arxiv: 2605.07661 · v1 · submitted 2026-05-08 · 💻 cs.LG · cs.CV

Recognition: 2 theorem links

· Lean Theorem

Stochastic Transition-Map Distillation for Fast Probabilistic Inference

George Rapakoulias , Peter Garud , Lingjiong Zhu , Panagiotis Tsiotras

Authors on Pith no claims yet

Pith reviewed 2026-05-11 02:21 UTC · model grok-4.3

classification 💻 cs.LG cs.CV

keywords diffusion modelsstochastic inferencemodel distillationfast samplingSDEprobabilistic generationimage generationWasserstein distance

0 comments

The pith

Diffusion model sampling accelerates to one or few steps while preserving full probabilistic structure by distilling the entire transition map of the sampling SDE.

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

The paper proposes Stochastic Transition-Map Distillation as a teacher-free way to speed up diffusion-based image generation. Standard diffusion models rely on many iterative steps from a stochastic differential equation, which makes inference slow. STMD instead captures the complete transition probabilities of that SDE rather than only its mean, then fits them with a conditional Mean Flow model. This produces fast stochastic samplers that keep the diversity and coverage properties needed for tasks like inverse problems or posterior sampling. The approach avoids bi-level optimization and trajectory caching, and comes with explicit Wasserstein convergence bounds that are tested on MNIST, CIFAR-10, and CelebA.

Core claim

STMD distills the full transition map associated with the sampling SDE by parameterizing those transitions with a conditional Mean Flow model, which yields a one- or few-step stochastic sampler that retains the transition structure of the underlying diffusion process. The method requires no pretrained teacher, bi-level optimization, or trajectory simulation, and is supported by derived Wasserstein-distance convergence bounds.

What carries the argument

Conditional Mean Flow model that parameterizes the complete probabilistic transitions of the sampling SDE instead of only the posterior mean.

If this is right

One- or few-step sampling becomes possible while retaining the stochastic character required for downstream probabilistic tasks.
Diffusion posterior sampling and inverse-problem solvers can use the distilled sampler directly without retraining the underlying diffusion model.
Energy-based fine-tuning of diffusion models can be combined with the fast stochastic sampler for controlled generation.
Training scales efficiently because no teacher network or cached trajectories are needed.
Wasserstein convergence guarantees provide a quantitative way to monitor how well the distilled transitions match the original SDE.

Where Pith is reading between the lines

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

The same transition-map distillation idea could be applied to other SDE-driven generative processes such as stochastic normalizing flows.
Real-time image editing pipelines that currently use slow diffusion sampling might adopt STMD to reach interactive speeds while keeping output variety.
The Wasserstein bounds could be turned into a practical regularizer during training to enforce diversity preservation explicitly.

Load-bearing premise

A conditional Mean Flow model can accurately capture and distill the full probabilistic transition map of the underlying SDE without loss of structure or diversity, and the Wasserstein convergence bounds translate to practical preservation of stochastic properties on image data.

What would settle it

Train an STMD one-step sampler on CIFAR-10, generate equal numbers of samples from it and from the original multi-step diffusion model, then measure both FID score and sample diversity statistics; if the distilled sampler produces statistically indistinguishable coverage and variance, the claim holds, while collapse in diversity or large FID gap would falsify it.

Figures

Figures reproduced from arXiv: 2605.07661 by George Rapakoulias, Lingjiong Zhu, Panagiotis Tsiotras, Peter Garud.

**Figure 2.** Figure 2: Stochastic Transition-Map Distillation. Consider a forward diffusion process dxt = − 1 2 βtxt dt + p βt dwt, x0 ∼ p0, (14) where wt is a standard d-dimensional Brownian motion and βt ≥ 0 is a scalar function of time. To avoid confusion with the Mean Flow variable zs, we will denote probability densities associated with the diffusion variable xt using pt, and denote the diffusion time with t. The SDE (14… view at source ↗

**Figure 3.** Figure 3: (a): Unconditional MNIST samples using various [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: (a): Unconditional CIFAR10 samples using various [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Unconditional generation on the CelebA dataset. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

