pith. sign in

arxiv: 2605.26800 · v1 · pith:FVJUQPOZnew · submitted 2026-05-26 · 🧮 math.ST · stat.TH

Accelerated Schr\"odinger-F\"ollmer samplers

Haotian Lin , Xiaojie Wang , Xiaoyan Zhang This is my paper

Pith reviewed 2026-07-01 16:20 UTC · model grok-4.3

classification 🧮 math.ST stat.TH
keywords Schrödinger-Föllmer samplerRunge-Kutta schemeWasserstein convergencediffusion samplingmultimodal distributionsdata-driven samplingstochastic discretizationHölder continuity
0
0 comments X

The pith

A stochastic Runge-Kutta scheme for the Schrödinger-Föllmer sampler reaches O(h^{3/2} |ln h|) convergence in L2-Wasserstein distance.

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

The paper introduces a stochastic Runge-Kutta discretization to speed up the Schrödinger-Föllmer sampler for drawing from complex high-dimensional multimodal distributions. It establishes a convergence rate of O(h^{3/2} |ln h|) in the L2-Wasserstein metric, which improves on the O(h) rate of the standard Euler discretization. The analysis proceeds despite the drift having only 1/2-Hölder continuity in time by deriving careful error bounds that absorb the resulting time singularities into a logarithmic factor. The same framework is extended to the data-driven setting that replaces the target density with an empirical measure. Numerical experiments are used to illustrate the practical gains in accuracy per step.

Core claim

The stochastic Runge-Kutta Schrödinger-Föllmer sampler (SRKSFS) is proved to converge at rate O(h^{3/2} |ln h|) in the L2-Wasserstein distance. This rate holds for a diffusion whose drift is merely 1/2-Hölder continuous in time; the proof relies on delicate error estimates that control the singularities arising from time derivatives of the drift at the cost of the extra logarithmic factor. The construction is further extended to the case in which the target distribution is available only through samples rather than an explicit density.

What carries the argument

Stochastic Runge-Kutta discretization of the Schrödinger-Föllmer diffusion, paired with tailored error estimates that accommodate the 1/2-Hölder time regularity of the drift.

If this is right

  • Fewer discretization steps are needed to reach a given accuracy level compared with Euler-based samplers.
  • The method applies directly to high-dimensional multimodal targets without requiring differentiability of the drift in time.
  • Data-driven sampling becomes feasible when only samples from the target are available.
  • The same acceleration technique can be applied to other diffusion-based samplers that share similar regularity limitations.

Where Pith is reading between the lines

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

  • The approach may lower the computational budget required for posterior sampling in Bayesian models with expensive likelihoods.
  • Further rate improvements could be obtained by smoothing the drift or by using higher-order Runge-Kutta variants once additional regularity is available.
  • The log factor suggests a natural next target: identifying conditions under which the logarithmic penalty can be removed entirely.
  • Implementation in existing stochastic differential equation solvers would allow direct benchmarking against other accelerated samplers.

Load-bearing premise

The drift of the underlying diffusion is only 1/2-Hölder continuous in time and therefore not differentiable.

What would settle it

Numerical computation of the L2-Wasserstein distance between the law of the SRKSFS output and the target measure for a sequence of decreasing step sizes h, verifying whether the observed scaling follows O(h^{3/2} |ln h|) or falls back to a lower order.

Figures

Figures reproduced from arXiv: 2605.26800 by Haotian Lin, Xiaojie Wang, Xiaoyan Zhang.

Figure 1
Figure 1. Figure 1: Sampling of Gaussian Circle with different step sizes. [PITH_FULL_IMAGE:figures/full_fig_p025_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Mean-square convergence rates of Gaussian Circle with different schemes. [PITH_FULL_IMAGE:figures/full_fig_p026_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean-square convergence rates of Gaussian Cross under different schemes. [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sampling of Gaussian Cross using different algorithms. [PITH_FULL_IMAGE:figures/full_fig_p027_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sampling of the Circular Gaussian Mixture using different algorithms. [PITH_FULL_IMAGE:figures/full_fig_p028_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sampling of the Clayton copula ( [PITH_FULL_IMAGE:figures/full_fig_p030_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Sampling of the Clayton copula (d = 5) using different algorithms. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Images generated by a deep generative model with latent dimension [PITH_FULL_IMAGE:figures/full_fig_p031_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Images generated by a deep generative model with latent dimension [PITH_FULL_IMAGE:figures/full_fig_p031_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Sampling from Moons with SRKSFS. Image generation from empirical distributions. To assess the scalability of our method in high-dimensional spaces, we consider image generation tasks using the MNIST [9] and CIFAR-10 datasets [23]. In this setting, the target distribution is defined entirely by the training images, and 33 [PITH_FULL_IMAGE:figures/full_fig_p033_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Sampling from S-curve with SRKSFS. our framework generates novel samples directly from the empirical distribution without density estimation or latent space normalization. The MNIST data set consists of gray-valued digital images, each with 28 × 28 pixels showing one handwritten digit. The generated images for MNIST are presented on the right of [PITH_FULL_IMAGE:figures/full_fig_p034_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Samples and data from MNIST [PITH_FULL_IMAGE:figures/full_fig_p035_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Samples and data from CIFAR-10. 35 [PITH_FULL_IMAGE:figures/full_fig_p035_13.png] view at source ↗
read the original abstract

