pith. sign in

arxiv: 2606.21907 · v1 · pith:BCMBNLGQnew · submitted 2026-06-20 · 📊 stat.ME · stat.ML

Nonparametric Deconvolution and Denoising using Simulation Based Inference

Ritwik Vashistha , Abhra Sarkar , Arya Farahi This is my paper

Pith reviewed 2026-06-26 12:03 UTC · model grok-4.3

classification 📊 stat.ME stat.ML
keywords nonparametric deconvolutiondenoisingsimulation-based inferenceconvolutional MMDadditive measurement errorempirical Bayeslikelihood-free inferenceSobolev smoothness
0
0 comments X

The pith

A likelihood-free framework uses convolutional MMD to recover latent densities and denoise signals from additive noise via simulation.

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

The paper establishes a simulation-based method for nonparametric deconvolution and empirical Bayes denoising when latent signals are obscured by additive measurement error with known distribution. It optimizes a generative model of the latent density by minimizing a convolutional maximum mean discrepancy between the observed data and the noise-convolved model output, supporting flexible models such as mixtures and flows. Theoretical results extend the approach to nonparametric settings with finite-sample bounds and L2 convergence rates that follow the classical dependence on noise smoothness. A sympathetic reader would care because this handles cases where likelihoods are intractable yet simulation of the noise process is feasible, addressing a frequent challenge in scientific data with measurement error.

Core claim

The authors claim that a convolutional maximum mean discrepancy loss yields a differentiable, simulation-based objective for learning a latent generative model by matching the observed data distribution to the noise-convolved model distribution. This framework performs nonparametric density deconvolution and empirical Bayes denoising for homoscedastic or heteroscedastic multivariate noise. They extend convMMD to nonparametric estimation, proving finite-sample bounds for empirical sieve minimizers and L2 convergence rates under Sobolev smoothness that recover polynomial rates for ordinary-smooth noise and logarithmic rates for super-smooth noise.

What carries the argument

The convMMD loss, which measures discrepancy between the empirical distribution of noisy observations and the distribution of a candidate latent density after convolution with the known noise.

If this is right

  • The learned density serves as an empirical prior for posterior denoising of individual latent values.
  • The approach is compatible with sieve classes such as Gaussian mixtures and normalizing flows.
  • It applies to both homoscedastic and heteroscedastic multivariate noise.
  • Convergence rates are polynomial for ordinary-smooth noise and logarithmic for super-smooth noise.

Where Pith is reading between the lines

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

  • The simulation requirement could support extensions to settings where the noise distribution is only approximately known but still simulatable.
  • Integration with high-capacity deep generative models might enable scaling to higher-dimensional latent spaces beyond the sieve classes considered.
  • The framework's reliance on convolution suggests it may inform other inverse problems where forward simulation is easier than likelihood evaluation.

Load-bearing premise

The measurement noise must be additive with a known distribution that can be accurately simulated or convolved with the latent model.

What would settle it

A Monte Carlo experiment in which the estimated latent density fails to converge in L2 norm to the true density at the predicted rate as the number of observations increases, when the latent density satisfies the assumed Sobolev smoothness and the noise is ordinary-smooth.

Figures

Figures reproduced from arXiv: 2606.21907 by Abhra Sarkar, Arya Farahi, Ritwik Vashistha.

Figure 1
Figure 1. Figure 1: Rate validation. L2 error follows the ordinary-smooth and supersmooth rates predicted by Theorem 3.6, while Latent MMD2 follows the parametric N −1 rate predicted by Corollary 3.4 The right panels validate Corollary 3.4. In both the ordinary-smooth and supersmooth settings, the latent squared MMD error, MMD2 (p, qθbN ), decays at the parametric N −1 rate, indicating that convMMD achieves fast convergence i… view at source ↗
Figure 2
Figure 2. Figure 2: Benchmarking denoising. The last column assumes the true density is perfectly known. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative Comparison of Estimated Prior for d = 3 across different methods (columns) and different noise [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Image denoising performance. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
read the original abstract

