pith. sign in

arxiv: 2606.30230 · v1 · pith:E7VQMQ5Knew · submitted 2026-06-29 · 🧮 math.OC · cs.LG

A Distributionally Robust Framework for Learned Reconstructions in Inverse Problems

Floor van Maarschalkerwaart , Subhadip Mukherjee , Christoph Brune , Marcello Carioni This is my paper

Pith reviewed 2026-06-30 05:22 UTC · model grok-4.3

classification 🧮 math.OC cs.LG
keywords distributionally robust optimizationinverse problemslearned reconstructionTikhonov regularizationWasserstein DROambiguity setsconditional perturbations
0
0 comments X

The pith

A structured distributionally robust optimization framework for inverse problems yields an explicit worst-case risk bound that induces Tikhonov regularization on the Lipschitz constant of the reconstruction operator.

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

The paper develops a distributionally robust optimization approach for learned reconstructions in inverse problems by restricting the ambiguity set to structured perturbations aligned with the data-acquisition process, such as changes only to the conditional distribution P(Y|X). This avoids the uniform perturbation of the full joint distribution used in standard Wasserstein DRO, which the authors argue is overly conservative and ignores measurement physics. Strong duality is established for the general case, with explicit finite-dimensional dual representations derived for joint, marginal, and conditional perturbations. The central result is an explicit worst-case risk bound that enforces Tikhonov regularization on the Lipschitz constant of the operator while remaining less conservative for well-posed problems. Experiments on deblurring and sinogram-to-CT tasks show the resulting operators improve robustness and stability over standard DRO and MSE training, and reduce to low-rank forms akin to truncated SVD in the linear case.

Core claim

By constraining perturbations to subsets such as P(Y|X) that align with the data-acquisition process, the framework models uncertainty in the forward operator and noise model more faithfully and accommodates any noise model expressible as a stochastic forward operator. Strong duality holds for this general formulation, and explicit finite-dimensional dual representations are derived for perturbations in the joint, marginal, and conditional distributions. A central result is an explicit worst-case risk bound that induces Tikhonov regularization on the Lipschitz constant of the reconstruction operator and is less conservative relative to standard DRO for well-posed problems.

What carries the argument

The structured ambiguity set restricted to perturbations aligned with the data-acquisition process (such as on the conditional distribution P(Y|X)), which carries the argument by enabling derivation of the explicit worst-case risk bound and the induced Tikhonov regularization on the reconstruction operator's Lipschitz constant.

If this is right

  • In the linear setting the learned operator becomes effectively low-rank, truncating at the intrinsic dimension of the data and recovering a data-driven analogue of truncated-SVD regularization.
  • The framework accommodates any noise model expressible as a stochastic forward operator.
  • Numerical experiments demonstrate improved robustness, stability, and interpretability over standard DRO and MSE baselines on deblurring and sinogram-to-CT reconstruction.
  • Strong duality holds for the general formulation with explicit finite-dimensional dual representations for joint, marginal, and conditional perturbations.

Where Pith is reading between the lines

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

  • The structured approach could be tested on other inverse problems with known forward physics, such as MRI or seismic imaging, to check whether the Lipschitz regularization effect persists beyond the reported tasks.
  • It may connect to existing robust optimization methods in control theory that also exploit conditional uncertainty structures rather than joint perturbations.
  • If the Lipschitz bound is tight in practice, it could motivate hybrid training objectives that combine the DRO risk with explicit Lipschitz penalties even outside the distributionally robust setting.

Load-bearing premise

Meaningful distributional uncertainty in inverse problems can be faithfully captured by restricting the ambiguity set to structured perturbations on subsets such as the conditional distribution P(Y|X) that align with the data-acquisition process.

What would settle it

A counterexample in which the derived explicit worst-case risk bound fails to hold or in which the method proves more conservative than standard DRO on a well-posed inverse problem would falsify the central claim.

Figures

Figures reproduced from arXiv: 2606.30230 by Christoph Brune, Floor van Maarschalkerwaart, Marcello Carioni, Subhadip Mukherjee.

