pith. sign in

arxiv: 2509.01685 · v2 · pith:L3OCR3OWnew · submitted 2025-09-01 · 📊 stat.ML · cs.LG· math.OC· stat.CO

Preconditioned Regularized Wasserstein Proximal Sampling

Hong Ye Tan , Stanley Osher , Wuchen Li This is my paper

Pith reviewed 2026-05-21 23:31 UTC · model grok-4.3

classification 📊 stat.ML cs.LGmath.OCstat.CO
keywords particle samplingWasserstein proximal operatorpreconditioningGibbs distributionscore approximationconvergence analysisCole-Hopf transformationBayesian inference
0
0 comments X

The pith

Preconditioning the regularized Wasserstein proximal operator enables efficient particle sampling from Gibbs distributions with step-size independent bias for quadratic potentials.

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

The paper introduces a preconditioned sampling method that evolves particles to sample from a Gibbs distribution by approximating the score function using a regularized Wasserstein proximal operator. This operator is derived through a Cole-Hopf transformation applied to coupled anisotropic heat equations, resulting in a kernel-based formulation that also resembles a modified self-attention mechanism. For quadratic potentials, the authors provide a non-asymptotic convergence analysis in discrete time and show that the bias depends on the regularization parameter but remains independent of the step size. This approach demonstrates improved acceleration and stability in experiments on log-concave and non-log-concave distributions, as well as in applications like image deconvolution and neural network training.

Core claim

The central claim is that by preconditioning the regularized Wasserstein proximal sampling and using the score from the proximal operator derived via Cole-Hopf transformation, one obtains a practical particle-based sampling algorithm with explicit bias characterization for quadratic cases that is independent of step-size.

What carries the argument

The preconditioned regularized Wasserstein proximal operator, derived by Cole-Hopf transformation on coupled anisotropic heat equations, which provides the score approximation and kernel formulation for the sampling dynamics.

If this is right

  • Convergence analysis shows non-asymptotic rates for quadratic potentials with bias controlled solely by regularization.
  • The method achieves acceleration and particle stability on toy examples and real applications like Bayesian image deconvolution.
  • Variable preconditioning matrices improve performance in non-convex settings such as Bayesian neural network training.
  • The diffusion component interprets as a modified self-attention block, potentially linking to transformer models.

Where Pith is reading between the lines

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

  • This framework might generalize the convergence results to non-quadratic potentials if similar bias independence can be shown.
  • Connecting the proximal operator to self-attention could inspire new hybrid models combining sampling and attention mechanisms.
  • Independent bias from step-size suggests robustness in choosing discretization parameters for practical implementations.

Load-bearing premise

The score function of the target distribution can be well approximated by the numerically tractable score of the regularized Wasserstein proximal operator obtained from the Cole-Hopf transformation.

What would settle it

Running the sampling algorithm on a quadratic potential with varying step sizes but fixed regularization and checking if the stationary distribution bias remains constant while changing regularization alters it.

Figures

Figures reproduced from arXiv: 2509.01685 by Hong Ye Tan, Stanley Osher, Wuchen Li.