Latent signals are often obscured by measurement noise, yet encode the underlying laws and dynamics of complex systems; learning both the signals and their distributions remains a central challenge in scientific inference. The noise is often non-negligible, and the likelihoods for expressive generative models are often intractable. We utilize a convolutional maximum mean discrepancy (convMMD) loss and propose a likelihood-free framework for nonparametric density deconvolution and empirical Bayes denoising under additive measurement error. Our method learns a latent generative model by matching the observed data distribution to the noise-convolved model distribution. This yields a differentiable, simulation-based objective for multivariate homoscedastic or heteroscedastic noise, compatible with expressive sieve classes such as Gaussian mixtures and normalizing flows. The learned density then serves as an empirical prior for posterior denoising of individual latent values. Theoretically, we extend convMMD from parametric to nonparametric estimation, proving finite-sample bounds for empirical sieve minimizers and $L_2$ convergence rates under Sobolev smoothness. These rates recover the classical inverse-problem dependence: polynomial for ordinary-smooth and logarithmic for super-smooth noises. Our method provides a practical, theoretically grounded approach to deconvolution and denoising under generative latent distribution models.

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

2 major / 1 minor

Summary. The paper proposes a likelihood-free nonparametric framework for density deconvolution and empirical Bayes denoising under additive measurement error. It employs a convolutional maximum mean discrepancy (convMMD) loss to match the observed data distribution to the noise-convolved latent generative model distribution, supporting sieve classes such as Gaussian mixtures and normalizing flows. The learned density serves as an empirical prior for posterior denoising. Theoretically, the work extends convMMD to nonparametric estimation by proving finite-sample bounds for empirical sieve minimizers and L2 convergence rates under Sobolev smoothness, with rates that recover the classical inverse-problem dependence (polynomial for ordinary-smooth noise, logarithmic for super-smooth noise).

Significance. If the theoretical results hold, the contribution is significant as it provides a simulation-based approach to nonparametric inverse problems with explicit finite-sample and rate guarantees that align with known minimax behavior in deconvolution. The method's compatibility with expressive generative models and heteroscedastic noise, combined with the likelihood-free objective, addresses a practical gap in scientific inference where likelihoods are intractable.

major comments (2)
  1. [Abstract] Abstract: The central claim that finite-sample bounds for empirical sieve minimizers and L2 convergence rates under Sobolev smoothness are proved, recovering classical inverse-problem rates, is load-bearing. However, the abstract provides no statement of the required assumptions on the sieve class, the kernel, or the noise convolution operator, preventing assessment of whether the extension from parametric to nonparametric convMMD is valid.
  2. [Method description] Method description (as summarized): The simulation-based objective requires that the additive noise distribution is known and can be accurately simulated or convolved with the latent model. This assumption is load-bearing for consistency of the density estimator; the manuscript should explicitly address robustness when the noise is only approximately known, as this directly impacts the claimed convergence rates.
minor comments (1)
  1. [Notation] Clarify notation for the multivariate convMMD kernel and how the convolution is implemented in the objective for heteroscedastic noise.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment below and indicate the revisions we will make to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that finite-sample bounds for empirical sieve minimizers and L2 convergence rates under Sobolev smoothness are proved, recovering classical inverse-problem rates, is load-bearing. However, the abstract provides no statement of the required assumptions on the sieve class, the kernel, or the noise convolution operator, preventing assessment of whether the extension from parametric to nonparametric convMMD is valid.

    Authors: We agree that the abstract would benefit from a brief statement of the key assumptions to clarify the scope of the theoretical claims. The full paper states these in the theorem statements and proofs (characteristic kernel, sieve with sufficient approximation capacity for the Sobolev ball, and standard smoothness conditions on the convolution operator for ordinary- and super-smooth noise). We will revise the abstract to note that the results hold under these standard assumptions on the kernel, sieve, and noise operator. revision: yes

  2. Referee: [Method description] Method description (as summarized): The simulation-based objective requires that the additive noise distribution is known and can be accurately simulated or convolved with the latent model. This assumption is load-bearing for consistency of the density estimator; the manuscript should explicitly address robustness when the noise is only approximately known, as this directly impacts the claimed convergence rates.

