A PDE-constrained Optimization Approach to Optimal Trajectory Planning under Uncertainty via Reflected Schr\"odinger Bridges

Dante Kalise; Wenxin Liu

arxiv: 2511.14355 · v2 · pith:FXWRM5VKnew · submitted 2025-11-18 · 🧮 math.OC

A PDE-constrained Optimization Approach to Optimal Trajectory Planning under Uncertainty via Reflected Schr\"odinger Bridges

Dante Kalise , Wenxin Liu This is my paper

Pith reviewed 2026-05-21 19:06 UTC · model grok-4.3

classification 🧮 math.OC

keywords Schrödinger bridgePDE-constrained optimizationtrajectory planning under uncertaintyreflecting boundary conditionsHopf-Cole transformationfinite element discretizationFokker-Planck equationstochastic differential equations

0 comments

The pith

Reflected Schrödinger bridges for uncertain trajectory planning reduce to advection-diffusion equations solved by finite elements.

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

The paper proposes solving optimal trajectory planning under uncertainty using a reflected Schrödinger bridge formulation. This is interpreted as the mean-field limit of an energy-optimal stochastic process with reflecting boundaries. The resulting nonlinear system is transformed using the Hopf-Cole transformation into forward and backward advection-diffusion equations. These equations are discretized with finite elements, where the weak form naturally enforces the reflecting conditions. This avoids the collision detection issues of particle methods and is demonstrated on three-dimensional maze problems with good mass conservation and convergence.

Core claim

The authors establish that the Schrödinger bridge problem with reflecting boundary conditions, arising as the mean-field limit of an energy-optimal SDE, leads to a coupled nonlinear Fokker-Planck and Hamilton-Jacobi-Bellman system. Application of the Hopf-Cole transformation recasts this as a pair of forward-backward advection-diffusion equations subject to appropriate boundary and initial-terminal conditions. The system is then solved numerically using a standard finite element discretization that incorporates the Neumann boundary conditions for reflections in its weak formulation.

What carries the argument

Hopf-Cole transformation of the mean-field Fokker-Planck and Hamilton-Jacobi-Bellman system into advection-diffusion equations with weak enforcement of reflecting boundaries.

Load-bearing premise

The Schrödinger bridge problem accurately captures the mean-field limit of the energy-optimal evolution of a particle under a stochastic differential equation with nonlinear drift and reflecting boundary conditions.

What would settle it

A direct comparison where the optimal controls from the PDE solution, when inserted into reflected SDE simulations, do not reproduce the prescribed terminal density or fail to conserve probability mass.

Figures

Figures reproduced from arXiv: 2511.14355 by Dante Kalise, Wenxin Liu.

**Figure 2.** Figure 2: provides comprehensive 3D temporal evolution at 4 time points, showing the density ρ successfully transfers from entrance to exit while maintaining boundary confinement throughout. The optimal velocity field u ∗ = ε ∇lnϕ ϕ ( [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: The optimal velocity field u ∗ (t,x) along the spiral exhibits a strong vertical component aligned with the ascent. B. Transport with prior drift (ν ̸= 0) We introduce a prior drift field ν satisfying ∇· v = 0 in Ω, v · n = 0 on ∂Ω, by projecting a physically meaningful initial velocity field V onto the divergence-free subspace via a Helmholtz–Hodge decomposition. Specifically, we solve the Poisson equatio… view at source ↗

**Figure 5.** Figure 5: RSBP with prior drift: reflected SDE validation with [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