**Figure 6.** Figure 6: Image inpainting on the CelebA dataset. From left to right: original image, masked image, [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

**Figure 7.** Figure 7: MNIST generation results. 1https://github.com/Gsunshine/meanflow 2https://huggingface.co/stabilityai/sd-vae-ft-ema 20 [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗

**Figure 8.** Figure 8: CIFAR10 generation results. D.3 Additional CelebA images [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗

**Figure 9.** Figure 9: CelebA generation with ninf = 4, nmf = 2 21 [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗

read the original abstract

Diffusion models achieve strong generation quality, diversity, and distribution coverage, but their performance often comes with expensive inference. In this work, we propose Stochastic Transition-Map Distillation (STMD), a teacher-free framework for accelerating diffusion model inference while preserving probabilistic sample generation. In contrast to score-based diffusion models, whose denoising parametrization models the mean of the posterior distribution, STMD distills the full transition map associated with the sampling stochastic differential equation (SDE). We parameterize these SDE transitions with a conditional Mean Flow model, yielding a one- or few-step stochastic sampler that retains the transition structure of the underlying diffusion process. This perspective is especially useful for downstream tasks that require stochastic inference, such as diffusion posterior sampling, inverse problems, and energy-based fine-tuning. Compared to recent distillation methods, STMD requires no pretrained teacher, bi-level optimization, or trajectory simulation and caching, enabling efficient and scalable training. We derive convergence bounds for our method in the Wasserstein distance, providing a strong theoretical foundation for our approach, and validate STMD on various image generation examples on the MNIST, CIFAR-10, and CelebA datasets.

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

STMD tries to distill the full SDE transition kernel teacher-free via conditional mean flow for one/few-step stochastic sampling, but the Wasserstein bounds alone do not clearly guarantee retained variance and noise structure.

read the letter

The main new piece here is the teacher-free distillation of the complete transition map of the sampling SDE, parameterized as a conditional mean flow model instead of the usual score or posterior mean. This avoids pretrained teachers, bi-level optimization, and trajectory caching, which is a practical advantage for scaling. The Wasserstein convergence bounds and the positioning for downstream stochastic tasks like posterior sampling are also clear positives if the derivations check out. Experiments on MNIST, CIFAR-10, and CelebA are standard starting points for this kind of work. The soft spot is exactly the one in the stress test: Wasserstein distance controls marginal closeness but does not automatically preserve the variance or noise correlations inside the transition kernel. The abstract describes the model as capturing the full transition structure, yet gives no detail on the training loss or how stochasticity is enforced beyond the parameterization itself. If the objective reduces to regressing the conditional mean, the sampler can satisfy the bound while collapsing diversity, which would undermine the claim of retaining the underlying diffusion process. Without seeing the exact loss, the bound assumptions, or metrics that directly test stochastic fidelity (not just FID or sample quality), it is hard to know whether the method delivers on the probabilistic part. This is aimed at people working on fast inference for diffusion models in inverse problems or energy-based settings. It is coherent enough on its own terms to deserve a serious referee, though the review will likely focus on whether the stochastic properties survive the distillation.

Referee Report

2 major / 2 minor

Summary. The paper proposes Stochastic Transition-Map Distillation (STMD), a teacher-free framework that distills the full transition map of the sampling SDE in diffusion models by parameterizing it with a conditional Mean Flow model. This yields a one- or few-step stochastic sampler that aims to retain the underlying diffusion process's transition structure, enabling fast probabilistic inference for tasks like posterior sampling and inverse problems. The authors derive convergence bounds in Wasserstein distance and provide empirical validation on MNIST, CIFAR-10, and CelebA image generation tasks, emphasizing the absence of pretrained teachers, bi-level optimization, or trajectory caching.