    Authors: The method assumes the noise distribution is known and simulatable, which is standard for simulation-based deconvolution to ensure the convMMD objective is well-defined. We will add a dedicated remark or subsection discussing robustness to approximate noise knowledge, including how misspecification affects consistency and rates, along with practical guidance on sensitivity analysis. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The derivation relies on a simulation-based convMMD objective that explicitly convolves or simulates from a known additive noise distribution to match the observed data distribution, then applies standard sieve estimation theory to obtain finite-sample bounds and Sobolev L2 rates. These steps are conditioned on external, simulatable inputs and classical inverse-problem analysis rather than any self-referential fit, prediction-by-construction, or load-bearing self-citation. The central claims remain independent of the target results.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the additive noise model with known distribution and the ability to simulate from expressive sieve classes; these are standard domain assumptions rather than new inventions.

free parameters (1)
  • sieve complexity
    The dimension or number of components in the sieve class (e.g., Gaussian mixture components or flow architecture) must be chosen to control approximation error.
axioms (1)
  • domain assumption Additive measurement error with known noise distribution
    The observed data is modeled as the convolution of the latent distribution with the noise; invoked throughout the method description.

pith-pipeline@v0.9.1-grok · 5746 in / 1345 out tokens · 36696 ms · 2026-06-26T12:03:50.127782+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

66 extracted references · 2 linked inside Pith

  1. [1]

    Chapman and Hall/CRC, 2006

    Raymond J Carroll, David Ruppert, Leonard A Stefanski, and Ciprian M Crainiceanu.Measurement error in nonlinear models: a modern perspective. Chapman and Hall/CRC, 2006

    2006

  2. [2]

    Deconvolution in astronomy: A review.Publications of the Astronomical Society of the Pacific, 114:1051–1069, 2002

    Jean-Luc Starck, Eric Pantin, and Fionn Murtagh. Deconvolution in astronomy: A review.Publications of the Astronomical Society of the Pacific, 114:1051–1069, 2002

    2002

  3. [3]

    Visal Sok, Adam Muzzin, Pascale Jablonka, Z Cemile Marsan, Vivian YY Tan, Leo Alcorn, Danilo Marchesini, and Mauro Stefanon. Finite-resolution deconvolution of multiwavelength imaging of 20,000 galaxies in the cosmos field: The evolution of clumpy galaxies over cosmic time.The Astrophysical Journal, 924:7, 2022

    2022

  4. [4]

    The decade cosmic shear project i: A new weak lensing shape catalog of 107 million galaxies

    D Anbajagane, C Chang, Z Zhang, CY Tan, M Adamow, LF Secco, MR Becker, PS Ferguson, A Drlica-Wagner, RA Gruendl, et al. The decade cosmic shear project i: A new weak lensing shape catalog of 107 million galaxies. arXiv preprint arXiv:2502.17674, 2025

    arXiv 2025

  5. [5]

    Weak lensing shear calibration with simulations of the hsc survey.Monthly Notices of the Royal Astronomical Society, 481:3170–3195, 2018

    Rachel Mandelbaum, François Lanusse, Alexie Leauthaud, Robert Armstrong, Melanie Simet, Hironao Miyatake, Joshua E Meyers, James Bosch, Ryoma Murata, Satoshi Miyazaki, et al. Weak lensing shear calibration with simulations of the hsc survey.Monthly Notices of the Royal Astronomical Society, 481:3170–3195, 2018

    2018

  6. [6]

    Eddington’s demon: inferring galaxy mass functions and other distributions from uncertain data.Monthly Notices of the Royal Astronomical Society, 474:5500–5522, 2018

    Danail Obreschkow, Steven G Murray, Aaron SG Robotham, and Tobias Westmeier. Eddington’s demon: inferring galaxy mass functions and other distributions from uncertain data.Monthly Notices of the Royal Astronomical Society, 474:5500–5522, 2018

    2018

  7. [7]

    Measurement error and the replication crisis.Science, 355:584–585, 2017

    Eric Loken and Andrew Gelman. Measurement error and the replication crisis.Science, 355:584–585, 2017