Figure 1
Figure 1. Figure 1: Evolution of the preconditioned regularized Wasserstein proximal WProxI T =0.2,I for the 2- dimensional standard Gaussian, done with 5 particles and step-size of η = 0.1. The bandwidth is automatically determined by the regularization. As suggested by theory, the PRWPO of the particles approaches the standard Gaussian. 6. Experiments. In this section, we present various two-dimensional toy examples to demo… view at source ↗
Figure 2
Figure 2. Figure 2: Densities of the PRWPO WProxI 0.2I for the 2-dimensional standard Gaussian at iteration 100, done with n ∈ {3, 4, 5, 6} particles and a step-size of η = 0.1. We observe that the density of the Wasserstein proximal gradually becomes more spherical and Gaussian-like. 0 25 50 75 100 125 150 175 200 Iteration 10 1 10 0 10 1 KL( k, ) Bimodal KL distance ULA MALA MLA BRWP PBRWP [PITH_FULL_IMAGE:figures/full_fig… view at source ↗
Figure 3
Figure 3. Figure 3: Evolution of the KL divergence between baselines and BRWP-based methods for the bimodal distribution. Applied with 100 particles with initial distribution N (0, I), fixed step-size η = 0.1 and regularization parameter T = 0.05. We observe that the BRWP-based methods converge more smoothly even in later iterations, and the particles better approximate the target distribution. where e1 is the first standard … view at source ↗
Figure 4
Figure 4. Figure 4: Evolution of the various methods for the bimodal distribution at different iterations, with initial distribution N (0, I), which is contained between the moons. Evaluated with 100 particles, at iterations 2, 5 and 10 in the top, middle, and bottom rows respectively. Observe the stable behavior of BRWP and PBRWP, as opposed to the randomness of the Langevin methods. demonstrates that in this case, the preco… view at source ↗
Figure 5
Figure 5. Figure 5: Evolution of the KL divergence between baselines and BRWP-based methods for the bimodal distribution, initialized with large variance N (0, 6I). Applied with 100 particles, fixed step-size η = 0.1 and regularization parameter T = 0.05. In this case, the preconditioning accelerates the convergence. ULA MALA MLA BRWP PBRWP [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Evolution of the various methods for the bimodal distribution, with large initial variance N (0, 6I), which surrounds the moons. Evaluated with 100 particles, at iterations 2, 5 and 10 in the top to bottom rows respectively. We observe again the structure behavior at convergence of the noise-free BRWP and PBRWP methods. Moreover, the preconditioning affects the empirical covariance of the particles in diff… view at source ↗
Figure 7
Figure 7. Figure 7: Evolution of the various methods for the scaled annulus. Evaluated with 100 particles, at iterations 10, 50, and 200 from top to bottom respectively. We observe that PBRWP and MLA diffuse faster than their non-preconditioned counterparts. Moreover, PBRWP retains a similar level-set structure to BRWP. [18, 17]. This approximation has also been used in the other direction, e.g. [46] using integrals to approx… view at source ↗
Figure 8
Figure 8. Figure 8: Evolution of the KL divergence between baselines and BRWP-based methods for the scaled annulus. Applied with 100 particles, fixed step-size η = 0.1 and regularization parameter T = 0.05. We observe that the BRWP-based methods converge more smoothly. The resulting modified iteration using scaling and the Laplace approximation is as follows, where β = d −1/2 , X (k+1) = X (k) − η 2 M∇V (X (k) ) + η 2T  X (k… view at source ↗
Figure 9
Figure 9. Figure 9: Evolution of the high-dimensional modifications for the 50-dimensional Gaussian, at convergence in 1000 iterations. Evaluated with 50 particles, step-size η = 0.1 and regularizations T = 0.02, 0.2, 0.9 in the top, middle and bottom rows respectively. We observe little difference when using the Laplace approximation compared to the ground-truth, suggesting this is reasonable. The β = d −1/2 scaling increase… view at source ↗
Figure 10
Figure 10. Figure 10: Standard deviations for TV-regularized deconvolution. Run with 40 particles, and evaluated at iterations 20, 200 and 2000 in the top, middle and bottom rows respectively. We observe that the pixelwise variance of the noise-free methods BRWP and PBRWP are lower than their Langevin counterparts. In the 200 iteration regime, we see that the features of the noise-free BRWP and PBRWP methods have more constras… view at source ↗
Figure 11
Figure 11. Figure 11: Plot of the norm of the average of the points ∥ 1 N P i xi∥ with respect to iteration, for the target distribution N (0, I2) and identity preconditioner, T = 2, and various choices of N and step-size. The mean does not converge linearly, and instead has two separate regimes. where for a positive definite symmetric matrix M ∈ R d×d , we define the scaled (2, M) norm of a d × N matrix by (C.14) ∥A∥ 2 2,M :=… view at source ↗
read the original abstract

We consider sampling from a Gibbs distribution by evolving finitely many particles. We propose a preconditioned version of a recently proposed noise-free sampling method, governed by approximating the score function with the numerically tractable score of a regularized Wasserstein proximal operator. This is derived by a Cole--Hopf transformation on coupled anisotropic heat equations, yielding a kernel formulation for the preconditioned regularized Wasserstein proximal. The diffusion component of the proposed method is also interpreted as a modified self-attention block, as in transformer architectures. For quadratic potentials, we provide a discrete-time non-asymptotic convergence analysis and explicitly characterize the bias, which is dependent on regularization and independent of step-size. Experiments demonstrate acceleration and particle-level stability on various log-concave and non-log-concave toy examples to Bayesian total-variation regularized image deconvolution, and competitive/better performance on non-convex Bayesian neural network training when utilizing variable preconditioning matrices.