**Figure 7.** Figure 7: Mass preservation in the RSBP with/without prior. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗

read the original abstract

A computational PDE-constrained optimization approach is proposed for optimal trajectory planning under uncertainty by means of an associated Schroedinger Bridge Problem (SBP). The proposed SBP formulation is interpreted as the mean-field limit associated with the energy-optimal evolution of a particle governed by a stochastic differential equation (SDE) with nonlinear drift and reflecting boundary conditions, constrained to prescribed initial and terminal densities. The resulting mean-field system consists of a nonlinear Fokker-Planck equation coupled with a Hamilton-Jacobi-Bellman equation, subject to two-point boundary conditions in time and Neumann boundary conditions in space. Through the Hopf-Cole transformation, this nonlinear system is recast as a pair of forward-backward advection-diffusion equations, which are amenable to efficient numerical solution via a standard finite element discretization. The weak formulation naturally enforces reflecting boundary conditions without requiring explicit particle-boundary collision detection, thus circumventing the computational difficulties inherent to particle-based methods in complex geometries. Numerical experiments on challenging 3D maze configurations demonstrate fast convergence, mass conservation, and validate the optimal controls computed through reflected SDE simulations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a finite-element route to reflected Schrödinger bridges for uncertain trajectory planning, but the Hopf-Cole step for boundary conditions is the weakest link.

read the letter

The main point is a PDE-constrained method that recasts reflected Schrödinger bridge problems as forward-backward advection-diffusion equations solved by standard finite elements. The setup starts from the mean-field limit of an SDE with nonlinear drift and reflecting boundaries, then applies the Hopf-Cole transform to linearize the coupled Fokker-Planck and Hamilton-Jacobi-Bellman system. Numerical runs on 3D mazes show mass conservation and convergence, which is useful for avoiding explicit particle-boundary handling in complex domains. That practical angle is the clearest contribution, and the authors correctly identify the computational bottleneck that particle methods face in robotics-style settings. The work sits on established SBP and Fokker-Planck theory, so the novelty is mainly in the reflected-boundary extension plus the discretization choice rather than a wholly new theory. The experiments provide some evidence that the controls remain consistent with reflected SDE simulations, which is better than nothing. The soft spot is exactly the one flagged in the stress test. The original reflecting condition is a no-flux Neumann condition on density that mixes the drift and gradient terms. After the nonlinear Hopf-Cole change, that condition does not map to a simple homogeneous Neumann on the new variables, and the paper does not show the boundary integrals that would be needed to recover it in the weak form. Without an explicit verification or error analysis, it is not obvious that reflection is enforced correctly. The abstract also omits convergence rates or direct comparisons to other solvers, so the accuracy claim rests on the reported maze runs alone. This is aimed at researchers in stochastic optimal control and numerical methods for mean-field problems who need tools for bounded domains. A reader working on path planning under uncertainty would get a workable computational template from the experiments and the transformation outline. The paper is coherent enough on its own terms to deserve a serious referee, even if the boundary-condition step will need tightening.

Referee Report

2 major / 2 minor

Summary. The paper proposes a PDE-constrained optimization framework for optimal trajectory planning under uncertainty via reflected Schrödinger Bridge Problems (SBP). The SBP is interpreted as the mean-field limit of an energy-optimal SDE with nonlinear drift and reflecting boundary conditions, subject to prescribed initial and terminal densities. This yields a coupled nonlinear Fokker-Planck and Hamilton-Jacobi-Bellman system with two-point temporal boundary conditions and Neumann spatial boundary conditions. The Hopf-Cole transformation recasts the system as forward-backward advection-diffusion equations, which are discretized using standard finite elements; the weak form is asserted to naturally enforce the reflecting conditions without explicit collision handling. Numerical experiments on 3D maze geometries report fast convergence, mass conservation, and validation against reflected SDE simulations.