Significance. If the central claims hold, STMD would offer a scalable, teacher-free route to fast stochastic sampling in diffusion models while providing theoretical guarantees via Wasserstein bounds; this is particularly relevant for downstream applications that rely on preserving sample diversity and noise structure rather than deterministic mean predictions. The avoidance of trajectory simulation and caching is a practical strength compared to prior distillation approaches.

major comments (2)

[Abstract] Abstract: The claim that the conditional Mean Flow 'distills the full transition map' and 'retains the transition structure' is load-bearing for the probabilistic inference contribution, yet the description does not specify the training loss or objective used to match the SDE transition kernel p(x_{t-Δt}|x_t). If the loss effectively regresses only on conditional expectations (as is common in mean-flow parameterizations), the derived Wasserstein bounds on marginal distances would not necessarily prevent variance collapse or loss of noise correlations, undermining the distinction from standard score-based mean denoising.
[Theoretical Analysis] Theoretical section (convergence bounds): The Wasserstein convergence bounds are presented as a strong foundation, but it is unclear whether they are derived under the assumption that the Mean Flow exactly represents the full transition kernel or merely approximates it with separately controlled error. Without explicit control on higher moments or stochastic fidelity in the bound derivation, the bounds may not translate to practical retention of the SDE's probabilistic properties on image data.

minor comments (2)

[Abstract] The abstract and introduction would benefit from a brief explicit statement of the training loss function and how it differs from standard conditional expectation regression to support the 'full transition map' claim.
[Experiments] Empirical results on MNIST, CIFAR-10, and CelebA should include quantitative metrics for sample diversity (e.g., FID with variance across multiple runs or entropy measures) to demonstrate that stochastic properties are preserved beyond visual quality.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive report. We address each major comment below with clarifications drawn directly from the manuscript and indicate revisions that will be incorporated to improve clarity without altering the core claims or results.

read point-by-point responses

Referee: [Abstract] The claim that the conditional Mean Flow 'distills the full transition map' and 'retains the transition structure' is load-bearing for the probabilistic inference contribution, yet the description does not specify the training loss or objective used to match the SDE transition kernel p(x_{t-Δt}|x_t). If the loss effectively regresses only on conditional expectations (as is common in mean-flow parameterizations), the derived Wasserstein bounds on marginal distances would not necessarily prevent variance collapse or loss of noise correlations, undermining the distinction from standard score-based mean denoising.

Authors: The manuscript defines the training objective for the conditional Mean Flow explicitly as the minimization of the expected squared L2 error between the predicted transition and the true conditional mean of the SDE transition kernel, while the stochastic component (including variance and noise correlations) is preserved by injecting noise drawn from the known SDE transition variance schedule during sampling. This separation ensures the full transition structure is retained rather than collapsing to a deterministic mean prediction. The Wasserstein bounds then apply to the resulting stochastic process. We agree the abstract is too terse on this point and will revise it to state the loss and the explicit retention of stochasticity via the variance schedule. revision: yes
Referee: [Theoretical Analysis] The Wasserstein convergence bounds are presented as a strong foundation, but it is unclear whether they are derived under the assumption that the Mean Flow exactly represents the full transition kernel or merely approximates it with separately controlled error. Without explicit control on higher moments or stochastic fidelity in the bound derivation, the bounds may not translate to practical retention of the SDE's probabilistic properties on image data.

Authors: The bounds are derived under the assumption of a controlled approximation error in the Mean Flow's prediction of the transition mean (with the error term appearing explicitly in the proof), combined with the exact variance schedule of the underlying SDE. Because the diffusion transitions are Gaussian, the Wasserstein distance between the approximated and true transition kernels directly controls both the mean and variance discrepancies, which in turn bounds higher moments for this family of distributions. We will revise the theoretical section to make these assumptions and the Gaussian case explicit, including a short remark on how the error propagation preserves the probabilistic properties observed in the experiments. revision: yes

Circularity Check

0 steps flagged

No circularity: Wasserstein bounds and Mean Flow parameterization presented as independently derived

full rationale

The abstract claims derivation of convergence bounds in Wasserstein distance for the STMD method after parameterizing SDE transitions via a conditional Mean Flow model. No equations, self-citations, or fitted inputs are shown that reduce these bounds or the stochastic sampler to the training loss or inputs by construction. The framework is explicitly teacher-free and avoids trajectory simulation, positioning the theoretical results as self-contained first-principles derivations rather than renamings or self-referential fits. This matches the default expectation of no significant circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 2 invented entities