    2017

  8. [8]

    Bayesian-based iterative method of image restoration.Journal of the Optical Society of America, 62:55–59, 1972

    William Hadley Richardson. Bayesian-based iterative method of image restoration.Journal of the Optical Society of America, 62:55–59, 1972

    1972

  9. [9]

    An iterative technique for the rectification of observed distributions.Astronomical Journal, 79:745, 1974

    Leon B Lucy. An iterative technique for the rectification of observed distributions.Astronomical Journal, 79:745, 1974

    1974

  10. [10]

    A deep generative deconvolutional image model

    Yunchen Pu, Win Yuan, Andrew Stevens, Chunyuan Li, and Lawrence Carin. A deep generative deconvolutional image model. InArtificial Intelligence and Statistics, pages 741–750. PMLR, 2016

    2016

  11. [11]

    Digital image reconstruction: Deblurring and denoising.Annual Review of Astronomy and Astrophysics, 43:139–194, 2005

    RC Puetter, TR Gosnell, and Amos Yahil. Digital image reconstruction: Deblurring and denoising.Annual Review of Astronomy and Astrophysics, 43:139–194, 2005

    2005

  12. [12]

    Density deconvolution with normalizing flows

    Tim Dockhorn, James A Ritchie, Yaoliang Yu, and Iain Murray. Density deconvolution with normalizing flows. arXiv preprint arXiv:2006.09396, 2020

    arXiv 2006

  13. [13]

    Machines learn to infer stellar parameters just by looking at a large number of spectra.Monthly Notices of the Royal Astronomical Society, 501(4):6026–6041, 2021

    Nima Sedaghat, Martino Romaniello, Jonathan E Carrick, and François-Xavier Pineau. Machines learn to infer stellar parameters just by looking at a large number of spectra.Monthly Notices of the Royal Astronomical Society, 501(4):6026–6041, 2021

    2021

  14. [14]

    Data denoising with transfer learning in single-cell transcriptomics.Nature methods, 16(9):875–878, 2019

    Jingshu Wang, Divyansh Agarwal, Mo Huang, Gang Hu, Zilu Zhou, Chengzhong Ye, and Nancy R Zhang. Data denoising with transfer learning in single-cell transcriptomics.Nature methods, 16(9):875–878, 2019

    2019

  15. [15]

    Hyperspectral image denoising: From model-driven, data-driven, to model-data-driven.IEEE Transactions on Neural Networks and Learning Systems, 35(10):13143–13163, 2023

    Qiang Zhang, Yaming Zheng, Qiangqiang Yuan, Meiping Song, Haoyang Yu, and Yi Xiao. Hyperspectral image denoising: From model-driven, data-driven, to model-data-driven.IEEE Transactions on Neural Networks and Learning Systems, 35(10):13143–13163, 2023

    2023

  16. [16]

    Optimal rates of convergence for deconvolving a density.Journal of the American Statistical Association, 83:1184–1186, 1988

    Raymond J Carroll and Peter Hall. Optimal rates of convergence for deconvolving a density.Journal of the American Statistical Association, 83:1184–1186, 1988

    1988

  17. [17]

    A consistent nonparametric density estimator for the deconvolution problem

    Ming Chung Liu and Robert L Taylor. A consistent nonparametric density estimator for the deconvolution problem. Canadian Journal of Statistics, 17:427–438, 1989

    1989

  18. [18]

    Deconvolving kernel density estimators.Statistics, 21:169–184, 1990

    Leonard A Stefanski and Raymond J Carroll. Deconvolving kernel density estimators.Statistics, 21:169–184, 1990

    1990

  19. [19]

    On the optimal rates of convergence for nonparametric deconvolution problems.Annals of Statistics, 19:1257–1272, 1991

    Jianqing Fan. On the optimal rates of convergence for nonparametric deconvolution problems.Annals of Statistics, 19:1257–1272, 1991

    1991

  20. [20]

    Density estimation with heteroscedastic error.Bernoulli, 14:562–579, 2008

    Aurore Delaigle and Alexander Meister. Density estimation with heteroscedastic error.Bernoulli, 14:562–579, 2008

