Pith. sign in

REVIEW 5 minor 1 cited by

On the equivalence between static and dynamic optimal transport governed by linear control systems

T0 review · 0 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Under a controllability rank condition, static and dynamic optimal transport coincide for linear control systems.

desk verdict Solid extension of static-dynamic optimal transport equivalence to non-autonomous linear control systems with p>1, worth serious refereeing. read the letter →

arxiv 2505.17570 v2 pith:LO3GIF65 submitted 2025-05-23 math.OC math.AP

Amit Einav , Yue Jiang , Alpár R. Mészáros This is my paper · ORCID
classification math.OCmath.AP MSC 49Q2235Q4949J1549N80
keywords optimaltransportlinearcontrolsystemsBenamou-BrenierformulageneralizedKalmanrankconditioncontinuityequationend-pointmapsuperpositionprincipleWassersteindistance
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

This paper proves that for linear control systems, the static and dynamic formulations of optimal transport agree exactly. Given dynamics $\gamma'=M(t)\gamma+N(t)\alpha$ and an energy cost $c_p(x,y)$ equal to the minimal $p$-th power of the control needed to steer $x$ to $y$, the minimum of $\int c_p\,d\pi$ over transport plans equals the minimum of $\int_0^T\int|u(t,x)|^p\,d\rho_t(x)\,dt$ over measure flows satisfying the generalized continuity equation. The result holds for every $p>1$ under a generalized Kalman rank condition, and it includes existence of minimizers and an explicit way to turn a minimizer of one problem into a minimizer of the other.

What carries the argument

The argument is carried by the end-point map $E^x_{s,t}(\alpha)=\Phi(s,t)x+\int_s^t \Phi(\tau,t)N(\tau)\alpha(\tau)\,d\tau$, where $\Phi$ is the state-transition flow of the homogeneous system $\gamma'=M\gamma$. Under the generalized Kalman rank condition this map is surjective, so $c_p(x,y)$ is finite and the optimal control $\alpha^*_p(\cdot;x,y)$ is unique. The central mechanism linking the two transport problems is the bijection $E_{0,T}: A_p(\mu,\nu)\to\Pi(\mu,\nu)$ between probability measures on optimal trajectories and transference plans: pushing a static plan through the inverse bijection and disintegrating the resulting path measure produces a solution of the generalized continuity equation, and this construction is used in both directions to prove equality and convert minimizers.

What would settle it

Take the constant double-integrator system $d=2$, $n=1$, $M=\begin{pmatrix}0&1\\0&0\end{pmatrix}$, $N=(0,1)^\top$, $T=1$, $p=2$, and transport $\delta_0$ to $\delta_{(1,1)}$. The optimal control in the static problem is $\alpha(t)=4-6t$ with $c_2(0,(1,1))=4$, so the theorem predicts $D_2=4$ with an explicit dynamic minimizer. A direct numerical discretization of the generalized continuity equation that yields any value strictly below $4$ would refute Theorem 1.2.

Watch

Extended reading notes

Core claim

The paper establishes that for any $\mu,\nu\in P_p(\mathbb{R}^d)$ with $p>1$ and coefficient matrices $M,N$ satisfying Assumption 2.1, the static cost $C_p(\mu,\nu)=\min_{\pi\in\Pi(\mu,\nu)}\int c_p\,d\pi$ equals the dynamic cost $D_p(\mu,\nu)=\min_{(\rho,u)\in cADM(\mu,\nu)}\int_0^T\int|u|^p\,d\rho_t\,dt$. Here $c_p(x,y)$ is the minimal control energy to steer $x$ to $y$ through the linear control system, and $cADM(\mu,\nu)$ consists of measure flows solving $\partial_t\rho_t+\operatorname{div}(\rho_t(M(t)x+N(t)u_t(x)))=0$. Both minima are attained, and the proof gives a constructive bijection between optimal transference plans and optimal dynamic flows.