Review uses only the abstract; ledger entries are therefore high-level inferences from the stated method. Full paper would likely reveal additional fitted parameters in the Mean Flow model and SDE assumptions.

free parameters (1)

Conditional Mean Flow model parameters
The model is trained on data to approximate transitions, so its weights constitute fitted parameters central to the sampler.

axioms (1)

domain assumption The sampling process of diffusion models is governed by an SDE whose full transition map can be distilled into a simpler parametric form while preserving probabilistic structure.
Invoked as the foundation for parameterizing transitions with the Mean Flow model.

invented entities (2)

Stochastic Transition-Map Distillation (STMD) no independent evidence
purpose: Teacher-free acceleration framework that distills SDE transitions.
Newly proposed method name and procedure.
Conditional Mean Flow model no independent evidence
purpose: Parametric model that represents the full SDE transition map for fast stochastic sampling.
Core technical device introduced to replace iterative denoising.

pith-pipeline@v0.9.0 · 5511 in / 1656 out tokens · 56511 ms · 2026-05-11T02:21:31.486926+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We parameterize these SDE transitions with a conditional Mean Flow model, yielding a one- or few-step stochastic sampler... We derive convergence bounds for our method in the Wasserstein distance
IndisputableMonolith/Foundation/BranchSelection.lean branch_selection unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

STMD distills the full transition map associated with the sampling stochastic differential equation (SDE)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

200 extracted references · 200 canonical work pages · 3 internal anchors

[1]

2007 , publisher=

Stochastic processes and filtering theory , author=. 2007 , publisher=

work page 2007
[2]

Journal of Machine Learning Research , volume=

Normalizing flows for probabilistic modeling and inference , author=. Journal of Machine Learning Research , volume=. 2021 , publisher=

work page 2021
[3]

Chen, Ricky T. Q. and Rubanova, Yulia and Bettencourt, Jesse and Duvenaud, David K , booktitle =. Neural Ordinary Differential Equations , address=

work page
[4]

Advances in Neural Information Processing Systems , volume=

Dissecting Neural Odes , author=. Advances in Neural Information Processing Systems , volume=. 2020 , address=

work page 2020
[5]

2021 , address=

Onken, Derek and Fung, Samy Wu and Li, Xingjian and Ruthotto, Lars , booktitle=. 2021 , address=

work page 2021
[6]

Deep Generalized

Liu, Guan-Horng and Chen, Tianrong and So, Oswin and Theodorou, Evangelos , booktitle =. Deep Generalized

work page
[7]

2024 , address=

Liu, Guan-Horng and Lipman, Yaron and Nickel, Maximilian and Karrer, Brian and Theodorou, Evangelos A and Chen, Ricky TQ , booktitle=. 2024 , address=

work page 2024
[8]

Likelihood Training of

Chen, Tianrong and Liu, Guan-Horng and Theodorou, Evangelos A , booktitle=. Likelihood Training of. 2022 , month=

work page 2022
[9]

Proceedings of the National Academy of Sciences , volume=

A machine learning framework for solving high-dimensional mean field game and mean field control problems , author=. Proceedings of the National Academy of Sciences , volume=. 2020 , publisher=

work page 2020
[10]

International Conference on Machine Learning , pages=

Wasserstein generative adversarial networks , author=. International Conference on Machine Learning , pages=. 2017 , address=

work page 2017
[11]

Auto-encoding variational

Kingma, Diederik P and Welling, Max , booktitle=. Auto-encoding variational

work page
[12]

Foundations and Trends

An introduction to variational autoencoders , author=. Foundations and Trends. 2019 , publisher=

work page 2019
[13]

GAMM-Mitteilungen , volume=

An introduction to deep generative modeling , author=. GAMM-Mitteilungen , volume=. 2021 , publisher=

work page 2021
[14]

2013 , publisher=

Stochastic differential equations: an introduction with applications , author=. 2013 , publisher=

work page 2013
[15]

