Pith. sign in

REVIEW 3 major objections 6 minor 40 references

Flowing a dual co-state with each sample enforces nonlinear constraints in flow matching without projections or pseudoinverses.

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

2026-07-11 18:16 UTC pith:VYXE3WOR

load-bearing objection Clean dual-augmented flow-matching ODE with honest Lyapunov rates and real code; the nonlinear guarantee still rests on an unchecked small-variation Gram condition, and the experiments are still toys. the 3 major comments →

arxiv 2607.04513 v1 pith:VYXE3WOR submitted 2026-07-05 math.OC cs.LG

Constrained Flow Matching via Lagrangian Dual Flows

Vince Kurtz , Alexander Davydov This is my paper
classification math.OC cs.LG MSC 90C3065L0568T07
keywords flow matchingconstrained generationLagrangian dual dynamicsprimal-dual methodsinference-time constraintsequality and inequality constraintsgenerative modelingODE integration
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.

Flow matching generates samples by integrating a learned ODE, but robotics, planning, and physics often need those samples to obey hard nonlinear constraints known only at inference time. Existing fixes either handle only linear measurements or solve expensive projections and optimization subproblems at every denoising step. This paper introduces Lagrangian Dual Flows: the generative ODE is simply augmented with a dual co-state (a continuously integrated Lagrange multiplier) and, for inequalities, a soft-projected slack. Under standard regularity, constraint qualification, and a small-variation condition on the constraint Gram matrix, the residual is driven to zero at an explicit algebraic rate as generation ends. Sampling still uses ordinary off-the-shelf ODE solvers and only one extra vector-Jacobian product per step, opening a direct link between generative flow matching and classical primal-dual optimization dynamics.

Core claim

Lagrangian Dual Flows augment a pretrained flow-matching ODE with a dual co-state whose dynamics accumulate the constraint residual, scaled by (1-t)^{-p}. For p=2, under smoothness, constraint qualification, bounded trajectories, and a small-variation condition on the Gram matrix of the constraint Jacobian, both equality residuals and inequality violations (via a soft-projected slack) converge to zero at rate O((1-t)^α) as t o1−, without requiring pseudoinverses, hard projections, or optimization subproblems during integration.

What carries the argument

The Lagrangian dual flow: the joint ODE that evolves the sample x_t together with dual variable λ_t (and slack s_t for inequalities) via ẋ = v_θ(x,t) − J^⊤(λ + c g) and λ̇ = g(x)/(1-t)^p (or the soft-projected analog for inequalities). The dual acts as a continuously integrated Lagrange multiplier that drives residual to zero while a modest penalty keeps the dynamics well-behaved.

Load-bearing premise

The constraint Jacobian’s Gram matrix must stay close enough to a fixed positive-definite reference along the whole trajectory so the Lyapunov dissipation margin stays positive; this is automatic for affine constraints but is not guaranteed a priori for highly curved nonlinear ones.

What would settle it

Integrate the dual flow with p=2 and high-accuracy ODE control on a smooth nonlinear equality whose Gram matrix varies so much that the small-variation constant ρ exceeds 1 for every reference matrix; if the residual does not go to zero as t o1−, the rate claim fails for that class of constraints.

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

If this is right

  • Nonlinear equality and inequality constraints known only at inference can be enforced with an augmented ODE and a single vector-Jacobian product per step.
  • Affine constraints automatically satisfy the small-variation assumption, so the algebraic rate holds under weaker conditions.
  • Sampling cost stays comparable to unconstrained flow matching, avoiding per-step linear solves of pseudoinverse methods and nested ODE or optimization solves of projection methods.
  • Soft projection preserves slack nonnegativity and global Lipschitz continuity, so standard ODE integrators remain valid.
  • When specialized to linear constraints the dual correction recovers a structure related to pseudoinverse guidance while remaining applicable to nonlinear ones.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same dual-flow coupling could be attached to reverse SDEs in score-based generative models, not only deterministic flow matching.
  • Online robotics and physics settings where constraints change after training stand to gain most, because no retraining or architecture change is required.
  • Stiffness-versus-tightness tradeoffs suggest hybrid schedules that ramp dual gain or switch p near t=1 to keep step counts low.
  • Mature continuous-time primal-dual accelerations could be imported to reduce the number of denoising steps needed for a target residual.