Figure 1
Figure 1. Figure 1: Structured perturbations in the product space [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic positioning Structured DRO (center) between distributionally robust optimization (left) and inverse problems (right). DRO optimizes against the worst case over an ambiguity ball Bε(µ ∗ ), inducing regularization made explicit through strong duality. Inverse problems stabilize a reconstruction H† of an ill-posed forward operator H by adding a regularizer R to the data-fidelity term. Structured DRO… view at source ↗
Figure 3
Figure 3. Figure 3: Percentage gap between the joint and P(Y |X)-DRO bounds as the Lipschitz constant LG of the fixed inverse grows, for the diagonal operator (left) and the backward heat equation (right). The gap is largest at small LG and closes monotonically, since the P(Y |X) bound replaces p 1 + L 2 G by LG. The two problems start at nearly the same LG but different gaps (27.6% vs. 3.9%) because the heat-equation inverse… view at source ↗
Figure 4
Figure 4. Figure 4: Differentiation: reconstruction metrics of the DRO inverse compared to the Tikhonov baseline [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Spectral analysis of DRO and Tikhonov inverses. Left: singular values. Right: RMSE under [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: MNIST reconstructions (left) and saliency maps (right) under increasing Gaussian noise. [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: MNIST reconstruction metrics under increasing Gaussian noise with panels and shaded [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: MNIST reconstructions and saliency under increasing Poisson noise. Reconstruction columns [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: MNIST reconstruction metrics under increasing Poisson noise with panels and shaded bands [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: MNIST reconstructions and saliency maps for different blurring kernels. Reconstruction [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: MNIST reconstruction metrics across the five test kernels of Figure 10. Left: RMSE. Right: [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: MNIST reconstructions for extreme radii. Columns (left to right): ground truth, noisy [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: MNIST reconstruction metrics under increasing Gaussian noise for extreme radii with [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: CIFAR-10 reconstructions and saliency under increasing Gaussian noise. Reconstruction [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: CIFAR-10 reconstruction metrics under increasing Gaussian noise with panels and shaded [PITH_FULL_IMAGE:figures/full_fig_p021_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: MNIST robust simulator: simulated measurements and saliency maps under increasing [PITH_FULL_IMAGE:figures/full_fig_p022_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: MNIST robust simulation metrics under increasing Gaussian noise with panels and shaded [PITH_FULL_IMAGE:figures/full_fig_p022_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: 2DeteCT CT reconstruction metrics as the photon count [PITH_FULL_IMAGE:figures/full_fig_p023_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: 2DeteCT CT reconstructions, saliency, and difference at photon counts from high-dose [PITH_FULL_IMAGE:figures/full_fig_p024_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: MNIST reconstructions and saliency under increasing Gaussian noise. Comparison of ro [PITH_FULL_IMAGE:figures/full_fig_p030_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: MNIST reconstruction metrics. Comparison of robustness in [PITH_FULL_IMAGE:figures/full_fig_p030_21.png] view at source ↗
read the original abstract

Learned reconstruction operators for inverse problems are typically trained under a fixed noise model, and generalize poorly when the distribution during testing differs from the one assumed during training. Distributionally robust optimization (DRO) addresses this by optimizing against the worst-case distribution within a prescribed ambiguity set, but standard Wasserstein DRO perturbs the full joint distribution uniformly, which can be overly conservative and ignores the physics of the measurement process. We develop a structured DRO framework in which the ambiguity set is restricted to structured perturbations aligned with the data-acquisition process. This allows us to learn data-driven reconstruction operators that remain robust to distributional shifts. By constraining perturbations to subsets such as $P(Y|X)$, our framework models uncertainty in the forward operator and noise model more faithfully, accommodating any noise model expressible as a stochastic forward operator. We establish strong duality for this general formulation and derive explicit finite-dimensional dual representations for perturbations in the joint, marginal, and conditional distributions. A central result is an explicit worst-case risk bound that induces Tikhonov regularization on the Lipschitz constant of the reconstruction operator, and is less conservative relative to standard DRO for well-posed problems. Numerical experiments on deblurring and sinogram-to-CT reconstruction demonstrate improved robustness, stability, and interpretability over standard DRO and MSE baselines. In the linear setting, the learned operator becomes effectively low-rank, truncating at the intrinsic dimension of the data and recovering a data-driven analogue of truncated-SVD regularization.

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

3 major / 2 minor

Summary. The manuscript proposes a structured distributionally robust optimization (DRO) framework for training learned reconstruction operators in inverse problems. By restricting the ambiguity set to structured perturbations (e.g., on the conditional P(Y|X) or other acquisition-aligned subsets) rather than the full joint distribution, the approach aims to model uncertainty in the forward operator and noise more faithfully. The paper establishes strong duality, derives explicit finite-dimensional dual representations for joint/marginal/conditional cases, and presents a central result: an explicit worst-case risk bound that induces Tikhonov regularization on the Lipschitz constant of the reconstruction operator and is less conservative than standard Wasserstein DRO for well-posed problems. In the linear case, the learned operator is shown to become effectively low-rank. Numerical experiments on deblurring and sinogram-to-CT reconstruction are used to demonstrate improved robustness and stability over standard DRO and MSE baselines.