Advances in Neural Information Processing Systems , volume=

A variational perspective on diffusion-based generative models and score matching , author=. Advances in Neural Information Processing Systems , volume=. 2021 , address=

work page 2021
[16]

International Conference on Learning Representations , year=

Score-Based Generative Modeling through Stochastic Differential Equations , author=. International Conference on Learning Representations , year=

work page
[17]

Automatica , volume=

Stochastic optimal control via forward and backward stochastic differential equations and importance sampling , author=. Automatica , volume=. 2018 , publisher=

work page 2018
[18]

A gentle introduction to concentration inequalities , author=. Dept. Comput. Sci., Cornell Univ., Tech. Rep , year=

work page
[19]

International Conference on Machine Learning , pages=

Invertible residual networks , author=. International Conference on Machine Learning , pages=. 2019 , organization=

work page 2019
[20]

International Conference on Learning Representations , year=

Convex Potential Flows: Universal Probability Distributions with Optimal Transport and Convex Optimization , author=. International Conference on Learning Representations , year=

work page
[21]

International Conference on Machine Learning , pages=

Input convex neural networks , author=. International Conference on Machine Learning , pages=. 2017 , address=

work page 2017
[22]

arXiv preprint arXiv:2012.12348 , year=

An overview on deep learning-based approximation methods for partial differential equations , author=. arXiv preprint arXiv:2012.12348 , year=

work page arXiv 2012
[23]

International Conference on Learning Representations , address=

Flow Matching for Generative Modeling , author=. International Conference on Learning Representations , address=

work page
[24]

International Conference on Learning Representations , address=

Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author=. International Conference on Learning Representations , address=

work page
[25]

Transactions on Automatic Control , pages=

Optimal covariance steering for discrete-time linear stochastic systems , author=. Transactions on Automatic Control , pages=

work page
[26]

Discrete-Time Optimal Covariance Steering via Semidefinite Programming , year=

Rapakoulias, George and Tsiotras, Panagiotis , booktitle=. Discrete-Time Optimal Covariance Steering via Semidefinite Programming , year=

work page
[27]

Optimal Steering of a Linear Stochastic System to a Final Probability Distribution, Part

Chen, Yongxin and Georgiou, Tryphon T and Pavon, Michele , journal=. Optimal Steering of a Linear Stochastic System to a Final Probability Distribution, Part. 2015 , publisher=

work page 2015
[28]

Control Systems Letters , volume=

Optimal covariance control for stochastic systems under chance constraints , author=. Control Systems Letters , volume=. 2018 , publisher=

work page 2018
[29]

34th Conference on Uncertainty in Artificial Intelligence , pages=

Sylvester normalizing flows for variational inference , author=. 34th Conference on Uncertainty in Artificial Intelligence , pages=. 2018 , address=

work page 2018
[30]

Transactions on Automatic Control , volume=

Optimal transport over a linear dynamical system , author=. Transactions on Automatic Control , volume=. 2016 , publisher=

work page 2016
[31]

Optimal steering of a linear stochastic system to a final probability distribution, Part

Chen, Yongxin and Georgiou, Tryphon T and Pavon, Michele , journal=. Optimal steering of a linear stochastic system to a final probability distribution, Part. 2015 , publisher=

work page 2015
[32]

Balci, Isin M and Bakolas, Efstathios , journal=. Exact

work page
[33]

Transactions on Control Systems Technology , publisher=

A neural network approach for high-dimensional optimal control applied to multiagent path finding , author=. Transactions on Control Systems Technology , publisher=

work page
[34]

Distributionally Robust Covariance Steering with Optimal Risk Allocation , year=

Renganathan, Venkatraman and Pilipovsky, Joshua and Tsiotras, Panagiotis , booktitle=. Distributionally Robust Covariance Steering with Optimal Risk Allocation , year=

work page
[35]

Automatica , volume=

Finite-horizon covariance control for discrete-time stochastic linear systems subject to input constraints , author=. Automatica , volume=. 2018 , publisher=

work page 2018
[36]

Wasserstein proximal algorithms for the

Caluya, Kenneth F and Halder, Abhishek , journal=. Wasserstein proximal algorithms for the. 2021 , publisher=

work page 2021
[37]