Load-bearing premise

The result stands on the generalized Kalman rank condition $\operatorname{rank}(R)=d$ at $T^-$; without it the end-point map may fail to be surjective, the static cost $c_p$ can be infinite for some pairs, and the equivalence between the static and dynamic problems becomes vacuous.

Editorial extensions

If this is right

  • Both variational problems attain their minima: for every $\mu,\nu\in P_p(\mathbb{R}^d)$ with $p>1$ and coefficients satisfying Assumption 2.1, there exist an optimal transference plan $\pi^*$ and an optimal dynamic pair $(\rho^*,u^*)$.
  • Any minimizer on one side can be converted into a minimizer on the other: push $\pi^*$ through the inverse of $E_{0,T}$ and disintegrate to get $(\rho^*,u^*)$, or superpose a dynamic minimizer to recover a static plan.
  • The cost $c_p$ satisfies two-sided quantitative bounds $K_1|y-\Phi(0,T)x|^p \le c_p(x,y) \le K_2|y-\Phi(0,T)x|^p$, making $c_p^{1/p}$ globally Lipschitz and $c_p$ continuous.
  • The optimal control $\alpha^*_p(t;x,y)$ is unique and continuous in all variables, with the Lagrange-multiplier representation $j_p(\alpha^*_p)=\frac1p N(t)^\top\Phi(t,T)^\top\xi_p(x,y)$.
  • For $p=2$ the Lagrange multiplier becomes explicit, yielding $\alpha^*_2(t;x,y)=N(t)^\top\Phi(t,T)^\top\mathcal{M}^{-1}(y-\Phi(0,T)x)$ and explicit energy bounds.

Reading between the lines

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

  • Beyond the paper, the same bijection-and-disintegration recipe should apply to any control-affine dynamics whose end-point map is surjective and whose cost has a unique continuous optimal control; the linear structure is used heavily for those two properties, so nonlinear extensions will likely need a regularity or convexity substitute.
  • Beyond the paper, the norm-induced metric $d_p(x,y)=c_p(x,\Phi(0,T)y)^{1/p}$ suggests a family of control-kinetic Wasserstein distances, and one could test whether optimal dynamic flows are constant-speed geodesics for $d_p$, as in the classical $W_p$ case.
  • Beyond the paper, the explicit minimizer-conversion recipe has a numerical corollary: solve the static problem with cost $c_p$ and then lift the optimal plan through $E_{0,T}$ to obtain a dynamic control field, potentially avoiding a direct discretization of the continuity equation.
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

0 major / 5 minor

Summary. The paper establishes an equivalence between the static Kantorovich problem with cost c_p(x,y), defined as the minimal L^p control cost driving the linear non-autonomous system γ'=M(t)γ+N(t)α from x to y, and the dynamic problem in which a measure flow ρ_t is transported by the controlled vector field M(t)x+N(t)u(t,x). Under a generalized Kalman rank condition (Assumption 2.1), the authors prove that the two infima coincide and that minimizers exist in both formulations (Theorem 1.2). The proof is constructive: it first analyzes the end-point map and the optimal control α_p^*(·;x,y), establishes Lipschitz/continuity properties of the cost, constructs path measures concentrated on optimal trajectories via a superposition principle, and then proves the two inequalities D_p≤C_p and D_p≥C_p using a bijection between path measures and transport plans.