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 / 2 minor

Summary. The paper proposes a preconditioned regularized Wasserstein proximal sampling method for drawing particles from a Gibbs distribution. The score is approximated via the numerically tractable score of a regularized Wasserstein proximal operator, obtained through a Cole-Hopf transformation applied to coupled anisotropic heat equations that yields an explicit kernel formulation. The diffusion step is interpreted as a modified self-attention mechanism. For quadratic potentials the authors supply a discrete-time non-asymptotic convergence analysis together with an explicit bias characterization that depends on the regularization parameter but is claimed to be independent of step-size. Experiments on log-concave and non-log-concave toy problems, Bayesian total-variation image deconvolution, and non-convex Bayesian neural-network training are reported to demonstrate acceleration and particle stability.

Significance. If the non-asymptotic analysis and the step-size-independent bias result for quadratic potentials are rigorously established, the work would provide a concrete advance in noise-free particle-based sampling with explicit error controls and a novel link to attention mechanisms. The explicit bias formula tied only to regularization is a clear strength that could guide practical tuning; the reported experiments on both synthetic and applied tasks add empirical support.

major comments (1)
  1. [Quadratic potentials analysis] Quadratic-potential analysis (discrete-time non-asymptotic convergence section): the central claim that the bias is independent of step-size h rests on an unverified cancellation inside the one-step proximal map. Standard Euler or proximal discretizations of score-based dynamics produce O(h) or O(h^2) bias; the paper must supply an explicit expansion of the update for a quadratic V (or an equivalent algebraic verification) to confirm that the kernel obtained from the Cole-Hopf transformation exactly cancels the leading h-dependent term. Without this verification the independence statement is not yet load-bearing.
minor comments (2)
  1. [Abstract] The abstract states that the bias 'is dependent on regularization and independent of step-size' but provides no quantitative error bounds or statement of the precise norm in which convergence is measured; adding one sentence with the leading-order bias expression would improve clarity.
  2. [Method and analysis sections] Notation for the preconditioning matrix and the regularization parameter should be introduced once in a dedicated 'Notation' paragraph rather than scattered across the method and analysis sections.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the thorough review and constructive feedback on our work. The single major comment raises a valid point about the need for explicit verification in the quadratic-potential analysis. We address it directly below and will revise the manuscript to incorporate the requested algebraic details.

read point-by-point responses

  1. Referee: [Quadratic potentials analysis] Quadratic-potential analysis (discrete-time non-asymptotic convergence section): the central claim that the bias is independent of step-size h rests on an unverified cancellation inside the one-step proximal map. Standard Euler or proximal discretizations of score-based dynamics produce O(h) or O(h^2) bias; the paper must supply an explicit expansion of the update for a quadratic V (or an equivalent algebraic verification) to confirm that the kernel obtained from the Cole-Hopf transformation exactly cancels the leading h-dependent term. Without this verification the independence statement is not yet load-bearing.

    Authors: We agree that an explicit algebraic verification is required to make the step-size independence rigorous. The current manuscript derives the kernel via the Cole-Hopf transformation on the coupled anisotropic heat equations and states that the resulting bias depends only on the regularization parameter, but it does not include a term-by-term expansion of the one-step update for quadratic V. In the revision we will add this expansion in the discrete-time non-asymptotic convergence section. Specifically, we will substitute the quadratic potential into the proximal map, apply the explicit kernel formula, and show that all O(h) and O(h^2) bias contributions cancel identically, leaving a remainder that depends solely on the regularization strength. This verification will be presented as a self-contained lemma with the algebraic steps spelled out. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper derives the preconditioned regularized Wasserstein proximal operator via Cole-Hopf transformation applied to coupled anisotropic heat equations, yielding an explicit kernel formulation. This PDE step is a standard external technique with independent mathematical grounding and does not reduce to a self-definition or fitted input. The discrete-time non-asymptotic convergence analysis for quadratic potentials explicitly characterizes bias as a function of the regularization parameter (independent of step-size h), without evidence that the bias term is obtained by construction from the discretization or from a self-citation chain. No load-bearing uniqueness theorem, ansatz smuggling, or renaming of known results is present; the central claims remain self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on a regularization parameter that controls bias and an assumption that the proximal score approximates the true score well enough for effective sampling.

free parameters (1)
  • regularization parameter
    Controls the tractability of the proximal operator and determines the bias in the sampling dynamics.
axioms (1)
  • domain assumption The score function can be approximated by the score of a regularized Wasserstein proximal operator.
    This approximation governs the particle evolution and is derived from the Cole-Hopf transformation.