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 / 6 minor

Summary. The paper proposes Lagrangian Dual Flows (LDF), which enforce inference-time equality and inequality constraints on a pretrained flow-matching model by augmenting the generative ODE with dual (and, for inequalities, soft-projected slack) variables. Sampling remains a single off-the-shelf ODE integration; no per-step pseudoinverses, projections, or optimization subproblems are required. Under smoothness, constraint qualification, bounded trajectories, and a small-variation condition on the constraint Gram matrix, Theorems 1 and 3 prove that residual violation decays as O((1−t)^α) as t→1− for dual exponent p=2. Algorithms, a Lyapunov analysis after the change of time τ=−log(1−t), and experiments on a nonlinear unit-circle constraint, a half-plane inequality, and MNIST inpainting are provided.

Significance. If the guarantees hold as stated, LDF supplies a simple, theoretically grounded alternative to projection- and pseudoinverse-based constrained generation that is especially attractive for nonlinear constraints arising in robotics, planning, and physics. Strengths include: (i) an explicit algebraic rate rather than a soft penalty residual; (ii) a globally Lipschitz soft-projection for inequalities that preserves nonnegativity without discontinuous projected dynamics (Lemma 2); (iii) careful ownership of the p≠2 regime and of the feasibility-only (not conditional-sampling) objective; and (iv) released JAX/Diffrax code. The bridge to classical primal-dual/saddle-point flows is a genuine conceptual contribution for the flow-matching literature.

major comments (3)
  1. Assumption (A4)/(B4) (small variation of Jg Jg⊤ or Jh Jh⊤ about a fixed A∗ with ρ small enough to keep the Lyapunov dissipation margin positive) is load-bearing for the algebraic rate on nonlinear constraints (Theorems 1 and 3, §4.3–4.4 and Appendices B–C). The paper correctly notes that (A4)/(B4) holds automatically for affine constraints, but the motivating applications and the unit-circle star experiment (§5.1, g(x)=∥x∥2−1) are nonlinear. No a-priori bound, post-hoc numerical check of sup_t ∥A(xt)−A∗∥, or diagnostic that ρ<1 (resp. ρ<κ0) is reported along those trajectories. Without such evidence the central nonlinear guarantee remains conditional on an unverified trajectory-dependent hypothesis. Either verify (A4)/(B4) on the nonlinear examples, restrict the rate claim more sharply to affine/near-affine constraints, or supply a structural condition on g that implies (A4).
  2. The contribution list and abstract claim “tighter constraint satisfaction than pseudoinverse and penalty-based baselines.” Table 1 does not uniformly support this: for the star with 10 steps, Dual Flows residual 3.3×10−2 exceeds Pseudoinverse 1.2×10−2; for MNIST with 10 steps, Dual Flows 3.1×10−1 exceeds Pseudoinverse 1.2×100 only after the latter is already large, while PCFM is orders of magnitude tighter in every row. The speed advantage over PCFM is clear and valuable; the residual comparison should be restated to match the table (e.g., competitive residual at substantially lower cost, or tighter than penalty-only as in Fig. 3).
  3. Assumption (A3)/(B3) (bounded trajectory of (x,λ) or (x,s,λ) on [0,1)) is used both to justify the comparison-lemma bounds and, for inequalities, to control the forcing term. For the dual variable this is not a priori; the equality proof obtains (1−t)∥λt∥→0 only after the Lyapunov argument, while the inequality proof must establish boundedness of η en route (Appendix B, Step 4). The manuscript should clarify that (A3) is partly a standing hypothesis on x (and s) rather than a fully closed a-priori bound, and note any practical safeguards (e.g., the optional ReLU clipping mentioned for s).
minor comments (6)
  1. §4.1, Eq. (7): the pseudoinverse form is written with Jg† Jg; a brief remark that this is the orthogonal projector onto the tangent space would help readers less familiar with constrained dynamics.
  2. Fig. 1 bottom row: generation paths are informative; adding a short note on integrator (midpoint, dt=0.01) in the caption would make the comparison self-contained.
  3. §5.1.2 / Fig. 2: the dependence of residual on c and p is clear; stating the default (c,p) used in Table 1 and Fig. 1 would improve reproducibility of the main tables.
  4. Appendix C.1, Proposition 6: the scalar p=1 counter-example is useful; a one-sentence pointer from the main text (§4.3, “for p∈[1,2)”) would help readers find it.
  5. Typos / polish: “Lagrage” → “Lagrange” (§4.1); “Schr¨odinger” encoding; “ast→1−” spacing in several theorem statements; arXiv id in the preprint header is 2607.04513 while the abstract block is fine.
  6. Related work: a short comparison to continuous-time projected/proximal gradient dynamics already cited [34–36] could further situate the soft-projection (9b).