Significance. If the equivalence holds, this is a substantial extension of the Benamou–Brenier formula to linear control systems with time-dependent coefficients and arbitrary p>1, generalizing the quadratic p=2 result of [CGP17] and complementing the nonlinear control-affine results of [ELLO23, Ela25]. The paper's distinctive contributions are the constructive treatment of the optimal control α_p^*, the continuity and quantitative estimates for c_p, and the explicit bijection between measures on optimal trajectories and transport plans (Lemma 4.1). The proof is detailed and essentially self-contained, with functional-analytic arguments that are checkable from the text; the main fragility is explicitly isolated in Assumption 2.1(iii), the generalized Kalman rank condition, whose failure removes the theorem's scope rather than invalidating the argument. The paper also provides a clear recipe for converting minimizers between the static and dynamic problems, which is a useful and concrete byproduct.

minor comments (5)
  1. [§3, Eqs. (3.5) and (3.10)] The constant C_p used in the bounds for c_p^{1/p} and c_p is introduced in (2.24) only for the explicit p=2 case in Remark 2.6, but it is subsequently invoked for general p in the proof of Theorem 3.2 and in the moment estimates of Theorem 1.1. The argument is unaffected because Corollary 2.9(ii) gives c_p(x,y)≤K_2|y-Φ(0,T)x|^p for all p>1, so the same inequalities hold with K_2^{1/p} and K_2 respectively; the references to (2.24) should be replaced accordingly.
  2. [§3, Lemma 3.1 proof] The displayed formula for γ(t) in the proof of Lemma 3.1 uses the integrand Φ(τ,T)N(τ)α_p^*(τ;γ(0),γ(T)) integrated from 0 to T, but the correct variation-of-constants formula is γ(t)=Φ(0,t)γ(0)+∫_0^t Φ(τ,t)N(τ)α_p^*(τ;γ(0),γ(T))dτ. The convergence argument works with the corrected formula, so this is a typographical slip rather than a substantive gap.
  3. [§1, last paragraph of the introduction] The sentence 'In what follows we describe the guiding ideas in the proofs Theorems 1.1, 1.3 and 1.3' should read 'Theorems 1.1, 1.2 and 1.3'.
  4. [§1, definition of cADM around (1.7)] In the displayed definition of cADM(μ,ν) the integrability conditions are written with ∫_0^1, while the time horizon throughout the paper is T; these integrals should be over [0,T] to be consistent with the rest of the manuscript.
  5. [§2, Lemma 2.8, triangle inequality step] In the proof of the triangle inequality, 'Taking the infimum over the right hand side gives us...' should specify that the infimum is taken over both admissible pairs (γ_1,α_1) and (γ_2,α_2); with that clarification the step is correct.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the static–dynamic equivalence is established by two independent inequalities and external (non-self) theorems; the only self-citations are introductory and non-load-bearing.

full rationale

The central result Theorem 1.2 is proved by two independent inequalities, neither of which imports the conclusion. Lemma 4.2 (Section 4) shows D_p <= C_p: starting from a static minimizer pi* (which exists by continuity of c_p, Corollary 2.9, and compactness of Pi(mu,nu)), the bijection E_{0,T} (Lemma 4.1) yields eta* in A_p(mu,nu), and the construction of Theorem 1.1/Remark 3.2 gives (rho*,u*) in cADM(mu,nu) with dynamic cost bounded above by the integral of c_p against pi* (inequality (3.19)); this is an upper bound, not an identity. Lemma 4.4 gives the reverse inequality D_p >= C_p using the external, standard AGS08 superposition principle: any (rho,u) in cADM(mu,nu) lifts to a path measure eta whose trajectories satisfy the control ODE with control u(t,gamma(t)), and then by the very definition of c_p as an infimum over all admissible (gamma,alpha) in (1.5), the pathwise estimate integral|u(t,gamma(t))|^p dt >= c_p(gamma(0),gamma(T)) holds; pushing forward by (e_0,e_T) produces a plan in Pi(mu,nu) with static cost at most the dynamic cost. Neither direction assumes the equality it proves: c_p is defined by the ODE control problem (1.5), D_p by the continuity equation (1.7), and the chain C_p = D_p closes only through the conjunction of the two inequalities. The load-bearing external inputs are [Son98] (controllability, re-proved under weaker smoothness in Theorem 2.3), [Rif14] (Lagrange multiplier theorem, Theorem 2.10), and [AGS08] (disintegration and superposition principles, Theorems 3.3 and 4.3); none of these is authored by the present authors. The only self-citations ([GPM22] and [AGPM25], both involving co-author Meszaros) appear in the introductory literature survey and play no role in any proof step. Minor technical slips — using the p=2 bound (2.24) in (3.10) and in the proof of Lemma 3.1 where the general-p bound of Corollary 2.9(i) is needed — are repairable cross-reference errors, not circularity, and they do not affect the validity of the inequalities. Finding: no significant circularity; the derivation is self-contained apart from standard external theorems.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The central claim rests on standard theorems (Lagrange multipliers, superposition principle, Prokhorov, narrow compactness) and on Assumption 2.1, the generalized Kalman rank condition that guarantees controllability. There are no fitted parameters or invented quantities.