pith-pipeline@v0.9.0 · 5691 in / 1225 out tokens · 55687 ms · 2026-05-21T23:31:11.567637+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

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

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

51 extracted references · 51 canonical work pages · 1 internal anchor

  1. [1]

    M. V. Afonso, J. M. Bioucas-Dias, and M. A. Figueiredo, Fast image recovery using variable splitting and constrained optimization, IEEE transactions on image processing, 19 (2010), pp. 2345–2356

    work page 2010

  2. [2]

    Benamou and Y

    J.-D. Benamou and Y. Brenier , A computational fluid mechanics solution to the Monge-Kantorovich mass transfer problem, Numerische Mathematik, 84 (2000), pp. 375–393

    work page 2000

  3. [3]

    Bleistein and R

    N. Bleistein and R. A. Handelsman , Asymptotic expansions of integrals , Ardent Media, 1975

    work page 1975

  4. [4]

    Bond-Taylor, A

    S. Bond-Taylor, A. Leach, Y. Long, and C. G. Willcocks, Deep generative modelling: A comparative review of VAEs, GANs, normalizing flows, energy-based and autoregressive models , IEEE transactions on pattern analysis and machine intelligence, (2021)

    work page 2021

  5. [5]

    Z. I. Botev, J. F. Grotowski, and D. P. Kroese , Kernel density estimation via diffusion , Annals of Statistics, 38 (2010), pp. 2916–2957

    work page 2010

  6. [6]

    J. A. Carrillo, K. Craig, and F. S. Patacchini , A blob method for diffusion , Calculus of Variations and Partial Differential Equations, 58 (2019), pp. 1–53

    work page 2019

  7. [7]

    A unified perspective on the dynamics of deep transformers.arXiv preprint arXiv:2501.18322,

    V. Castin, P. Ablin, J. A. Carrillo, and G. Peyr ´e, A unified perspective on the dynamics of deep 26 H. Y. TAN, S. OSHER, AND W. LI transformers, arXiv preprint arXiv:2501.18322, (2025)

    work page arXiv 2025

  8. [8]

    R. T. Chen, Y. Rubanova, J. Bettencourt, and D. K. Duvenaud , Neural ordinary differential equations, Advances in neural information processing systems, 31 (2018)

    work page 2018

  9. [9]

    L. Condat, A primal–dual splitting method for convex optimization involving Lipschitzian, proximable and linear composite terms , Journal of optimization theory and applications, 158 (2013), pp. 460–479

    work page 2013

  10. [10]

    Durmus and ´E

    A. Durmus and ´E. Moulines, High-dimensional Bayesian inference via the unadjusted Langevin algorithm, Bernoulli, 25 (2019), pp. 2854–2882

    work page 2019

  11. [11]

    Durmus, E

    A. Durmus, E. Moulines, and M. Pereyra , Efficient Bayesian computation by proximal Markov chain Monte Carlo: when Langevin meets Moreau , SIAM Journal on Imaging Sciences, 11 (2018), pp. 473–506

    work page 2018

  12. [12]

    Gelman, J

    A. Gelman, J. B. Carlin, H. S. Stern, and D. B. Rubin , Bayesian data analysis , Chapman and Hall/CRC, 1995

    work page 1995

  13. [13]

    Geshkovski, C

    B. Geshkovski, C. Letrouit, Y. Polyanskiy, and P. Rigollet , A mathematical perspective on transformers, Bulletin of the American Mathematical Society, 62 (2025), pp. 427–479

    work page 2025

  14. [14]

    Gramacki, Nonparametric kernel density estimation and its computational aspects , vol

    A. Gramacki, Nonparametric kernel density estimation and its computational aspects , vol. 37, Springer, 2018

    work page 2018

  15. [15]

    Habring, A

    A. Habring, A. Falk, and T. Pock , Diffusion at absolute zero: Langevin sampling using successive Moreau envelopes, in 2025 IEEE Statistical Signal Processing Workshop (SSP), IEEE, 2025, pp. 61–65

    work page 2025

  16. [16]

    Habring, M

    A. Habring, M. Holler, and T. Pock , Subgradient langevin methods for sampling from nonsmooth potentials, SIAM Journal on Mathematics of Data Science, 6 (2024), pp. 897–925

    work page 2024

  17. [17]