Circularity Check

0 steps flagged

No circularity: rates follow from Lyapunov analysis of the stated dual ODEs under explicit assumptions, not from fitted targets or self-citation identities.

full rationale

The paper defines the Lagrangian dual flow ODEs (8a–8b) and soft-projected inequality dynamics (9a–9c), then proves residual decay via a standard exponential time change, a constructed Lyapunov function, and the comparison lemma under Assumptions (A1)–(A4)/(B1)–(B4). The claimed algebraic rates are properties of those dynamics, not quantities fitted to data or defined as the residual itself. Free parameters c and p are hyperparameters; they are not identified with the target residual. Classical primal-dual citations (Arrow–Hurwicz, Platt–Barr) and one related continuous-time optimization citation by a coauthor supply inspiration for the form of the dynamics, not a uniqueness theorem or ansatz that forces Theorems 1 and 3. Numerical comparisons are empirical benchmarks, not predictions forced by construction. The skeptic’s concern about unverified small-variation (A4)/(B4) on nonlinear trajectories is an assumption/correctness risk, not circularity. Derivation chain is self-contained.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 2 invented entities

The central feasibility claims rest on standard ODE/Lyapunov machinery plus domain assumptions on the pretrained velocity field and constraints, plus two free hyperparameters (penalty c and dual exponent p). No new physical entities are postulated; the dual and slack states are classical optimization constructs adapted to finite-horizon flow matching via the (1−t)^{-p} rescaling.

free parameters (2)
  • penalty weight c
    Hand-chosen scalar controlling the quadratic residual pull and soft-projection strength; experiments show residual and stiffness depend strongly on c (Figs. 2–3, 6).
  • dual time-rescaling exponent p
    Hand-chosen exponent in λ̇ = residual/(1−t)^p; theory focuses on p=2, with incomplete general guarantees for other p and numerical stiffness for p>2.
axioms (5)
  • domain assumption Smoothness of g/h and v_θ with uniform Jacobian and velocity bounds along trajectories (A1/B1).
    Standing regularity for chain-rule residual dynamics and forcing bounds in the Lyapunov analysis.
  • domain assumption Uniform full-rank constraint qualification: a_min I ⪯ J J^T ⪯ a_max I along the trajectory (A2/B2).
    Needed so dual corrections span the constraint normal space and Gram matrices stay well-conditioned.
  • domain assumption Bounded trajectories of (x,λ) or (x,s,λ) on [0,1) (A3/B3).
    Rules out finite-time blowup; for inequalities the paper notes an optional ReLU cap to enforce compact slack.
  • ad hoc to paper Small-variation of the constraint Gram matrix about a fixed A_* with dissipation margin (A4/B4).
    Extra trajectory condition enabling the Lyapunov comparison argument for nonlinear constraints; automatic only for affine constraints.
  • standard math Comparison lemma / standard Lyapunov ODE estimates (Khalil).
    Used to convert dV/dτ inequalities into explicit algebraic rates in τ-time.
invented entities (2)
  • Lagrangian Dual Flow generative sampler (dual-augmented flow-matching ODE with (1−t)^{-p} dual dynamics) independent evidence
    purpose: Couples classical primal-dual continuous-time dynamics to finite-horizon flow matching so feasibility is achieved at t→1 without nested solves.
    The dual variable itself is classical; the entity is the specific generative construction and time rescaling for [0,1] flow matching.
  • Soft-projected slack dynamics for inequality LDF independent evidence
    purpose: Keep slack nonnegative with globally Lipschitz right-hand side, avoiding discontinuous projected dynamical systems.
    Continuous-time analog of projected/proximal gradient steps; nonnegativity proved in Lemma 2.