Significance. If the central claims hold, the work offers a potentially efficient PDE-based alternative to particle methods for stochastic trajectory planning in complex domains with uncertainty. It integrates mean-field limits, optimal control, and finite-element discretization in a coherent way. The explicit mean-field SDE interpretation and the avoidance of particle-boundary collision detection in the weak form would be notable strengths if rigorously justified. The numerical demonstration on 3D mazes provides practical evidence of feasibility, though the absence of convergence rates or baseline comparisons limits the strength of the validation.

major comments (2)

[Abstract / transformation section] Abstract and the transformation step: the claim that the weak formulation of the forward-backward advection-diffusion equations 'naturally enforces' reflecting boundary conditions must be verified explicitly. Because the Hopf-Cole transformation is nonlinear (typically of the form involving exp(-V/(2σ²)) and a related density variable), the original no-flux Neumann condition on the density (which couples the nonlinear drift and gradient terms) does not in general map to a homogeneous Neumann condition on the transformed variables. The boundary integrals arising in the weak form therefore require separate justification to confirm they reproduce the reflecting dynamics without additional terms or modifications.
[Numerical experiments] Numerical validation section: while mass conservation and convergence are reported for the 3D maze experiments, the manuscript lacks quantitative error bounds, convergence rates with respect to mesh size or time step, or comparisons against established particle-based reflected SDE solvers or other SBP methods. These omissions make it difficult to assess the accuracy and efficiency claims for the finite-element approach.

minor comments (2)

[Formulation] Notation for the transformed variables (e.g., the precise definitions of the Hopf-Cole pair) should be introduced with explicit equations early in the formulation section to improve readability.
[Numerical method] The manuscript would benefit from a brief discussion of how the two-point boundary conditions in time are enforced numerically after the transformation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our manuscript. We address each major comment below, indicating the revisions we intend to make to strengthen the presentation.

read point-by-point responses

Referee: Abstract and the transformation step: the claim that the weak formulation of the forward-backward advection-diffusion equations 'naturally enforces' reflecting boundary conditions must be verified explicitly. Because the Hopf-Cole transformation is nonlinear (typically of the form involving exp(-V/(2σ²)) and a related density variable), the original no-flux Neumann condition on the density (which couples the nonlinear drift and gradient terms) does not in general map to a homogeneous Neumann condition on the transformed variables. The boundary integrals arising in the weak form therefore require separate justification to confirm they reproduce the reflecting dynamics without additional terms or modifications.

Authors: We acknowledge that the referee correctly identifies a point requiring explicit verification. The Hopf-Cole transformation is indeed nonlinear, and while the manuscript derives the weak form directly from the transformed forward-backward advection-diffusion equations (with the Neumann conditions inherited from the original system), a dedicated step-by-step check of the boundary integrals under this transformation was not included. In the revised manuscript we will add a short subsection after the transformation derivation that explicitly computes the boundary terms for our specific form of the Hopf-Cole map and shows that they vanish, thereby confirming that the reflecting (no-flux) dynamics are preserved in the weak formulation without extra modifications or collision-handling terms. revision: yes
Referee: Numerical validation section: while mass conservation and convergence are reported for the 3D maze experiments, the manuscript lacks quantitative error bounds, convergence rates with respect to mesh size or time step, or comparisons against established particle-based reflected SDE solvers or other SBP methods. These omissions make it difficult to assess the accuracy and efficiency claims for the finite-element approach.

Authors: We agree that quantitative error analysis and baseline comparisons would strengthen the numerical section. The present experiments already demonstrate mass conservation to machine precision and qualitative agreement between the computed controls and independent reflected-SDE trajectories. In the revision we will augment the numerical results with (i) L² error norms computed against a reference solution obtained on a highly refined mesh, (ii) observed convergence rates under successive uniform mesh refinements for both the forward and backward equations, and (iii) direct runtime and accuracy comparisons against a standard particle-based reflected SDE solver on the same 3D maze geometries. These additions will allow readers to assess the accuracy and efficiency claims more rigorously while still highlighting the practical advantage of the PDE approach in avoiding explicit particle-boundary collision detection. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on standard transformations and numerical methods

full rationale

The paper's chain starts from the SBP mean-field limit of the reflected SDE, yields the nonlinear HJB-FP system with Neumann BCs, applies the Hopf-Cole transformation to obtain forward-backward advection-diffusion equations, and discretizes via standard weak-form FEM. Each step is a direct mathematical or numerical operation whose validity can be checked independently against the original SDE and boundary conditions; no step reduces the claimed result to a fitted parameter, self-referential definition, or load-bearing self-citation. The assertion that the weak form naturally enforces reflection follows from the transformed operators and is presented as a consequence rather than an input. The formulation is therefore self-contained against external benchmarks such as direct SDE simulation.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests primarily on standard PDE and stochastic control theory plus one key domain assumption linking the SBP to the mean-field SDE limit; no free parameters or invented entities are introduced in the abstract.