    2008

  21. [21]

    Discrete-transform approach to deconvolution problems.Biometrika, 92:135–148, 2005

    Peter Hall and Peihua Qiu. Discrete-transform approach to deconvolution problems.Biometrika, 92:135–148, 2005

    2005

  22. [22]

    Variational inference with normalizing flows

    Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. InInternational Conference on Machine Learning, pages 1530–1538. PMLR, 2015. 11 Nonparametric Deconvolution and Denoising

    2015

  23. [23]

    Glow: generative flow with invertible 1 × 1 convolutions

    Diederik P Kingma and Prafulla Dhariwal. Glow: generative flow with invertible 1 × 1 convolutions. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 10236– 10245, 2018

    2018

  24. [24]

    Normalizing flows for probabilistic modeling and inference.Journal of Machine Learning Research, 22:1–64, 2021

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshminarayanan. Normalizing flows for probabilistic modeling and inference.Journal of Machine Learning Research, 22:1–64, 2021

    2021

  25. [25]

    Stochastic forward–backward deconvolution: Training diffusion models with finite noisy datasets

    Haoye Lu, Qifan Wu, and Yaoliang Yu. Stochastic forward–backward deconvolution: Training diffusion models with finite noisy datasets. InInternational Conference on Machine Learning, pages 40741–40768. PMLR, 2025

    2025

  26. [26]

    Multivariate, heteroscedastic empirical Bayes via nonparametric maximum likelihood.Journal of the Royal Statistical Society Series B: Statistical Methodology, 87:1–32, 2025

    Jake A Soloff, Adityanand Guntuboyina, and Bodhisattva Sen. Multivariate, heteroscedastic empirical Bayes via nonparametric maximum likelihood.Journal of the Royal Statistical Society Series B: Statistical Methodology, 87:1–32, 2025

    2025

  27. [27]

    Extreme deconvolution: Inferring complete distribution functions from noisy, heterogeneous and incomplete observations.Annals of Applied Statistics, 5:1657–1677, 2011

    Jo Bovy, David W Hogg, and Sam T Roweis. Extreme deconvolution: Inferring complete distribution functions from noisy, heterogeneous and incomplete observations.Annals of Applied Statistics, 5:1657–1677, 2011

    2011

  28. [28]

    Convolutional maximum mean discrepancy for inference in noisy data.arXiv preprint arXiv:2604.12022, 2026

    Ritwik Vashistha, Jeff M Phillips, Abhra Sarkar, and Arya Farahi. Convolutional maximum mean discrepancy for inference in noisy data.arXiv preprint arXiv:2604.12022, 2026

    Pith/arXiv arXiv 2026

  29. [29]

    An empirical bayes approach to statistics

    Herbert E Robbins. An empirical bayes approach to statistics. InBreakthroughs in Statistics: Foundations and Basic Theory, pages 388–394. Springer

  30. [30]

    Bayes, oracle bayes and empirical bayes.Statistical Science, 34:177–201, 2019

    Bradley Efron. Bayes, oracle bayes and empirical bayes.Statistical Science, 34:177–201, 2019

    2019

  31. [31]

    Borgwardt, Malte J

    Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Schölkopf, and Alexander Smola. A kernel two-sample test.Journal of Machine Learning Research, 13:723–773, 2012

    2012

  32. [32]

    Hilbert space embeddings and metrics on probability measures.Journal of Machine Learning Research, 11:1517–1561, 2010

    Bharath K Sriperumbudur, Arthur Gretton, Kenji Fukumizu, Bernhard Schölkopf, and Gert RG Lanckriet. Hilbert space embeddings and metrics on probability measures.Journal of Machine Learning Research, 11:1517–1561, 2010

    2010

  33. [33]

    Approximation and convergence properties of generative adversarial learning.Advances in Neural Information Processing Systems, 30, 2017

    Shuang Liu, Olivier Bousquet, and Kamalika Chaudhuri. Approximation and convergence properties of generative adversarial learning.Advances in Neural Information Processing Systems, 30, 2017

    2017

  34. [34]