pith-pipeline@v1.1.0-grok45 · 27727 in / 3383 out tokens · 33739 ms · 2026-07-11T18:16:39.011803+00:00 · methodology

0 comments
read the original abstract

Flow matching is a powerful tool for generative modeling, but emerging applications in robotics, planning, and physics require inference-time constraints on generated outputs. Such constraints are often complex and highly nonlinear. As a result, methods designed for linear constraints like image inpainting are rarely sufficient, and projection or optimization-based alternatives can be prohibitively expensive. In this paper, we introduce Lagrangian Dual Flows, a new family of constrained generation techniques based on Lagrangian dual dynamics. By simply flowing a dual co-state alongside generated samples, we can guarantee nonlinear constraint satisfaction without expensive optimization subproblems, pseudoinverses, or projection steps during the denoising process. The resulting constrained generation algorithms are simple, effective, and open new theoretical connections between flow matching and primal-dual methods in numerical optimization.

Figures

Figures reproduced from arXiv: 2607.04513 by Alexander Davydov, Vince Kurtz.

Figure 1
Figure 1. Figure 1: Generated star examples subject to a unit circle constraint. Compares our approach, [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Average constraint violation for the star [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison with a simple penalty method (e.g., λ = 0). Flowed Lagrangian dy￾namics allows us to further reduce constraint vi￾olations while limiting the stiffness of the ODE. The resulting data is shown in [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Samples from a pretrained “star” model, constrained to the right half of the plane using Al￾gorithm 2. Image inpainting, where only a certain subset of pixels are generated by a pre-trained model, is a special case of equality constraints g(x) = 0. In this case, g(x) is a linear constraint specify￾ing the value of fixed pixels. We train a ∼6M parameter UNet to generate MNIST digits, then apply an inpaintin… view at source ↗
Figure 5
Figure 5. Figure 5: A pre-trained model generates MNIST handwritten digits. We then apply pseudoinverse [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: For MNIST image inpainting, constraint tightness saturates around c = 20 with a fixed￾step midpoint integrator. Error-controlled integra￾tion (in this case adaptive-step Heun-Euler) [38] allows us to further tighten constraint satisfaction. Lagrangian dual flows provide a simple, com￾putationally efficient means to apply inference￾time constraints to pre-trained flow matching policies. By flowing a dual co… view at source ↗
Figure 7
Figure 7. Figure 7: The original PCFM algorithm (left, [7, Algorithm 1]) produces extremely noisy generation paths for our star examples, subject to a unit circle constraint. We compare with an improved baseline (right), where increasing the number of projection steps avoids overshoot and results in more reasonable generation paths. We wrap the dynamics of the sample x, Lagrange multipliers λ, and any slack variables s into a… 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

40 extracted references · 4 linked inside Pith

  1. [1]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  2. [2]