axioms (2)

domain assumption The Schrödinger bridge problem formulation corresponds to the mean-field limit of energy-optimal particle evolution under an SDE with nonlinear drift and reflecting boundary conditions.
Explicitly stated as the interpretation of the SBP in the abstract.
standard math The Hopf-Cole transformation recasts the coupled nonlinear Fokker-Planck/HJB system into forward-backward advection-diffusion equations while preserving the boundary conditions.
Invoked in the abstract to enable the finite-element discretization.

pith-pipeline@v0.9.0 · 5720 in / 1461 out tokens · 50864 ms · 2026-05-21T19:06:27.054642+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

[1]

Distributed optimal control for multi-agent trajectory optimization,

G. Foderaro, S. Ferrari, and T. A. Wettergren, “Distributed optimal control for multi-agent trajectory optimization,”Automatica, vol. 50, no. 1, pp. 149–154, 2014

work page 2014
[2]

Computation and control of unstable steady states for mean field multiagent systems,

S. Bicego, D. Kalise, and G. A. Pavliotis, “Computation and control of unstable steady states for mean field multiagent systems,”Proc. R. Soc. A, vol. 481, no. 2311, p. 20240476, 2025

work page 2025
[3]

Maze solving by chemotactic droplets,

I. Lagzi, S. Soh, P. J. Wesson, K. P. Browne, and B. A. Grzybowski, “Maze solving by chemotactic droplets,”J. Am. Chem. Soc., vol. 132, no. 4, pp. 1198–1199, 2010

work page 2010
[4]

Maze solving using temperature-induced marangoni flow,

P. Lovass, M. Branicki, R. T ´oth, A. Braun, K. Suzuno, D. Ueyama, and I. Lagzi, “Maze solving using temperature-induced marangoni flow,” RSC Adv., vol. 5, no. 60, pp. 48 563–48 568, 2015

work page 2015
[5]

Villani,Topics in Optimal Transportation

C. Villani,Topics in Optimal Transportation. American Mathematical Society, 2009, vol. 58

work page 2009
[6]

Sur la th ´eorie relativiste de l’ ´electron et l’interpr´etation de la m´ecanique quantique,

E. Schr ¨odinger, “Sur la th ´eorie relativiste de l’ ´electron et l’interpr´etation de la m´ecanique quantique,”Ann. Inst. Henri Poincar ´e, vol. 2, no. 4, pp. 269–310, 1932

work page 1932
[7]

Stochastic control liaisons: Richard Sinkhorn meets Gaspard Monge on a Schr ¨odinger bridge,

Y . Chen, T. T. Georgiou, and M. Pavon, “Stochastic control liaisons: Richard Sinkhorn meets Gaspard Monge on a Schr ¨odinger bridge,” SIAM Rev., vol. 63, no. 2, pp. 249–313, 2021

work page 2021
[8]

Gradient flow algorithms for density propagation in stochastic systems,

K. F. Caluya and A. Halder, “Gradient flow algorithms for density propagation in stochastic systems,”IEEE Trans. Automat. Control, vol. 65, no. 10, pp. 3991–4004, 2020

work page 2020
[9]

Wasserstein proximal algorithms for the Schr ¨odinger bridge problem: Density control with nonlinear drift,

