); a brief disambiguation would reduce ambiguity when reading (2.4)–(2.5) versus (2.11)–(2.13).
  3. Figure 5 caption and surrounding text: the push-forward uses five Euler substeps per interval and clipping; stating the sensitivity of the visualized marginals to these choices (or reporting a single-step comparison) would strengthen reproducibility.
  4. References: the static cluster-moment OT paper [12] is the direct methodological antecedent; ensuring that the dynamic multi-block construction is clearly distinguished from a mere product of independent static relaxations (already stated in §1.4) could be reinforced once more when (Mom) is introduced.
  5. Typographical: “Schr¨ odinger” appears with a broken umlaut in several places (abstract/related work); standardize to “Schrödinger” or “Schrodinger”.

Circularity Check

1 steps flagged

No significant circularity: Benamou–Brenier equivalence and dual recovery are self-contained first-principles arguments; self-citation to the authors’ static OT paper supplies only the static building blocks.

specific steps
  1. self citation load bearing [Section 1.4 Related work; Section 2 opening paragraph]
    "more directly, the cluster marginal and moment relaxations in [12] provide the closest methodological antecedent to our approach for high-dimensional static transport. ... The construction is motivated by the marginal and cluster moment relaxations for static optimal transport in [12]"

    The authors cite their own prior static OT paper [12] as the source of the local-marginal and cluster-moment machinery. This is ordinary methodological self-citation and is not load-bearing for the paper’s new dynamic claims (Propositions 3.1–3.2), which are proved independently of [12]. It therefore contributes only a minor score increment.

full rationale

The central claims (Proposition 3.1: sequential-coupling problem (3.2) has optimal value exactly W_p^p(µ,ν) and recovers grid-time measures of a Benamou–Brenier geodesic; Proposition 3.2 recovers velocity from dual potentials under connected-support assumptions) are proved by elementary reduction to one-step Wasserstein problems, concatenation of geodesics, and standard Kantorovich dual uniqueness, without any fitted parameters or load-bearing self-citation. The dual of the moment relaxation is derived as an SOS certificate of the exact dual (dual 3.2), again from first principles. Numerical validation uses independent closed-form Gaussian geodesics or Monte-Carlo samples generated by separate Glauber chains; the kernel-fitting step (Section 4) is an explicit least-squares projection onto a parametric family and is not presented as a first-principles prediction. The only self-citation of note is to the authors’ static OT paper [12], which supplies the static marginal/moment building blocks; it is not used to force uniqueness or to forbid alternatives for the dynamic claims. The user-chosen cluster graph G is a genuine practical limitation of the outer approximations (flagged by the authors in Section 6), but it is not load-bearing for the exact equivalence statements, which hold before any spatial relaxation. Score 1 reflects a single non-load-bearing self-citation with an otherwise independent derivation.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper rests on standard convex duality and optimal-transport theory plus modeling choices (cluster partition, graph, basis degree, parametric kernel family) that are left to the user. No new physical entities are postulated; free parameters are algorithmic hyperparameters rather than fitted physical constants.

free parameters (4)
  • cluster partition C and cluster-graph edge set E (or graph radius r)
    User-chosen sparsity pattern that determines which local marginals/moments are retained; directly controls both computational cost and tightness of the relaxation.
  • moment degree r / monomial degree deg_mon
    Truncation order of the cluster bases; chosen by hand (e.g., deg_mon=11) and affects both accuracy and SDP size.
  • regularization strength λ and ridge floor in Glauber fitting
    Penalty and numerical floor used when converting recovered marginals into Glauber parameters via least-squares or log-odds regression.
  • maximum-entropy moment-matching tolerance ε_mom and grid resolution
    Discretization and soft-matching parameters used to reconstruct continuous low-dimensional marginals from recovered moments.
axioms (4)
  • standard math Standard Borel state space and existence of regular conditional probabilities (disintegration of couplings into Markov kernels).
    Used to equate the kernel formulation (1.1) with the sequential-coupling formulation (1.5).
  • standard math Benamou–Brenier duality and uniqueness of Kantorovich potentials under absolute continuity and connected support (Lemma 3.3, Prop. 3.2).
    Invoked to recover the velocity field from dual potentials of the time-discrete problem.
  • domain assumption Locality/sparsity of interactions is sufficiently preserved that the projected cluster constraints remain a useful outer approximation.
    Implicit throughout Section 2 and listed as an open question in Section 6; without it the lower bounds can be arbitrarily loose.
  • ad hoc to paper Chosen parametric family (e.g., single-site Glauber) is rich enough to realize the recovered local statistics.
    Required for the kernel-fitting step in Section 4; the paper does not prove that a perfect fit always exists.