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models.Advances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  3. [3]

    Pyramidal flow matching for efficient video generative modeling

    Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. InInternational Conference on Learning Representations (ICLR), pages 23378– 23402, 2025

    2025

  4. [4]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.π 0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    FlowPolicy: Enabling fast and robust 3D flow-based policy via consistency flow matching for robot manipulation

    Qinglun Zhang, Zhen Liu, Haoqiang Fan, Guanghui Liu, Bing Zeng, and Shuaicheng Liu. FlowPolicy: Enabling fast and robust 3D flow-based policy via consistency flow matching for robot manipulation. InAAAI Conference on Artificial Intelligence, volume 39, pages 14754– 14762, 2025

    2025

  6. [6]

    Generative predictive control: Flow matching policies for dynamic and difficult-to-demonstrate tasks.arXiv preprint arXiv:2502.13406, 2025

    Vince Kurtz and Joel W Burdick. Generative predictive control: Flow matching policies for dynamic and difficult-to-demonstrate tasks.arXiv preprint arXiv:2502.13406, 2025

    arXiv 2025

  7. [7]

    Physics-constrained flow matching: Sampling generative models with hard con- straints

    Utkarsh, Pengfei Cai, Alan Edelman, Rafael Gomez-Bombarelli, and Christopher Vincent Rackauckas. Physics-constrained flow matching: Sampling generative models with hard con- straints. InAdvances in Neural Information Processing Systems (NeurIPS), 2025. 12

    2025

  8. [8]

    Physics vs distributions: Pareto optimal flow matching with physics constraints

    Giacomo Baldan, Qiang Liu, Alberto Guardone, and Nils Thuerey. Physics vs distributions: Pareto optimal flow matching with physics constraints. InInternational Conference on Learn- ing Representations (ICLR), 2026

    2026

  9. [9]

    Pseudoinverse-guided diffu- sion models for inverse problems

    Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffu- sion models for inverse problems. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  10. [10]

    Muckley, Ricky T

    Ashwini Pokle, Matthew J. Muckley, Ricky T. Q. Chen, and Brian Karrer. Training-free linear image inverses via flows.Transactions on Machine Learning Research (TMLR), 2024

    2024

  11. [11]

    HardFlow: Hard-constrained sampling for flow- matching models via trajectory optimization.arXiv preprint arXiv:2511.08425, 2025

    Zeyang Li, Kaveh Alim, and Navid Azizan. HardFlow: Hard-constrained sampling for flow- matching models via trajectory optimization.arXiv preprint arXiv:2511.08425, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    Arrow, Leonid Hurwicz, and Hirofumi Uzawa.Studies in Linear and Non-Linear Programming

    Kenneth J. Arrow, Leonid Hurwicz, and Hirofumi Uzawa.Studies in Linear and Non-Linear Programming. Stanford University Press, 1958

    1958

  13. [13]

    Constrained differential optimization

    John Platt and Alan Barr. Constrained differential optimization. InAdvances in Neural Infor- mation Processing Systems (NeurIPS), 1987

    1987

  14. [14]

    Improving diffusion models for inverse problems using manifold constraints

    Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. InAdvances in Neural Information Process- ing Systems (NeurIPS), 2022

    2022

  15. [15]

    Pyramidal denoising diffusion probabilistic models.arXiv preprint arXiv:2208.01864, 2022

    Dohoon Ryu and Jong Chul Ye. Pyramidal denoising diffusion probabilistic models.arXiv preprint arXiv:2208.01864, 2022

    Pith/arXiv arXiv 2022

  16. [16]

    On the guidance of flow matching

    Ruiqi Feng, Chenglei Yu, Wenhao Deng, Peiyan Hu, and Tailin Wu. On the guidance of flow matching. InInternational Conference on Machine Learning (ICML), 2025

    2025

  17. [17]

    Improving diffusion models for inverse problems using optimal posterior covariance