    F. Han, S. Osher, and W. Li, Convergence of noise-free sampling algorithms with regularized Wasserstein proximals, arXiv preprint arXiv:2409.01567, (2024)

    work page arXiv 2024

  18. [18]

    F. Han, S. Osher, and W. Li , Splitting regularized Wasserstein proximal algorithms for nonsmooth sampling problems, arXiv preprint arXiv:2502.16773, (2025)

    work page arXiv 2025

  19. [19]

    F. Han, S. Osher, and W. Li , Tensor train based sampling algorithms for approximating regularized Wasserstein proximal operators, SIAM/ASA Journal on Uncertainty Quantification, 13 (2025), pp. 775– 804

    work page 2025

  20. [20]

    J. R. Hershey and P. A. Olsen, Approximating the Kullback Leibler divergence between Gaussian mixture models, in 2007 IEEE International Conference on Acoustics, Speech and Signal Processing-ICASSP’07, vol. 4, IEEE, 2007, pp. IV–317

    work page 2007

  21. [21]

    Hsieh, A

    Y.-P. Hsieh, A. Kavis, P. Rolland, and V. Cevher , Mirrored Langevin dynamics, Advances in Neural Information Processing Systems, 31 (2018)

    work page 2018

  22. [22]

    Hwang, Fadam: Adam is a natural gradient optimizer using diagonal empirical Fisher information , arXiv preprint arXiv:2405.12807, (2024)

    D. Hwang, Fadam: Adam is a natural gradient optimizer using diagonal empirical Fisher information , arXiv preprint arXiv:2405.12807, (2024)

    work page arXiv 2024

  23. [23]

    Jiang, Mirror Langevin Monte Carlo: the case under isoperimetry , Advances in Neural Information Processing Systems, 34 (2021), pp

    Q. Jiang, Mirror Langevin Monte Carlo: the case under isoperimetry , Advances in Neural Information Processing Systems, 34 (2021), pp. 715–725

    work page 2021

  24. [24]

    D. P. Kingma and J. Ba , Adam: A method for stochastic optimization , arXiv preprint arXiv:1412.6980, (2014)

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  25. [25]

    Krauth, Statistical mechanics: algorithms and computations , vol

    W. Krauth, Statistical mechanics: algorithms and computations , vol. 13, OUP Oxford, 2006

    work page 2006

  26. [26]

    Kubo, Stochastic Liouville equations , Journal of Mathematical Physics, 4 (1963), pp

    R. Kubo, Stochastic Liouville equations , Journal of Mathematical Physics, 4 (1963), pp. 174–183

    work page 1963

  27. [27]

    Kunstner, P

    F. Kunstner, P. Hennig, and L. Balles , Limitations of the empirical fisher approximation for natural gradient descent, Advances in neural information processing systems, 32 (2019)

    work page 2019

  28. [28]

    Laumont, V

    R. Laumont, V. D. Bortoli, A. Almansa, J. Delon, A. Durmus, and M. Pereyra , Bayesian imaging using plug & play priors: when Langevin meets Tweedie , SIAM Journal on Imaging Sciences, 15 (2022), pp. 701–737

    work page 2022

  29. [29]

    W. Li, S. Liu, and S. Osher , A kernel formula for regularized Wasserstein proximal operators , Research in the Mathematical Sciences, 10 (2023), p. 43

    work page 2023

  30. [30]

    W. Li, W. Liu, J. Chen, L. Wu, P. D. Flynn, W. Ding, and P. Chen , Reducing mode collapse with Monge–Kantorovich optimal transport for generative adversarial networks , IEEE Transactions on Cybernetics, (2023)

    work page 2023

  31. [31]

    D. J. MacKay, Bayesian neural networks and density networks , Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 354 (1995), pp. 73–80

    work page 1995

  32. [32]

    Nijkamp, R

    E. Nijkamp, R. Gao, P. Sountsov, S. Vasudevan, B. Pang, S.-C. Zhu, and Y. N. Wu , MCMC PRECONDITIONED REGULARIZED WASSERSTEIN PROXIMAL SAMPLING METHODS 27 should mix: learning energy-based model with neural transport latent space MCMC. , in International Conference on Learning Representations (ICLR 2022)., 2022

    work page 2022

  33. [33]

    K. B. Petersen, M. S. Pedersen, et al. , The matrix cookbook, Technical University of Denmark, 7 (2008), p. 510

    work page 2008

  34. [34]

    Risken, Fokker-Planck equation, in The Fokker-Planck equation: methods of solution and applications, Springer, 1989, pp