——, “Wasserstein proximal algorithms for the Schr ¨odinger bridge problem: Density control with nonlinear drift,”IEEE Trans. Automat. Control, vol. 67, no. 3, pp. 1163–1178, 2022. (a) Convergence in∆x (b) Convergence in∆t Fig. 6: Convergence in the RSBP with prior drift. (a) Mass conservation (ν=0) (b) Mass conservation (ν̸=0) Fig. 7: Mass preservation in...

work page 2022
[10]

Reflected Schr ¨odinger bridge: Density control with path con- straints,

——, “Reflected Schr ¨odinger bridge: Density control with path con- straints,” inProc. 2021 American Control Conference (ACC). IEEE, 2021, pp. 1137–1142

work page 2021
[11]

A fictitious-play finite-difference method for linearly solvable mean field games,

Inoue, Daisuke, Ito, Yuji, Kashiwabara, Takahito, Saito, Norikazu, and Yoshida, Hiroaki, “A fictitious-play finite-difference method for linearly solvable mean field games,”ESAIM Math. Model. Numer . Anal., vol. 57, no. 4, pp. 1863–1892, 2023

work page 2023
[12]

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

J.-D. Benamou and Y . Brenier, “A computational fluid mechanics solution to the Monge-Kantorovich mass transfer problem,”Numer . Math., vol. 84, no. 3, pp. 375–393, 2000

work page 2000
[13]

Mean field control hierarchy,

G. Albi, Y .-P. Choi, M. Fornasier, and D. Kalise, “Mean field control hierarchy,”Appl. Math. Optim., vol. 76, no. 1, pp. 93–135, 2017

work page 2017
[14]

On the implementation of a primal- dual algorithm for second order time-dependent mean field games with local couplings,

Brice ˜no-Arias, L., Kalise, D., Kobeissi, Z., Lauri `ere, M., Mateos Gonz´alez, ´A., and Silva, F. J., “On the implementation of a primal- dual algorithm for second order time-dependent mean field games with local couplings,”ESAIM Proc. Surv., vol. 65, pp. 330–348, 2019

work page 2019
[15]

Entropic and displacement interpolation: A computational approach using the hilbert metric,

Y . Chen, T. Georgiou, and M. Pavon, “Entropic and displacement interpolation: A computational approach using the hilbert metric,” SIAM J. Appl. Math., vol. 76, no. 6, pp. 2375–2396, 2016

work page 2016
[16]

Quarteroni and A

A. Quarteroni and A. Valli,Unsteady Advection-Diffusion Problems. Berlin, Heidelberg: Springer Berlin Heidelberg, 1994, pp. 405–427

work page 1994

[1] [1]

Distributed optimal control for multi-agent trajectory optimization,

G. Foderaro, S. Ferrari, and T. A. Wettergren, “Distributed optimal control for multi-agent trajectory optimization,”Automatica, vol. 50, no. 1, pp. 149–154, 2014

work page 2014

[2] [2]

Computation and control of unstable steady states for mean field multiagent systems,

S. Bicego, D. Kalise, and G. A. Pavliotis, “Computation and control of unstable steady states for mean field multiagent systems,”Proc. R. Soc. A, vol. 481, no. 2311, p. 20240476, 2025

work page 2025

[3] [3]

Maze solving by chemotactic droplets,

I. Lagzi, S. Soh, P. J. Wesson, K. P. Browne, and B. A. Grzybowski, “Maze solving by chemotactic droplets,”J. Am. Chem. Soc., vol. 132, no. 4, pp. 1198–1199, 2010

work page 2010

[4] [4]

Maze solving using temperature-induced marangoni flow,

P. Lovass, M. Branicki, R. T ´oth, A. Braun, K. Suzuno, D. Ueyama, and I. Lagzi, “Maze solving using temperature-induced marangoni flow,” RSC Adv., vol. 5, no. 60, pp. 48 563–48 568, 2015

work page 2015

[5] [5]

Villani,Topics in Optimal Transportation

C. Villani,Topics in Optimal Transportation. American Mathematical Society, 2009, vol. 58

work page 2009

[6] [6]