2020 AAS/AIAA Astrodynamics Specialist Conference , pages=

Chance-constrained covariance control for low-thrust minimum-fuel trajectory optimization , author=. 2020 AAS/AIAA Astrodynamics Specialist Conference , pages=

work page 2020
[38]

Density Steering of

Balci, Isin and Bakolas, Efstathios , booktitle=. Density Steering of

work page
[39]

Mosek modeling cookbook , author=

work page
[40]

2019 , publisher=

Peyré, Gabriel and Cuturi, Marco , title=. 2019 , publisher=

work page 2019
[41]

International Conference on Learning Representations , address=

Spectral Normalization for Generative Adversarial Networks , author=. International Conference on Learning Representations , address=

work page
[42]

Improved Training of

Gulrajani, Ishaan and Ahmed, Faruk and Arjovsky, Martin and Dumoulin, Vincent and Courville, Aaron C , booktitle =. Improved Training of

work page
[43]

International Conference on Learning Representations , address=

Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations , address=

work page
[44]

Paszke, Adam and Gross, Sam and Massa, Francisco and Lerer, Adam and Bradbury, James and Chanan, Gregory and Killeen, Trevor and Lin, Zeming and Gimelshein, Natalia and Antiga, Luca and Desmaison, Alban and K\". Py. Advances in Neural Information Processing Systems , volume=. 2019 , pages=

work page 2019
[45]

Advances in Neural Information Processing Systems , volume=

Parallelized stochastic gradient descent , author=. Advances in Neural Information Processing Systems , volume=

work page
[46]

arXiv preprint arXiv:2310.02485 , year=

Computationally Efficient Chance Constrained Covariance Control with Output Feedback , author=. arXiv preprint arXiv:2310.02485 , year=

work page arXiv
[47]

Density Control of Interacting Agent Systems , year=

Chen, Yongxin , journal=. Density Control of Interacting Agent Systems , year=

work page
[48]

arXiv preprint arXiv:1904.02272 , year=

Finite horizon density control for static state feedback linearizable systems , author=. arXiv preprint arXiv:1904.02272 , year=

work page arXiv 1904
[49]

American Control Conference , pages=

Finite horizon density steering for multi-input state feedback linearizable systems , author=. American Control Conference , pages=. 2020 , organization=

work page 2020
[50]

Advances in Neural Information Processing Systems , volume=

Denoising diffusion probabilistic models , author=. Advances in Neural Information Processing Systems , volume=

work page
[51]

Bernstein, Dennis S , year=

work page
[52]

arXiv preprint arXiv:2212.00398 , year=

Distributed model predictive covariance steering , author=. arXiv preprint arXiv:2212.00398 , year=

work page arXiv
[53]

Robotics: Science and Systems XIX , address=

Distributed Hierarchical Distribution Control for Very-Large-Scale Clustered Multi-Agent Systems , author=. Robotics: Science and Systems XIX , address=

work page
[54]

arXiv preprint arXiv:2301.06227 , year=

General Distribution Steering: A Sub-Optimal Solution by Convex Optimization , author=. arXiv preprint arXiv:2301.06227 , year=

work page arXiv
[55]

Safe High-Performance Autonomous Off-Road Driving Using Covariance Steering Stochastic Model Predictive Control , year=

Knaup, Jacob and Okamoto, Kazuhide and Tsiotras, Panagiotis , journal=. Safe High-Performance Autonomous Off-Road Driving Using Covariance Steering Stochastic Model Predictive Control , year=

work page
[56]

On the relation between optimal transport and

Chen, Yongxin and Georgiou, Tryphon T and Pavon, Michele , journal=. On the relation between optimal transport and. 2016 , publisher=

work page 2016
[57]

Diffusion

De Bortoli, Valentin and Thornton, James and Heng, Jeremy and Doucet, Arnaud , booktitle=. Diffusion. 2021 , address=

work page 2021
[58]

Inverse Problems , volume=

Stable architectures for deep neural networks , author=. Inverse Problems , volume=. 2017 , publisher=

work page 2017
[59]

Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

Deep residual learning for image recognition , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

work page
[60]

Berkeley, California , volume=

Derivations for linear algebra and optimization , author=. Berkeley, California , volume=

work page
[61]

Learning for Dynamics and Control Conference , address=

Probabilistic verification of ReLU neural networks via characteristic functions , author=. Learning for Dynamics and Control Conference , address=. 2023 , organization=

work page 2023
[62]

American Control Conference , pages=

Distribution steering for discrete-time linear systems with general disturbances using characteristic functions , author=. American Control Conference , pages=

work page
[63]

Control Systems Magazine , volume=

Controlling uncertainty , author=. Control Systems Magazine , volume=. 2021 , publisher=

work page 2021
[64]

Stochastic control liaisons:

Chen, Yongxin and Georgiou, Tryphon T and Pavon, Michele , journal=. Stochastic control liaisons:. 2021 , publisher=

work page 2021
[65]

2019 , school=

LARGE SCALE STOCHASTIC CONTROL: ALGORITHMS, OPTIMALITY AND STABILITY , author=. 2019 , school=

work page 2019
[66]

2020 , publisher=

Bakshi, Kaivalya and Fan, David D and Theodorou, Evangelos A , journal=. 2020 , publisher=

work page 2020
[67]

Annual Review of Control, Robotics, and Autonomous Systems , volume=

Optimal transport in systems and control , author=. Annual Review of Control, Robotics, and Autonomous Systems , volume=. 2021 , publisher=

work page 2021
[68]

Automatica , volume=

Minimum-energy covariance controllers , author=. Automatica , volume=. 1997 , publisher=

work page 1997
[69]

International Journal of Control , volume=

Covariance control theory , author=. International Journal of Control , volume=. 1987 , month = jul, publisher=

work page 1987
[70]

Science Robotics , volume=

Neural-fly enables rapid learning for agile flight in strong winds , author=. Science Robotics , volume=. 2022 , publisher=

work page 2022
[71]

Conference on Decision and Control , pages=

A Physics-Informed Neural Networks Framework to Solve the Infinite-Horizon Optimal Control Problem , author=. Conference on Decision and Control , pages=. 2023 , address=

work page 2023
[72]

Annals of Probability , volume=

Lazy random walks and optimal transport on graphs , author=. Annals of Probability , volume=

work page
[73]

Advances in Computational Intelligence Systems: Contributions Presented at the 18th UK Workshop on Computational Intelligence, September 5-7, 2018, Nottingham, UK , pages=

A study on CNN transfer learning for image classification , author=. Advances in Computational Intelligence Systems: Contributions Presented at the 18th UK Workshop on Computational Intelligence, September 5-7, 2018, Nottingham, UK , pages=. 2019 , organization=

work page 2018
[74]

Improving and generalizing flow-based generative models with minibatch optimal transport

Improving and generalizing flow-based generative models with minibatch optimal transport , author=. arXiv preprint arXiv:2302.00482 , year=

work page internal anchor Pith review arXiv
[75]

arXiv preprint arXiv:2307.03672 , year=

Simulation-free Schr " odinger bridges via score and flow matching , author=. arXiv preprint arXiv:2307.03672 , year=

work page arXiv
[76]

2019 , publisher=

Applied Stochastic Differential Equations , author=. 2019 , publisher=

work page 2019
[77]

Random fields and diffusion processes , author=. Lect. Notes Math , volume=

work page
[78]

Diffusion

Shi, Yuyang and De Bortoli, Valentin and Campbell, Andrew and Doucet, Arnaud , booktitle=. Diffusion

work page
[79]

Entropic and displacement interpolation: a computational approach using the

Chen, Yongxin and Georgiou, Tryphon and Pavon, Michele , journal=. Entropic and displacement interpolation: a computational approach using the. 2016 , publisher=

work page 2016
[80]

The Data-Driven

Pavon, Michele and Trigila, Giulio and Tabak, Esteban G , journal=. The Data-Driven. 2021 , publisher=

work page 2021

Showing first 80 references.