pith-pipeline@v1.1.0-grok45 · 35364 in / 2978 out tokens · 33360 ms · 2026-07-13T03:05:32.712722+00:00 · methodology

0 comments
read the original abstract

In this paper, we study the problem of optimizing Markov processes that interpolate between two prescribed probability distributions while minimizing a given cost. The main computational challenge is the curse of dimensionality: in high-dimensional state spaces, representing the full distribution is intractable. To address this, we reformulate the problem in terms of sequential couplings and develop convex relaxations based on local marginals and cluster moments. These relaxations exploit locality and sparse interaction structure, provide computable lower bounds, and recover low-order statistics of the intermediate laws. We identify dynamic optimal transport as a special case of our Markov process optimization problem and develop a procedure for recovering the underlying Benamou--Brenier dynamics from the relaxed solution. We also show that the procedure extends to more general Markov processes and illustrate it with a constrained process between Ising models.

Figures

Figures reproduced from arXiv: 2607.09423 by Hongyi Zhang, Tianyun Tang, YueHaw Khoo.

Figure 1
Figure 1. Figure 1: Benamou–Brenier dynamics between Gaussian distributions in dimension [PITH_FULL_IMAGE:figures/full_fig_p027_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Effect of the cluster graph radius on the recovered covariance and velocity field. The [PITH_FULL_IMAGE:figures/full_fig_p027_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison between moment relaxation and back-propagation for the Gaussian benchmark [PITH_FULL_IMAGE:figures/full_fig_p029_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Intermediate two-dimensional marginals for transport from a Gaussian law to a Ginzburg– [PITH_FULL_IMAGE:figures/full_fig_p031_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Velocity-induced push-forward for transport from a Gaussian law to a Ginzburg–Landau [PITH_FULL_IMAGE:figures/full_fig_p032_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Cluster decomposition and cluster graph for the 4 [PITH_FULL_IMAGE:figures/full_fig_p033_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Recovered marginals on cluster A1 = {1, 2} for the 1D Ising experiment with d = 30 spins. Across the five epochs, the recovered marginal transfers mass from aligned spin configurations, favored by the ferromagnetic source, to anti-aligned configurations, favored by the antiferromagnetic target. After obtaining the local marginals from the relaxation, we fit Glauber dynamics using the linear regression form… view at source ↗
Figure 8
Figure 8. Figure 8: Fitted Glauber dynamics reproduces the recovered two-spin marginals in the 1D Ising [PITH_FULL_IMAGE:figures/full_fig_p034_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Representative trajectory generated by the fitted Glauber dynamics in the 1D Ising [PITH_FULL_IMAGE:figures/full_fig_p035_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Fitted Glauber dynamics reproduces the recovered four-spin marginals in the 2D Ising [PITH_FULL_IMAGE:figures/full_fig_p036_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 3 canonical work pages

  1. [1]

    Mosek optimization toolbox for matlab

    Mosek Aps. “Mosek optimization toolbox for matlab”. In:User’s Guide and Reference Manual Version4.1 (2019), p. 116

    2019

  2. [2]

    Eustasio del Barrio, Alberto Gonz´ alez-Sanz, and Jean-Michel Loubes.Central Limit Theorems for General Transportation Costs. 2021. arXiv: 2102.06379 [math.ST].url: https://arxiv. org/abs/2102.06379

    Pith/arXiv arXiv 2021

  3. [3]

    A computational fluid mechanics solution to the Monge-Kantorovich mass transfer problem

    Jean-David Benamou and Yann Brenier. “A computational fluid mechanics solution to the Monge-Kantorovich mass transfer problem”. In:Numerische Mathematik84.3 (2000), pp. 375– 393

    2000

  4. [4]

    On the Relation Between Optimal Transport and Schr¨ odinger Bridges: A Stochastic Control Viewpoint

    Y. Chen, T. T. Georgiou, and M. Pavon. “On the Relation Between Optimal Transport and Schr¨ odinger Bridges: A Stochastic Control Viewpoint”. In:Journal of Optimization Theory and Applications169 (2016), pp. 671–691.doi:10.1007/s10957-015-0803-z

    doi:10.1007/s10957-015-0803-z 2016

  5. [5]