Sur la th ´eorie relativiste de l’ ´electron et l’interpr´etation de la m´ecanique quantique,

E. Schr ¨odinger, “Sur la th ´eorie relativiste de l’ ´electron et l’interpr´etation de la m´ecanique quantique,”Ann. Inst. Henri Poincar ´e, vol. 2, no. 4, pp. 269–310, 1932

work page 1932

[7] [7]

Stochastic control liaisons: Richard Sinkhorn meets Gaspard Monge on a Schr ¨odinger bridge,

Y . Chen, T. T. Georgiou, and M. Pavon, “Stochastic control liaisons: Richard Sinkhorn meets Gaspard Monge on a Schr ¨odinger bridge,” SIAM Rev., vol. 63, no. 2, pp. 249–313, 2021

work page 2021

[8] [8]

Gradient flow algorithms for density propagation in stochastic systems,

K. F. Caluya and A. Halder, “Gradient flow algorithms for density propagation in stochastic systems,”IEEE Trans. Automat. Control, vol. 65, no. 10, pp. 3991–4004, 2020

work page 2020

[9] [9]

Wasserstein proximal algorithms for the Schr ¨odinger bridge problem: Density control with nonlinear drift,

——, “Wasserstein proximal algorithms for the Schr ¨odinger bridge problem: Density control with nonlinear drift,”IEEE Trans. Automat. Control, vol. 67, no. 3, pp. 1163–1178, 2022. (a) Convergence in∆x (b) Convergence in∆t Fig. 6: Convergence in the RSBP with prior drift. (a) Mass conservation (ν=0) (b) Mass conservation (ν̸=0) Fig. 7: Mass preservation in...

work page 2022

[10] [10]

Reflected Schr ¨odinger bridge: Density control with path con- straints,

——, “Reflected Schr ¨odinger bridge: Density control with path con- straints,” inProc. 2021 American Control Conference (ACC). IEEE, 2021, pp. 1137–1142

work page 2021

[11] [11]

A fictitious-play finite-difference method for linearly solvable mean field games,

Inoue, Daisuke, Ito, Yuji, Kashiwabara, Takahito, Saito, Norikazu, and Yoshida, Hiroaki, “A fictitious-play finite-difference method for linearly solvable mean field games,”ESAIM Math. Model. Numer . Anal., vol. 57, no. 4, pp. 1863–1892, 2023

work page 2023

[12] [12]

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

J.-D. Benamou and Y . Brenier, “A computational fluid mechanics solution to the Monge-Kantorovich mass transfer problem,”Numer . Math., vol. 84, no. 3, pp. 375–393, 2000

work page 2000

[13] [13]

Mean field control hierarchy,

G. Albi, Y .-P. Choi, M. Fornasier, and D. Kalise, “Mean field control hierarchy,”Appl. Math. Optim., vol. 76, no. 1, pp. 93–135, 2017

work page 2017

[14] [14]

On the implementation of a primal- dual algorithm for second order time-dependent mean field games with local couplings,

Brice ˜no-Arias, L., Kalise, D., Kobeissi, Z., Lauri `ere, M., Mateos Gonz´alez, ´A., and Silva, F. J., “On the implementation of a primal- dual algorithm for second order time-dependent mean field games with local couplings,”ESAIM Proc. Surv., vol. 65, pp. 330–348, 2019

work page 2019

[15] [15]

Entropic and displacement interpolation: A computational approach using the hilbert metric,

Y . Chen, T. Georgiou, and M. Pavon, “Entropic and displacement interpolation: A computational approach using the hilbert metric,” SIAM J. Appl. Math., vol. 76, no. 6, pp. 2375–2396, 2016

work page 2016

[16] [16]

Quarteroni and A

A. Quarteroni and A. Valli,Unsteady Advection-Diffusion Problems. Berlin, Heidelberg: Springer Berlin Heidelberg, 1994, pp. 405–427

work page 1994