    H. Risken, Fokker-Planck equation, in The Fokker-Planck equation: methods of solution and applications, Springer, 1989, pp. 63–95

    work page 1989

  35. [35]

    G. O. Roberts and R. L. Tweedie, Exponential convergence of Langevin distributions and their discrete approximations, Bernoulli, (1996), pp. 341–363

    work page 1996

  36. [36]

    P. J. Rossky, J. D. Doll, and H. L. Friedman , Brownian dynamics as smart Monte Carlo simulation , The Journal of Chemical Physics, 69 (1978), pp. 4628–4633

    work page 1978

  37. [37]

    L. I. Rudin, S. Osher, and E. Fatemi, Nonlinear total variation based noise removal algorithms , Physica D: nonlinear phenomena, 60 (1992), pp. 259–268

    work page 1992

  38. [38]

    M. E. Sander, P. Ablin, M. Blondel, and G. Peyr ´e, Sinkformers: Transformers with doubly stochastic attention, in International Conference on Artificial Intelligence and Statistics, PMLR, 2022, pp. 3515–3530

    work page 2022

  39. [39]

    Santambrogio, Optimal transport for applied mathematicians , vol

    F. Santambrogio, Optimal transport for applied mathematicians , vol. 87, Springer, 2015

    work page 2015

  40. [40]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole , Score-based generative modeling through stochastic differential equations , in International Conference on Learning Representations, 2021

    work page 2021

  41. [41]

    Srivastava, L

    A. Srivastava, L. Valkov, C. Russell, M. U. Gutmann, and C. Sutton, Veegan: Reducing mode collapse in GANs using implicit variational learning , Advances in neural information processing systems, 30 (2017)

    work page 2017

  42. [42]

    A. M. Stuart, Inverse problems: a Bayesian perspective , Acta numerica, 19 (2010), pp. 451–559

    work page 2010

  43. [43]

    Tachella, M

    J. Tachella, M. Terris, S. Hurault, A. Wang, D. Chen, M.-H. Nguyen, M. Song, T. Davies, L. Davy, J. Dong, et al. , Deepinverse: A python package for solving imaging inverse problems with deep learning, arXiv preprint arXiv:2505.20160, (2025)

    work page arXiv 2025

  44. [44]

    H. Y. Tan, Z. Cai, M. Pereyra, S. Mukherjee, J. Tang, and C.-B. Sch ¨onlieb, Unsupervised training of convex regularizers using maximum likelihood estimation , Transactions on Machine Learning Research, (2024)

    work page 2024

  45. [45]

    H. Y. Tan, S. Osher, and W. Li , Noise-free sampling algorithms via regularized Wasserstein proximals , Research in the Mathematical Sciences, 11 (2024), p. 65

    work page 2024

  46. [46]

    R. J. Tibshirani, S. W. Fung, H. Heaton, and S. Osher, Laplace meets Moreau: Smooth approximation to infimal convolutions using Laplace’s method , Journal of Machine Learning Research, 26 (2025), pp. 1–36

    work page 2025

  47. [47]

    Van Kerm, Adaptive kernel density estimation , The Stata Journal, 3 (2003), pp

    P. Van Kerm, Adaptive kernel density estimation , The Stata Journal, 3 (2003), pp. 148–156

    work page 2003

  48. [48]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, Attention is all you need , Advances in neural information processing systems, 30 (2017)

    work page 2017

  49. [49]

    M. P. Wand and M. C. Jones , Kernel smoothing, CRC press, 1994

    work page 1994

  50. [50]

    log |˜Σk| |˜Σk+1| + Tr ˜Σ−1 ∞ (˜Σk+1 − ˜Σk) # = 1 2

    Y. Wang and W. Li, Accelerated information gradient flow, Journal of Scientific Computing, 90 (2022), pp. 1–47. 28 H. Y. TAN, S. OSHER, AND W. LI Appendix A. Proofs. A.1. Anisotropic Green’s function. Recall (3.1): (A.1) Gt,M(x, y) := 1 (4πβ −1t)d/2|M |1/2 e−β (x−y)⊤M −1(x−y) 4t . Proposition A.1. The anisotropic kernel Gt,M(x, y) is a Green’s function fo...

    work page 2022

  51. [51]

    // bias correction 6 M (k) = diag(1/(√ˆvk + ϵ)) ; // construct preconditioning matrix 7 end 8 return (M (k))K k=1

    work page