    Convex relaxation for Fokker–Planck equation

    Yian Chen, Yuehaw Khoo, and Lek-Heng Lim. “Convex relaxation for Fokker–Planck equation”. In:Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 481.2313 (2025), p. 20240001.url:https://doi.org/10.1098/rspa.2024.0001

    doi:10.1098/rspa.2024.0001 2025

  6. [6]

    Diffusion Schr¨ odinger Bridge with Applications to Score-Based Generative Modeling

    Valentin De Bortoli et al. “Diffusion Schr¨ odinger Bridge with Applications to Score-Based Generative Modeling”. In:Advances in Neural Information Processing Systems. Vol. 34. Curran Associates, Inc., 2021, pp. 17695–17709.url: https://proceedings.neurips.cc/paper/ 2021/hash/940392f5f32a7ade1cc201767cf83e31-Abstract.html

    2021

  7. [7]

    Approximation and sampling of multivariate probability distributions in the tensor train decomposition

    Sergey Dolgov et al. “Approximation and sampling of multivariate probability distributions in the tensor train decomposition”. In:Statistics and Computing30.3 (2020), pp. 603–625

    2020

  8. [8]

    Dynamical optimal transport of nonlinear control-affine systems

    Karthik Elamvazhuthi et al. “Dynamical optimal transport of nonlinear control-affine systems”. In:Journal of Computational Dynamics10.4 (2023), pp. 425–449.url: https : / / www . aimsciences.org/article/id/64f1b01f89e5d06909db800b

    2023

  9. [9]

    Chris Finlay et al.How to train your neural ODE: the world of Jacobian and kinetic regular- ization. 2020. arXiv:2002.02798 [stat.ML].url:https://arxiv.org/abs/2002.02798

    Pith/arXiv arXiv 2020

  10. [10]

    On the rate of convergence in Wasserstein distance of the empirical measure

    Nicolas Fournier and Arnaud Guillin. “On the rate of convergence in Wasserstein distance of the empirical measure”. In:Probability Theory and Related Fields162.3–4 (2015), pp. 707–738. doi:10.1007/s00440-014-0583-7

    doi:10.1007/s00440-014-0583-7 2015

  11. [11]

    On the translocation of masses

    L. V. Kantorovich. “On the translocation of masses”. In:C. R. (Doklady) Acad. Sci. URSS (N.S.)37 (1942), pp. 199–201

    1942

  12. [12]

    Yuehaw Khoo and Tianyun Tang.Convex relaxation approaches for high-dimensional optimal transport. 2025. arXiv:2511.13847 [math.OC].url:https://arxiv.org/abs/2511.13847

    Pith/arXiv arXiv 2025

  13. [13]

    Discrete Diffusion Schr¨ odinger Bridge Matching for Graph Trans- formation

    Jun Hyeong Kim et al. “Discrete Diffusion Schr¨ odinger Bridge Matching for Graph Trans- formation”. In:The Thirteenth International Conference on Learning Representations. 2025. arXiv:2410.01500 [cs.LG].url:https://openreview.net/forum?id=tQyh0gnfqW

    Pith/arXiv arXiv 2025

  14. [14]

    Nikita Kornilov et al.Optimal Flow Matching: Learning Straight Trajectories in Just One Step. 2024. arXiv:2403.13117 [stat.ML].url:https://arxiv.org/abs/2403.13117

    Pith/arXiv arXiv 2024

  15. [15]

    Convergent SDP-Relaxations in Polynomial Optimization with Sparsity

    Jean B. Lasserre. “Convergent SDP-Relaxations in Polynomial Optimization with Sparsity”. In:SIAM Journal on Optimization17.3 (2006), pp. 822–843.doi:10.1137/05064504X

    work page doi:10.1137/05064504x 2006

  16. [16]

    Jean Bernard Lasserre.Moments, positive polynomials and their applications. Vol. 1. World Scientific, 2009. 37

    2009

  17. [17]

    A survey of the Schr¨ odinger problem and some of its connections with optimal transport

    Christian L´ eonard. “A survey of the Schr¨ odinger problem and some of its connections with optimal transport”. In:Discrete and Continuous Dynamical Systems - A34.4 (2014), pp. 1533– 1574.url:https://arxiv.org/abs/1308.0215

    Pith/arXiv arXiv 2014

  18. [18]