assumptions (3)
  • standard math Superposition principle for continuity equations (Theorem 4.3, [AGS08, Theorem 8.2.1])
    Used in Lemma 4.4 to represent any solution of the generalized continuity equation as a superposition of paths solving the ODE with the given control.
  • standard math Lagrange multipliers rule for constrained minimization in Banach spaces (Theorem 2.10, [Rif14, Appendix B])
    Used in Theorem 2.14 to characterize the optimal control in the definition of c_p via a multiplier xi_p(x,y).
  • domain assumption Assumption 2.1 (generalized Kalman rank condition ensuring controllability)
    Imposed on M,N to guarantee that the endpoint map is surjective, so that c_p is finite and the transport problems are well-posed. This is a standing assumption on the control system, not derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the equivalence between static and dynamic optimal transport governed by linear control systems." pith.science (2026). https://pith.science/paper/LO3GIF65

@misc{pith2026250517570,
  author       = {Pith},
  title        = {Pith review of: On the equivalence between static and dynamic optimal transport governed by linear control systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LO3GIF65}},
  note         = {Machine review of arXiv:2505.17570}
}
abstract

In this paper we revisit a class of optimal transport problems associated to non-autonomous linear control systems. Building on properties of the cost functions on $\mathbb{R}^{d}\times\mathbb{R}^{d}$ derived from suitable variational problems, we show the equivalence between the static and dynamic versions of the corresponding transport problems. Our analysis is constructive in nature and relies on functional analytic properties of the end-point map and the fine properties of the optimal control functions. These lead to some new quantitative estimates which play a crucial role in our investigation.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. From Kinetic Theory to AI: a Rediscovery of High-Dimensional Divergences and Their Properties

    math-ph 2025-07 conditional novelty 5.0 of 10

    The authors extend Energy Distances to negative and higher orders, connect them to Fourier-based metrics, and construct scale-invariant divergence measures for high-dimensional machine learning.

Reference graph

Works this paper leans on

39 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ambrose, M

    D.M. Ambrose, M. Griffin-Pickering, and A.R. M\'esz\'aros. Kinetic-type mean field games with non-separable local hamiltonians. J. Lond. Math. Soc. (2) , to appear, 2025

    work page 2025

  2. [2]

    urich. Birkh\

    L. Ambrosio, N. Gigli, and G. Savar\'e. Gradient flows in metric spaces and in the space of probability measures . Lectures in Mathematics ETH Z\"urich. Birkh\"auser Verlag, Basel, second edition, 2008

    work page 2008

  3. [3]

    Agrachev and P

    A. Agrachev and P. Lee. Optimal transportation under nonholonomic constraints. Trans. Amer. Math. Soc. , 361(11):6019--6047, 2009

    work page 2009

  4. [4]

    Ambrosio and S

    L. Ambrosio and S. Rigot. Optimal mass transportation in the H eisenberg group. J. Funct. Anal. , 208(2):261--301, 2004

    work page 2004

  5. [5]

    Benamou and Y

    J.-D. Benamou and Y. Brenier. A computational fluid mechanics solution to the M onge- K antorovich mass transfer problem. Numer. Math. , 84(3):375--393, 2000

    work page 2000

  6. [6]