    Tweedie’s formula and selection bias.Journal of the American Statistical Association, 106:1602– 1614, 2011

    Bradley Efron. Tweedie’s formula and selection bias.Journal of the American Statistical Association, 106:1602– 1614, 2011

    2011

  35. [35]

    Rates of convergence for the Gaussian mixture sieve.Annals of Statistics, 28:1105–1127, 2000

    Christopher R Genovese and Larry Wasserman. Rates of convergence for the Gaussian mixture sieve.Annals of Statistics, 28:1105–1127, 2000

    2000

  36. [36]

    Entropies and rates of convergence for maximum likelihood and Bayes estimation for mixtures of normal densities.Annals of Statistics, 29:1233–1263, 2001

    Subhashis Ghosal and Aad W Van Der Vaart. Entropies and rates of convergence for maximum likelihood and Bayes estimation for mixtures of normal densities.Annals of Statistics, 29:1233–1263, 2001

    2001

  37. [37]

    Generative adversarial nets

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. InProceedings of the 28th International Conference on Neural Information Processing Systems-Volume 2, pages 2672–2680, 2014

    2014

  38. [38]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InInternational Conference on Machine Learning, pages 2256–2265. PMLR, 2015

    2015

  39. [39]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  40. [40]

    Score- based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score- based generative modeling through stochastic differential equations. InInternational Conference on Learning Representations, 2021

    2021

  41. [41]

    Coupling- based invertible neural networks are universal diffeomorphism approximators.Advances in Neural Information Processing Systems, 33:3362–3373, 2020

    Takeshi Teshima, Isao Ishikawa, Koichi Tojo, Kenta Oono, Masahiro Ikeda, and Masashi Sugiyama. Coupling- based invertible neural networks are universal diffeomorphism approximators.Advances in Neural Information Processing Systems, 33:3362–3373, 2020

    2020

  42. [42]

    Multilayer feedforward networks are universal approxi- mators.Neural Networks, 2:359–366, 1989

    Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approxi- mators.Neural Networks, 2:359–366, 1989

    1989

  43. [43]

    Neural spline flows

    Conor Durkan, Artur Bekasovs, Iain Murray, and George Papamakarios. Neural spline flows. In33rd Conference on Neural Information Processing Systems, pages 7511–7522. Neural Information Processing Systems Foundation, Inc, 2019

    2019

  44. [44]

    Finite sample properties of parametric mmd estimation: robustness to misspecification and dependence.Bernoulli, 28(1):181–213, 2022

    Badr-Eddine Chérief-Abdellatif and Pierre Alquier. Finite sample properties of parametric mmd estimation: robustness to misspecification and dependence.Bernoulli, 28(1):181–213, 2022. 12 Nonparametric Deconvolution and Denoising

    2022

  45. [45]

    Universal robust regression via maximum mean discrepancy.Biometrika, 111(1):71–92, 2024

    Pierre Alquier and Mathieu Gerber. Universal robust regression via maximum mean discrepancy.Biometrika, 111(1):71–92, 2024

    2024

  46. [46]

    On the rate of convergence in Wasserstein distance of the empirical measure

    Nicolas Fournier and Arnaud Guillin. On the rate of convergence in Wasserstein distance of the empirical measure. Probability Theory and Related Fields, 162:707–738, 2015

    2015

  47. [47]

    MMD GAN: Towards deeper understanding of moment matching network.Advances in Neural Information Processing Systems, 30:1–11, 2017

    Chun-Liang Li, Wei-Cheng Chang, Yu Cheng, Yiming Yang, and Barnabás Póczos. MMD GAN: Towards deeper understanding of moment matching network.Advances in Neural Information Processing Systems, 30:1–11, 2017

    2017

  48. [48]

    Noise2self: Blind denoising by self-supervision

    Joshua Batson and Loic Royer. Noise2self: Blind denoising by self-supervision. InInternational conference on machine learning, pages 524–533. PMLR, 2019

    2019

  49. [49]

    Training deep learning based denoisers without ground truth data

    Shakarim Soltanayev and Se Young Chun. Training deep learning based denoisers without ground truth data. Advances in neural information processing systems, 31, 2018