Sampling is a fundamental algorithmic task in wide-ranging applications across multiple disciplines such as scientific computing, statistics and machine learning. In this paper, an efficient stochastic Runge-Kutta scheme is proposed to accelerate the Schr\"odinger-F\"ollmer sampler, designed for sampling from complex and high-dimensional multimodal distributions. The resulting stochastic Runge-Kutta Schr\"odinger-F\"ollmer sampler (SRKSFS) is proved to achieve a convergence rate of order $\mathcal{O} ( h^{3/2} |\ln h|)$ in the $L^2$-Wasserstein distance, considerably improving the order $\mathcal{O}(h)$ of the existing Euler type sampler. Obtaining the enhanced convergence rate is, however, not trivial, by noting that the drift of the diffusion process is not differentiable but only $\frac{1}{2}$-H\"older continuity with respect to the time variable. To address the difficulty, we rely on delicate error estimates to overcome the singularity due to time derivatives of the drift, at the expense of the logarithmic factor. Furthermore, the framework is extended to data-driven Schr\"odinger-F\"ollmer generation with empirical measures, enabling data-driven sampling without known density. A variety of numerical experiments are reported to validate the effectiveness of the proposed sampling algorithms.

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.

Referee Report

1 major / 0 minor

Summary. The paper proposes a stochastic Runge-Kutta discretization (SRKSFS) of the Schrödinger-Föllmer diffusion to sample from complex multimodal distributions. It asserts a proof that this scheme attains an L²-Wasserstein convergence rate of O(h^{3/2} |ln h|), improving on the O(h) rate of the Euler-Maruyama sampler, despite the drift being merely 1/2-Hölder continuous in time. The framework is extended to data-driven sampling via empirical measures, and numerical experiments are provided to illustrate performance.

Significance. If the error estimates are correct, the result would constitute a concrete advance in high-order numerical methods for sampling, showing that Runge-Kutta schemes can recover an extra half-order (modulo log) even when the time regularity of the drift precludes standard Itô-Taylor expansions. The data-driven extension and the explicit handling of the time singularity are of interest to computational statistics and stochastic analysis.

major comments (1)
  1. [convergence analysis (Theorem establishing the SRKSFS rate)] The central convergence claim (abstract and the theorem establishing the O(h^{3/2}|ln h|) rate) rests entirely on bespoke error estimates that absorb the singularities arising from time derivatives of the 1/2-Hölder drift. The manuscript must supply the complete derivation of these estimates, including the precise application of the Burkholder-Davis-Gundy inequality to the stochastic integrals and the control of the Itô-Taylor remainder terms, so that it can be verified that no additional factors depending on the Hölder constant or on the time mesh enter the final bound.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The single major comment concerns the completeness of the convergence proof. We address it directly below and will incorporate the requested details in the revision.

read point-by-point responses

  1. Referee: The central convergence claim (abstract and the theorem establishing the O(h^{3/2}|ln h|) rate) rests entirely on bespoke error estimates that absorb the singularities arising from time derivatives of the 1/2-Hölder drift. The manuscript must supply the complete derivation of these estimates, including the precise application of the Burkholder-Davis-Gundy inequality to the stochastic integrals and the control of the Itô-Taylor remainder terms, so that it can be verified that no additional factors depending on the Hölder constant or on the time mesh enter the final bound.