    Wuchen Li and Stanley Osher.Constrained dynamical optimal transport and its Lagrangian formulation. 2018. arXiv: 1807.00937 [math.OC] .url: https://arxiv.org/abs/1807. 00937

    Pith/arXiv arXiv 2018

  19. [19]

    Olga Mula and Anthony Nouy.Moment-SoS Methods for Optimal Transport Problems. 2022. arXiv:2211.10742 [math.NA].url:https://arxiv.org/abs/2211.10742

    Pith/arXiv arXiv 2022

  20. [20]

    Optimal Transport with Proximal Splitting

    Nicolas Papadakis, Gabriel Peyr´ e, and Edouard Oudet. “Optimal Transport with Proximal Splitting”. In:SIAM Journal on Imaging Sciences7.1 (Jan. 2014), pp. 212–238.issn: 1936-4954. doi:10.1137/130920058.url:http://dx.doi.org/10.1137/130920058

    work page doi:10.1137/130920058.url:http://dx.doi.org/10.1137/130920058 2014

  21. [21]

    Parrilo and Rekha R

    Pablo A. Parrilo and Rekha R. Thomas, eds.Sum of Squares: Theory and Applications. Vol. 77. Proceedings of Symposia in Applied Mathematics. American Mathematical Society, 2020. isbn: 978-1-4704-5025-0

    2020

  22. [22]

    Yifan Peng et al.Tensor Density Estimator by Convolution-Deconvolution. 2025. arXiv: 2412.18964 [math.NA].url:https://arxiv.org/abs/2412.18964

    arXiv 2025

  23. [23]

    Filippo Santambrogio.Optimal Transport for Applied Mathematicians: Calculus of Varia- tions, PDEs, and Modeling. Vol. 87. Progress in Nonlinear Differential Equations and Their Applications. Cham: Birkh¨ auser, 2015.doi:10.1007/978-3-319-20828-2

    doi:10.1007/978-3-319-20828-2 2015

  24. [24]

    Diffusion Schr¨ odinger Bridge Matching

    Yuyang Shi et al. “Diffusion Schr¨ odinger Bridge Matching”. In:Advances in Neural Information Processing Systems. Vol. 36. Curran Associates, Inc., 2023, pp. 62183–62223.doi: 10.52202/ 075280-2717

    2023

  25. [25]

    Alexander Tong et al.Improving and generalizing flow-based generative models with minibatch optimal transport. 2024. arXiv: 2302.00482 [cs.LG] .url: https://arxiv.org/abs/2302. 00482

    Pith/arXiv arXiv 2024

  26. [26]

    Alexander Tong et al.TrajectoryNet: A Dynamic Optimal Transport Network for Modeling Cellular Dynamics. 2020. arXiv: 2002.04461 [stat.ML] .url: https://arxiv.org/abs/ 2002.04461

    Pith/arXiv arXiv 2020

  27. [27]

    C´ edric Villani.Optimal Transport: Old and New. Vol. 338. Grundlehren der mathematischen Wissenschaften. Springer, 2009

    2009

  28. [28]

    Sums of Squares and Semidefinite Program Relaxations for Polynomial Optimization Problems with Structured Sparsity

    Hayato Waki et al. “Sums of Squares and Semidefinite Program Relaxations for Polynomial Optimization Problems with Structured Sparsity”. In:SIAM Journal on Optimization17.1 (2006), pp. 218–242.doi:10.1137/050623802

    work page doi:10.1137/050623802 2006

  29. [29]

    Wei Wan et al.A scalable deep learning approach for solving high-dimensional dynamic optimal transport. 2022. arXiv:2205.07521 [cs.LG].url:https://arxiv.org/abs/2205.07521

    Pith/arXiv arXiv 2022

  30. [30]

    Sharp asymptotic and finite-sample rates of convergence of empirical measures in Wasserstein distance

    Jonathan Weed and Francis Bach. “Sharp asymptotic and finite-sample rates of convergence of empirical measures in Wasserstein distance”. In:Bernoulli25.4A (2019), pp. 2620–2648. doi:10.3150/18-BEJ1065

    doi:10.3150/18-bej1065 2019

  31. [31]

    Chen Xu, Xiuyuan Cheng, and Yao Xie.Computing high-dimensional optimal transport by flow neural networks. 2025. arXiv: 2305.11857 [stat.ML] .url: https://arxiv.org/abs/ 2305.11857. 38

    Pith/arXiv arXiv 2025