    2018

  50. [50]

    Blind universal bayesian image denoising with gaussian noise level learning.IEEE Transactions on Image Processing, 29:4885–4897, 2020

    Majed El Helou and Sabine Süsstrunk. Blind universal bayesian image denoising with gaussian noise level learning.IEEE Transactions on Image Processing, 29:4885–4897, 2020

    2020

  51. [51]

    Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004

    2004

  52. [52]

    A review of image denoising algorithms, with a new one

    Antoni Buades, Bartomeu Coll, and Jean-Michel Morel. A review of image denoising algorithms, with a new one. Multiscale Modeling & Simulation, 4:490–530, 2005

    2005

  53. [53]

    Image denoising by sparse 3-d transform-domain collaborative filtering.IEEE Transactions on Image Processing, 16:2080–2095, 2007

    Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering.IEEE Transactions on Image Processing, 16:2080–2095, 2007

    2080

  54. [54]

    Taming diffusion models for image restoration: a review.Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 383, 2025

    Ziwei Luo, Fredrik Gustafsson, Zheng Zhao, Jens Sjölund, and Thomas Schön. Taming diffusion models for image restoration: a review.Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 383, 2025

    2025

  55. [55]

    Spectral normalization for generative adversarial networks.arXiv preprint arXiv:1802.05957, 2018

    Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks.arXiv preprint arXiv:1802.05957, 2018

    Pith/arXiv arXiv 2018

  56. [56]

    How well generative adversarial networks learn distributions.Journal of Machine Learning Research, 22:1–41, 2021

    Tengyuan Liang. How well generative adversarial networks learn distributions.Journal of Machine Learning Research, 22:1–41, 2021

    2021

  57. [57]

    Error bounds for approximations with deep ReLU networks.Neural Networks, 94:103–114, 2017

    Dmitry Yarotsky. Error bounds for approximations with deep ReLU networks.Neural Networks, 94:103–114, 2017

    2017

  58. [58]

    MIT Press, 2016

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville.Deep Learning. MIT Press, 2016

    2016

  59. [59]

    On the method of bounded differences.Surveys in Combinatorics, 141:148–188, 1989

    Colin McDiarmid et al. On the method of bounded differences.Surveys in Combinatorics, 141:148–188, 1989

    1989

  60. [60]

    McGraw Hill, 1974

    Walter Rudin.Real and complex analysis. McGraw Hill, 1974

    1974

  61. [61]

    Elsevier, 2003

    Robert A Adams and John JF Fournier.Sobolev spaces. Elsevier, 2003

    2003

  62. [62]

    John Wiley & Sons, 1999

    Gerald B Folland.Real analysis: modern techniques and their applications. John Wiley & Sons, 1999

    1999

  63. [63]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations

  64. [64]

    A practical introduction to kernel discrepancies: Mmd, hsic & ksd.arXiv preprint arXiv:2503.04820, 2025

    Antonin Schrab. A practical introduction to kernel discrepancies: Mmd, hsic & ksd.arXiv preprint arXiv:2503.04820, 2025

    arXiv 2025

  65. [65]

    Mmd-fuse: Learning and combining kernels for two-sample testing without data splitting.Advances in Neural Information Processing Systems, 36:75151–75188, 2023

    Felix Biggs, Antonin Schrab, and Arthur Gretton. Mmd-fuse: Learning and combining kernels for two-sample testing without data splitting.Advances in Neural Information Processing Systems, 36:75151–75188, 2023. 13 Nonparametric Deconvolution and Denoising A Assumptions For clarity, we collect here all assumptions used in the paper. Theorems in the main text...

    2023

  66. [66]

    Here, we verify the sensitivity of the learned prior and downstream denoising performance on bandwidth choice

    This heuristic provides user with a principled data-driven default. Here, we verify the sensitivity of the learned prior and downstream denoising performance on bandwidth choice. Setup.We hold the full training procedure fixed and vary only a single multiplicative scale factor s∈ 0.10,0.25,0.50,0.75,1.00,1.50,2.00,3.00,5.00 applied uniformly to all nine b...

    2000