    Authors: We agree that the full derivation must be supplied for independent verification. The current manuscript contains the main steps of the error analysis but omits some intermediate calculations involving the Burkholder-Davis-Gundy inequality and the precise bounding of the Itô-Taylor remainders under the 1/2-Hölder time regularity. In the revised version we will add a self-contained appendix that spells out these steps in full, confirming that the final O(h^{3/2}|ln h|) bound does not introduce extraneous factors depending on the Hölder constant or the mesh beyond the logarithmic term already stated. revision: yes

Circularity Check

0 steps flagged

No circularity: convergence proof rests on independent error estimates

full rationale

The paper derives the O(h^{3/2}|ln h|) L^2-Wasserstein rate for the stochastic Runge-Kutta Schrödinger-Föllmer sampler via bespoke error estimates that explicitly handle the 1/2-Hölder time singularity in the drift. No step reduces a claimed result to a fitted parameter, self-definition, or load-bearing self-citation; the analysis is presented as a direct (if delicate) application of stochastic analysis tools such as Burkholder-Davis-Gundy and Itô-Taylor expansions. The derivation chain is therefore self-contained against external mathematical benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, invented entities, or non-standard axioms are mentioned; the work relies on background results from stochastic differential equations and numerical analysis for diffusion processes.

pith-pipeline@v0.9.1-grok · 5760 in / 1130 out tokens · 31316 ms · 2026-07-01T16:20:05.819894+00:00 · methodology

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

45 extracted references · 6 canonical work pages · 2 internal anchors

  1. [1]

    Stochastic interpolants: A uni- fying framework for flows and diffusions.Journal of Machine Learning Research, 26(209):1–80, 2025

    Michael Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A uni- fying framework for flows and diffusions.Journal of Machine Learning Research, 26(209):1–80, 2025. 34 Figure 12: Samples and data from MNIST. Figure 13: Samples and data from CIFAR-10. 35

    2025

  2. [2]

    Building Normalizing Flows with Stochastic In- terpolants

    Michael Albergo and Eric Vanden-Eijnden. Building Normalizing Flows with Stochastic In- terpolants. InThe International Conference on Learning Representations, 2023

    2023

  3. [3]

    Shifted composition iii: Local error framework for kl divergence.arXiv preprint arXiv:2412.17997, 2024

    Jason M Altschuler and Sinho Chewi. Shifted composition iii: Local error framework for kl divergence.arXiv preprint arXiv:2412.17997, 2024

    work page arXiv 2024

  4. [4]

    Sam Bond-Taylor, Adam Leach, Yang Long, and Chris G Willcocks. Deep generative mod- elling: A comparative review of vaes, gans, normalizing flows, energy-based and autoregressive models.IEEE transactions on pattern analysis and machine intelligence, 44(11):7327–7347, 2021

    2021

  5. [5]

    Your gan is secretly an energy-based model and you should use discrimi- nator driven latent sampling.Advances in Neural Information Processing Systems, 33:12275– 12287, 2020

    Tong Che, Ruixiang Zhang, Jascha Sohl-Dickstein, Hugo Larochelle, Liam Paull, Yuan Cao, and Yoshua Bengio. Your gan is secretly an energy-based model and you should use discrimi- nator driven latent sampling.Advances in Neural Information Processing Systems, 33:12275– 12287, 2020

    2020

  6. [6]

    Cheng, N

    Xiang Cheng, Niladri S Chatterji, Yasin Abbasi-Yadkori, Peter L Bartlett, and Michael I Jordan. Sharp convergence rates for Langevin dynamics in the nonconvex setting.arXiv preprint arXiv:1805.01648, 2018

    work page arXiv 2018

  7. [7]

    Global optimization via Schr¨ odinger-F¨ ollmer diffusion.SIAM J

    Yin Dai, Yuling Jiao, Lican Kang, Xiliang Lu, and Jerry Zhijian Yang. Global optimization via Schr¨ odinger-F¨ ollmer diffusion.SIAM J. Control Optim., 61(5):2953–2980, 2023

    2023

  8. [8]

    Arnak S Dalalyan. Theoretical guarantees for approximate sampling from smooth and log- concave densities.Journal of the Royal Statistical Society Series B: Statistical Methodology, 79(3):651–676, 2017

    2017

  9. [9]

    The mnist database of handwritten digit images for machine learning research [best of the web].IEEE signal processing magazine, 29(6):141–142, 2012

    Li Deng. The mnist database of handwritten digit images for machine learning research [best of the web].IEEE signal processing magazine, 29(6):141–142, 2012

    2012

  10. [10]

    Nonasymptotic convergence analysis for the unadjusted Langevin algorithm.The Annals of Applied Probability, 27(3):1551, 2017

    Alain Durmus and Eric Moulines. Nonasymptotic convergence analysis for the unadjusted Langevin algorithm.The Annals of Applied Probability, 27(3):1551, 2017

    2017

  11. [11]

    High-dimensional Bayesian inference via the Unadjusted Langevin Algorithm.Bernoulli, 25(4A), 2019

    Alain Durmus and ´Eric Moulines. High-dimensional Bayesian inference via the Unadjusted Langevin Algorithm.Bernoulli, 25(4A), 2019