    Bernard and B

    P. Bernard and B. Buffoni. The M onge problem for supercritical M a\ n\'e potentials on compact manifolds. Adv. Math. , 207(2):691--706, 2006

    work page 2006

  7. [7]

    Bernard and B

    P. Bernard and B. Buffoni. Optimal mass transportation and M ather theory. J. Eur. Math. Soc. (JEMS) , 9(1):85--121, 2007

    work page 2007

  8. [8]

    P. Bernard. Young measures, superposition and transport. Indiana Univ. Math. J. , 57(1):247--275, 2008

    work page 2008

Show all 39 references
  1. [9]

    Bonnet and H

    B. Bonnet and H. Frankowska. Necessary optimality conditions for optimal control problems in W asserstein spaces. Appl. Math. Optim. , 84:S1281--S1330, 2021

    2021

  2. [10]

    Bonnet and H

    B. Bonnet and H. Frankowska. Semiconcavity and sensitivity analysis in mean-field optimal control and applications. J. Math. Pures Appl. (9) , 157:282--345, 2022

    2022

  3. [11]

    Brigati, J

    G. Brigati, J. Maas, and F. Quattrocchi. Kinetic optimal transport ( OTIKIN ) -- P art 1: Second-order discrepancies between probability measures. arXiv:2502.15665 , 2025

    2025 arXiv

  4. [12]

    Bivas and M

    M. Bivas and M. Quincampoix. Optimal control for the evolution of deterministic multi-agent systems. J. Differential Equations , 269(3):2228--2263, 2020

    2020

  5. [13]

    Bonnet and F

    B. Bonnet and F. Rossi. Intrinsic L ipschitz regularity of mean-field optimal controls. SIAM J. Control Optim. , 59(3):2011--2046, 2021

    2011

  6. [14]

    Carmona and F

    R. Carmona and F. Delarue. Probabilistic theory of mean field games with applications. I , volume 83 of Probability Theory and Stochastic Modelling . Springer, Cham, 2018. Mean field FBSDEs, control, and games

    2018

  7. [15]

    Craig, K

    K. Craig, K. Elamvazhuthi, and H. Lee. A blob method for mean field control with terminal constraints. ESAIM Control Optim. Calc. Var. , 31:Paper No. 20, 46, 2025

    2025

  8. [16]

    Chen, T.T

    Y. Chen, T.T. Georgiou, and M. Pavon. Optimal transport over a linear dynamical system. IEEE Trans. Automat. Control , 62(5):2137--2152, 2017

    2017

  9. [17]

    Cavagnari, S

    G. Cavagnari, S. Lisini, C. Orrieri, and G. Savar\'e. Lagrangian, E ulerian and K antorovich formulations of multi-agent optimal control problems: equivalence and gamma-convergence. J. Differential Equations , 322:268--364, 2022

    2022

  10. [18]

    Cavagnari, A

    G. Cavagnari, A. Marigonda, K.T. Nguyen, and F.S. Priuli. Generalized control systems in the space of probability measures. Set-Valued Var. Anal. , 26(3):663--691, 2018

    2018

  11. [19]

    De Pascale, M

    L. De Pascale, M. S. Gelli, and L. Granieri. Minimal measures, one-dimensional currents and the M onge- K antorovich problem. Calc. Var. Partial Differential Equations , 27(1):1--23, 2006

    2006

  12. [20]

    Elamvazhuthi and M

    K. Elamvazhuthi and M. Jacobs. Optimal transport of linear systems over equilibrium measures. Automatica J. IFAC , 175:Paper No. 112222, 7, 2025

    2025

  13. [21]

    Elamvazhuthi

    K. Elamvazhuthi. B enamou-- B renier formulation of optimal transport for nonlinear control systems on R ^ d . arXiv:2407.16088v4 , 2025