    Xinyu Peng, Ziyang Zheng, Wenrui Dai, Nuoqian Xiao, Chenglin Li, Junni Zou, and Hongkai Xiong. Improving diffusion models for inverse problems using optimal posterior covariance. arXiv preprint arXiv:2402.02149, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Flower: A flow-matching solver for inverse problems

    Mehrsa Pourya, Bassam El Rawas, and Michael Unser. Flower: A flow-matching solver for inverse problems. InInternational Conference on Learning Representations (ICLR), 2025

    2025

  19. [19]

    Mccann, Marc L

    Hyungjin Chung, Jeongsol Kim, Michael T. Mccann, Marc L. Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  20. [20]

    Real-time execution of action chunking flow policies.Advances in Neural Information Processing Systems (NeurIPS), 38:33383– 33407, 2026

    Kevin Black, Manuel Galliker, and Sergey Levine. Real-time execution of action chunking flow policies.Advances in Neural Information Processing Systems (NeurIPS), 38:33383– 33407, 2026

    2026

  21. [21]

    Maddix, Abdul Fatir Ansari, Andrew Stuart, Michael W

    Chaoran Cheng, Boran Han, Danielle C. Maddix, Abdul Fatir Ansari, Andrew Stuart, Michael W. Mahoney, and Bernie Wang. Gradient-free generation for hard-constrained sys- tems. InInternational Conference on Learning Representations (ICLR), 2025

    2025

  22. [22]

    Wei Deng, Yu Chen, Nicole Tianjiao Yang, Hengrong Du, Qi Feng, and Ricky T. Q. Chen. Reflected Schr¨odinger bridge for constrained generative modeling. InUncertainty in Artificial Intelligence (UAI), 2024

    2024

  23. [23]

    Reflected flow matching

    Tianyu Xie, Yu Zhu, Longlin Yu, Tong Yang, Ziheng Cheng, Shiyue Zhang, Xiangyu Zhang, and Cheng Zhang. Reflected flow matching. InInternational Conference on Machine Learning (ICML), 2024

    2024

  24. [24]

    PolyFlow: Safe and efficient polytope-constrained flow matching with constraint embed- ding and projection-free update

    Jianming Ma, Qiyue Yang, Yang Zhang, Liyun Yan, Zhanxiang Cao, Yazhou Zhang, and Yue Gao. PolyFlow: Safe and efficient polytope-constrained flow matching with constraint embed- ding and projection-free update. InInternational Conference on Machine Learning (ICML), 2026

    2026

  25. [25]

    Springer Science & Business Media, 2012

    Anna Nagurney and Ding Zhang.Projected Dynamical Systems and Variational Inequalities With Applications, volume 2. Springer Science & Business Media, 2012. 13

    2012

  26. [26]

    SafeDiffuser: Safe planning with diffusion probabilistic models

    Wei Xiao, Tsun-Hsuan Wang, Chuang Gan, Ramin Hasani, Mathias Lechner, and Daniela Rus. SafeDiffuser: Safe planning with diffusion probabilistic models. InInternational Conference on Learning Representations (ICLR), 2025

    2025

  27. [27]

    SafeFlow: Safe robot motion planning with flow matching via control barrier functions.arXiv preprint arXiv:2504.08661, 2025

    Xiaobing Dai, Zewen Yang, Dian Yu, Fangzhou Liu, Hamid Sadeghian, Sami Haddadin, and Sandra Hirche. SafeFlow: Safe robot motion planning with flow matching via control barrier functions.arXiv preprint arXiv:2504.08661, 2025

    arXiv 2025

  28. [28]

    UniConFlow: A unified constrained flow-matching framework for certified motion planning.arXiv preprint arXiv:2506.02955, 2025