    2019

  12. [12]

    Reflection couplings and contraction rates for diffusions.Probability theory and related fields, 166(3):851–886, 2016

    Andreas Eberle. Reflection couplings and contraction rates for diffusions.Probability theory and related fields, 166(3):851–886, 2016

    2016

  13. [13]

    An entropy approach to the time reversal of diffusion processes

    Hans F¨ ollmer. An entropy approach to the time reversal of diffusion processes. InStochas- tic Differential Systems Filtering and Control: Proceedings of the IFIP-WG 7/1 Working Conference Marseille-Luminy, France, March 12–17, 1984, pages 156–163. Springer, 2005

    1984

  14. [14]

    Random fields and diffusion processes

    Hans F¨ ollmer. Random fields and diffusion processes. In ´Ecole d’ ´Et´ e de Probabilit´ es de Saint-Flour XV–XVII, 1985–87, pages 101–203. Springer, 2006

    1985

  15. [15]

    OUP Oxford, 2002

    Paul H Garthwaite, Ian T Jolliffe, and Byron Jones.Statistical inference. OUP Oxford, 2002

    2002

  16. [16]

    Chapman and Hall/CRC, 1995

    Andrew Gelman, John B Carlin, Hal S Stern, and Donald B Rubin.Bayesian data analysis. Chapman and Hall/CRC, 1995. 36

    1995

  17. [17]

    Number 25

    Jack K Hale.Asymptotic behavior of dissipative systems. Number 25. American Mathematical Soc., 2010

    2010

  18. [18]

    One-step data-driven generative model via schr¨ odinger bridge.arXiv preprint arXiv:2405.12453, 2024

    Hanwen Huang. One-step data-driven generative model via schr¨ odinger bridge.arXiv preprint arXiv:2405.12453, 2024

    work page arXiv 2024

  19. [19]

    Schr¨ odinger-F¨ ollmer sampler.IEEE Trans

    Jian Huang, Yuling Jiao, Lican Kang, Xu Liao, Jin Liu, and Yanyan Liu. Schr¨ odinger-F¨ ollmer sampler.IEEE Trans. Inform. Theory, 71(2):1283–1299, 2025

    2025

  20. [20]

    Multimodal conditional image synthesis with product-of-experts gans

    Xun Huang, Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Multimodal conditional image synthesis with product-of-experts gans. InEuropean conference on computer vision, pages 91–109. Springer, 2022

    2022

  21. [21]

    World Scientific Publishing Company, 2012

    Fima C Klebaner.Introduction to stochastic calculus with applications. World Scientific Publishing Company, 2012

    2012

  22. [22]

    Kloeden and Eckhard Platen.Numerical Solution of Stochastic Differential Equa- tions

    Peter E. Kloeden and Eckhard Platen.Numerical Solution of Stochastic Differential Equa- tions. Springer, Berlin, Heidelberg, 1992

    1992

  23. [23]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    2009

  24. [24]

    A sharp uniform-in-time error estimate for Stochastic Gradient Langevin Dynamics.CSIAM Transactions on Applied Mathematics, 6(4):711–759, 2025

    Lei Li and Yuliang Wang. A sharp uniform-in-time error estimate for Stochastic Gradient Langevin Dynamics.CSIAM Transactions on Applied Mathematics, 6(4):711–759, 2025

    2025

  25. [25]

    Sqrt (d) dimension dependence of Langevin monte carlo.The International Conference on Learning Representations, 2022

    Ruilin Li, Hongyuan Zha, and Molei Tao. Sqrt (d) dimension dependence of Langevin monte carlo.The International Conference on Learning Representations, 2022

    2022

  26. [26]

    Stochastic runge-kutta accelerates langevin monte carlo and beyond.Advances in neural information processing systems, 32, 2019

    Xuechen Li, Yi Wu, Lester Mackey, and Murat A Erdogdu. Stochastic runge-kutta accelerates langevin monte carlo and beyond.Advances in neural information processing systems, 32, 2019

    2019

  27. [27]

    Elsevier, 2007

    Xuerong Mao.Stochastic differential equations and applications. Elsevier, 2007

    2007

  28. [28]

    Springer, 2004

    Grigori N Milstein and Michael V Tretyakov.Stochastic numerics for mathematical physics, volume 39. Springer, 2004

    2004

  29. [29]

    Im- proved bounds for discretization of Langevin diffusions: Near-optimal rates without convexity

    Wenlong Mou, Nicolas Flammarion, Martin J Wainwright, and Peter L Bartlett. Im- proved bounds for discretization of Langevin diffusions: Near-optimal rates without convexity. Bernoulli, 28(3):1577–1601, 2022