    2025 arXiv

  14. [22]

    Elamvazhuthi, S

    K. Elamvazhuthi, S. Liu, W. Li, and S. Osher. Dynamical optimal transport of nonlinear control-affine systems. J. Comput. Dyn. , 10(4):425--449, 2023

    2023

  15. [23]

    Figalli and N

    A. Figalli and N. Juillet. Absolute continuity of W asserstein geodesics in the H eisenberg group. J. Funct. Anal. , 255(1):133--141, 2008

    2008

  16. [24]

    Fornasier, S

    M. Fornasier, S. Lisini, C. Orrieri, and G. Savar\'e. Mean-field optimal control as gamma-limit of finite agent controls. European J. Appl. Math. , 30(6):1153--1186, 2019

    2019

  17. [25]

    Figalli and L

    A. Figalli and L. Rifford. Mass transportation on sub- R iemannian manifolds. Geom. Funct. Anal. , 20(1):124--159, 2010

    2010

  18. [26]

    Griffin-Pickering and A.R

    M. Griffin-Pickering and A.R. M\'esz\'aros. A variational approach to first order kinetic mean field games with local couplings. Comm. Partial Differential Equations , 47(10):1945--2022, 2022

    1945

  19. [27]

    Hindawi, J.-B

    A. Hindawi, J.-B. Pomet, and L. Rifford. Mass transportation with LQ cost functions. Acta Appl. Math. , 113(2):215--229, 2011

    2011

  20. [28]

    Iacobelli

    M. Iacobelli. A new perspective on W asserstein distances for kinetic problems. Arch. Ration. Mech. Anal. , 244(1):27--50, 2022

    2022

  21. [29]

    Iacobelli and J

    M. Iacobelli and J. Junn\'e. Stability estimates for the V lasov- P oisson system in p -kinetic W asserstein distances. Bull. Lond. Math. Soc. , 56(7):2250--2267, 2024

    2024

  22. [30]

    C. Jimenez. Dynamic formulation of optimal transport problems. J. Convex Anal. , 15(3):593--622, 2008

    2008

  23. [31]

    Jimenez, A

    C. Jimenez, A. Marigonda, and M. Quincampoix. Optimal control of multiagent systems in the W asserstein space. Calc. Var. Partial Differential Equations , 59(2):Paper No. 58, 45, 2020

    2020

  24. [32]

    A. Klenke. Probability theory---a comprehensive course . Universitext. Springer, Cham, third edition, [2020] 2020

    2020

  25. [33]

    S. Park. A variational perspective on the dissipative H amiltonian structure of the V lasov-- F okker-- P lanck equation. arXiv:2406.13682v2 , 2025

    2025 arXiv

  26. [34]

    Pratelli

    A. Pratelli. Equivalence between some definitions for the optimal mass transport problem and for the transport density on manifolds. Ann. Mat. Pura Appl. (4) , 184(2):215--238, 2005

    2005

  27. [35]

    L. Rifford. Sub- R iemannian geometry and optimal transport . SpringerBriefs in Mathematics. Springer, Cham, 2014

    2014

  28. [36]

    Santambrogio

    F. Santambrogio. Optimal transport for applied mathematicians , volume 87 of Progress in Nonlinear Differential Equations and their Applications . Birkh\"auser/Springer, Cham, 2015. Calculus of variations, PDEs, and modeling

    2015

  29. [37]

    E.D. Sontag. Mathematical control theory , volume 6 of Texts in Applied Mathematics . Springer-Verlag, New York, second edition, 1998. Deterministic finite-dimensional systems

    1998

  30. [38]

    C. Villani. Topics in optimal transportation , volume 58 of Graduate Studies in Mathematics . American Mathematical Society, Providence, RI, 2003

    2003

  31. [39]

    C. Villani. Optimal transport , volume 338 of Grundlehren der mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences] . Springer-Verlag, Berlin, 2009. Old and new

    2009

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.