    Zewen Yang, Xiaobing Dai, Dian Yu, Zhijun Li, Majid Khadiv, Sandra Hirche, and Sami Haddadin. UniConFlow: A unified constrained flow-matching framework for certified motion planning.arXiv preprint arXiv:2506.02955, 2025

    arXiv 2025

  29. [29]

    Control barrier functions: Theory and applications

    Aaron D Ames, Samuel Coogan, Magnus Egerstedt, Gennaro Notomista, Koushil Sreenath, and Paulo Tabuada. Control barrier functions: Theory and applications. InEuropean Control Conference (ECC), pages 3420–3431, 2019

    2019

  30. [30]

    Stability of primal–dual gradient dynamics and applica- tions to network optimization.Automatica, 46(12):1974–1981, 2010

    Diego Feijer and Fernando Paganini. Stability of primal–dual gradient dynamics and applica- tions to network optimization.Automatica, 46(12):1974–1981, 2010

    1974

  31. [31]

    Saddle-point dynamics: Condi- tions for asymptotic stability of saddle points.SIAM Journal on Control and Optimization, 55(1):486–511, 2017

    Ashish Cherukuri, Bahman Gharesifard, and Jorge Cort ´es. Saddle-point dynamics: Condi- tions for asymptotic stability of saddle points.SIAM Journal on Control and Optimization, 55(1):486–511, 2017

    2017

  32. [32]

    On the exponential stability of primal-dual gradient dynamics.IEEE Control Systems Letters, 3(1):43–48, 2019

    Guannan Qu and Na Li. On the exponential stability of primal-dual gradient dynamics.IEEE Control Systems Letters, 3(1):43–48, 2019

    2019

  33. [33]

    Khalil.Nonlinear Systems, volume 3

    Hassan K. Khalil.Nonlinear Systems, volume 3. Prentice Hall, 2002

    2002

  34. [34]

    Exponential stability of globally projected dynamic systems.IEEE Transac- tions on Neural Networks, 14(2):426–431, 2003

    Xing-Bao Gao. Exponential stability of globally projected dynamic systems.IEEE Transac- tions on Neural Networks, 14(2):426–431, 2003

    2003

  35. [35]

    Jovanovi ´c

    Sepideh Hassan-Moghaddam and Mihailo R. Jovanovi ´c. Proximal gradient flow and Dou- glas–Rachford splitting dynamics: Global exponential stability via integral quadratic con- straints.Automatica, 123:109311, 2021

    2021

  36. [36]

    Time-varying convex optimization: A contraction and equilibrium tracking approach

    Alexander Davydov, Veronica Centorrino, Anand Gokhale, Giovanni Russo, and Francesco Bullo. Time-varying convex optimization: A contraction and equilibrium tracking approach. IEEE Transactions on Automatic Control, 70(11):7446–7460, 2025

    2025

  37. [37]

    PhD thesis, University of Oxford, 2021

    Patrick Kidger.On Neural Differential Equations. PhD thesis, University of Oxford, 2021

    2021

  38. [38]

    Springer, 1993

    Ernst Hairer, Gerhard Wanner, and Syvert P Nørsett.Solving Ordinary Differential Equations I: Nonstiff Problems. Springer, 1993

    1993

  39. [39]

    Diffusion models beat GANs on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), volume 34, pages 8780–8794, 2021. A Implementation Details Our implementation of LDF, along with baselines and code for reproducing all figures in this paper, is available atgithub.com/vincekurtz/constraine...

    2021

  40. [40]

    Going forward, we will use ′ to denote a derivative with respect toτ

    Sincep >2, we see thatQis strictly positive and non-decreasing, i.e., Q(τ)>0andQ ′(τ)≥0. Going forward, we will use ′ to denote a derivative with respect toτ. Define the energy function V(τ) = 1 Q(τ) (y′)2 +y 2.(48) SinceQ(τ)>0,V(τ)is always nonnegative and satisfies thaty 2 ≤V(τ), meaning that if we can show thatVis bounded, we will conclude thatyis as w...