    2022

  30. [30]

    Springer, 2006

    Roger B Nelsen.An introduction to copulas. Springer, 2006

    2006

  31. [31]

    Learning latent space energy-based prior model.Advances in Neural Information Processing Systems, 33:21994– 22008, 2020

    Bo Pang, Tian Han, Erik Nijkamp, Song-Chun Zhu, and Ying Nian Wu. Learning latent space energy-based prior model.Advances in Neural Information Processing Systems, 33:21994– 22008, 2020

    2020

  32. [32]

    Projected Langevin Monte Carlo algorithms in non-convex and super-linear setting.Journal of Computational Physics, 526:113754, 2025

    Chenxu Pang, Xiaojie Wang, and Yue Wu. Projected Langevin Monte Carlo algorithms in non-convex and super-linear setting.Journal of Computational Physics, 526:113754, 2025. 37

    2025

  33. [33]

    Scikit-learn: Machine learning in python.The Journal of Machine Learning Research, 12:2825–2830, 2011

    Fabian Pedregosa, Ga¨ el Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python.The Journal of Machine Learning Research, 12:2825–2830, 2011

    2011

  34. [34]

    Efficient multimodal sampling via tempered distribution flow

    Yixuan Qiu and Xiao Wang. Efficient multimodal sampling via tempered distribution flow. Journal of the American Statistical Association, 119(546):1446–1460, 2024

    2024

  35. [35]

    Unbiased estimation using a class of diffusion processes.Journal of Computational Physics, 472:111643, 2023

    Hamza Ruzayqat, Alexandros Beskos, Dan Crisan, Ajay Jasra, and Nikolas Kantas. Unbiased estimation using a class of diffusion processes.Journal of Computational Physics, 472:111643, 2023

    2023

  36. [36]

    Learning deep generative models.Annual Review of Statistics and Its Application, 2(1):361–385, 2015

    Ruslan Salakhutdinov. Learning deep generative models.Annual Review of Statistics and Its Application, 2(1):361–385, 2015

    2015

  37. [37]

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

    Erwin Schr¨ odinger. Sur la th´ eorie relativiste de l’´ electron et l’interpr´ etation de la m´ ecanique quantique. InAnnales de l’institut Henri Poincar´ e, volume 2, pages 269–310, 1932

    1932

  38. [38]

    Schuh and P

    Katharina Schuh and Peter A Whalley. Convergence of kinetic Langevin samplers for non- convex potentials.arXiv preprint arXiv:2405.09992, 2024

    work page arXiv 2024

  39. [39]

    Springer, 2015

    Timothy John Sullivan.Introduction to uncertainty quantification, volume 63. Springer, 2015

    2015

  40. [40]

    Multimodal sampling via Schr¨ odinger-F¨ ollmer samplers with temperatures.Journal of Complexity, 96:102052, 2026

    Xiaojie Wang and Xiaoyan Zhang. Multimodal sampling via Schr¨ odinger-F¨ ollmer samplers with temperatures.Journal of Complexity, 96:102052, 2026

    2026

  41. [41]

    Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms

    Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms.arXiv preprint arXiv:1708.07747, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  42. [42]

    Bin Yang and Xiaojie Wang. Non-asymptotic Error Bounds inW 2-Distance with Sqrt(d) Dimension Dependence and First Order Convergence for Langevin Monte Carlo beyond Log- Concavity.International Conference on Machine Learning, 2025

    2025

  43. [43]

    Accelerating Langevin Monte Carlo via Efficient Stochastic Runge--Kutta Methods beyond Log-Concavity

    Bin Yang and Xiaojie Wang. Accelerating Langevin Monte Carlo via Efficient Stochastic Runge–Kutta Methods beyond Log-Concavity.arXiv preprint arXiv:2605.07939, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  44. [44]

    Diffusion models: A comprehensive survey of methods and applications.ACM computing surveys, 56(4):1–39, 2023

    Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications.ACM computing surveys, 56(4):1–39, 2023

    2023

  45. [45]

    Deep generative molecular design reshapes drug discovery.Cell Reports Medicine, 3(12), 2022

    Xiangxiang Zeng, Fei Wang, Yuan Luo, Seung-gu Kang, Jian Tang, Felice C Lightstone, Evandro F Fang, Wendy Cornell, Ruth Nussinov, and Feixiong Cheng. Deep generative molecular design reshapes drug discovery.Cell Reports Medicine, 3(12), 2022. 38 A Proof of Proposition 3.2 Proof.By assumptions, the functiong β is of classC 4, and moreover,g β,∇g β,∇ 2gβ,∇ ...

    2022