Significance. If the derivations and modeling assumptions hold, the framework provides a more physically grounded alternative to standard DRO for handling distributional shifts in inverse problems, with the explicit dual forms and connection to classical regularization (Tikhonov on Lipschitz constant, data-driven truncated SVD) offering both theoretical and practical value. The numerical validation on concrete imaging tasks strengthens the contribution.

major comments (3)
  1. [§4] §4 (Duality results): The strong duality and explicit dual representations are derived under the structured ambiguity set restricted to perturbations of P(Y|X) (or similar subsets). However, there is no general theorem characterizing the conditions under which this restriction is without loss of modeling power for arbitrary distributional shifts in inverse problems, or when the resulting worst-case risk bound remains valid outside the linear case. This directly affects the claim that the bound is less conservative than joint Wasserstein DRO.
  2. [§5.1] §5.1 (Worst-case risk bound): The central claim that the explicit worst-case risk bound induces Tikhonov regularization on the Lipschitz constant of the reconstruction operator rests on the assumption that meaningful distributional uncertainty can be captured by the restricted ambiguity set. The manuscript does not provide a concrete test or counterexample analysis showing when the true shift lies outside this class, which would make the improvement over standard DRO fail to hold.
  3. [§6] §6 (Numerical experiments): The reported improvements in robustness for deblurring and CT tasks are presented as evidence of reduced conservatism, but the choice of ambiguity-set radius and the precise definition of the structured perturbation sets used in the baselines are not detailed enough to verify that the comparison isolates the effect of the P(Y|X) restriction rather than post-hoc parameter tuning.
minor comments (2)
  1. The notation for the reconstruction operator R and the forward operator is introduced early but occasionally reused with different subscripts in the dual derivations; a consistent table of symbols would improve readability.
  2. Figure captions for the CT reconstruction results could more explicitly state the noise models and shift types used in the test distributions.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript. We provide point-by-point responses below and indicate where revisions will be made to address the concerns.

read point-by-point responses

  1. Referee: [§4] §4 (Duality results): The strong duality and explicit dual representations are derived under the structured ambiguity set restricted to perturbations of P(Y|X) (or similar subsets). However, there is no general theorem characterizing the conditions under which this restriction is without loss of modeling power for arbitrary distributional shifts in inverse problems, or when the resulting worst-case risk bound remains valid outside the linear case. This directly affects the claim that the bound is less conservative than joint Wasserstein DRO.

    Authors: We agree that a comprehensive theorem establishing when the restriction to P(Y|X) perturbations is without loss of generality for all possible distributional shifts is absent from the manuscript. Our approach is specifically designed for inverse problems where the primary sources of uncertainty (forward operator variations and noise) are naturally modeled via the conditional distribution P(Y|X). The reduced conservatism claim is made in comparison to joint Wasserstein DRO when the distributional shift respects this structure, which aligns with the physics of the acquisition process. For the linear case, we derive the low-rank property explicitly. We will revise the manuscript to include a dedicated paragraph in §4 discussing the modeling assumptions and the conditions under which the framework is expected to be less conservative, without claiming universality. revision: partial

  2. Referee: [§5.1] §5.1 (Worst-case risk bound): The central claim that the explicit worst-case risk bound induces Tikhonov regularization on the Lipschitz constant of the reconstruction operator rests on the assumption that meaningful distributional uncertainty can be captured by the restricted ambiguity set. The manuscript does not provide a concrete test or counterexample analysis showing when the true shift lies outside this class, which would make the improvement over standard DRO fail to hold.

    Authors: The derivation of the Tikhonov regularization on the Lipschitz constant is specific to the linear setting and the structured ambiguity set. To address the request for concrete analysis, we will add a subsection or appendix with a counterexample where a shift in the joint distribution cannot be captured by P(Y|X) perturbations, illustrating the boundary of applicability. This will clarify when the improvement holds and when it may not, strengthening the presentation of the result. revision: yes

  3. Referee: [§6] §6 (Numerical experiments): The reported improvements in robustness for deblurring and CT tasks are presented as evidence of reduced conservatism, but the choice of ambiguity-set radius and the precise definition of the structured perturbation sets used in the baselines are not detailed enough to verify that the comparison isolates the effect of the P(Y|X) restriction rather than post-hoc parameter tuning.

    Authors: We acknowledge that the experimental section lacks sufficient detail on the ambiguity set radii and the exact construction of the perturbation sets. In the revised version, we will expand §6 with explicit parameter values, a clear description of how the structured sets are implemented versus the joint Wasserstein baselines, and additional experiments varying the radius to show that the observed improvements are not due to tuning. This will allow readers to reproduce and verify the comparisons. revision: yes

Circularity Check

0 steps flagged

No circularity: derivations of duality and risk bounds are mathematically self-contained

full rationale

The provided abstract and description show the central results (strong duality, finite-dimensional dual representations, and the explicit worst-case risk bound inducing Tikhonov regularization on the Lipschitz constant) are obtained by direct mathematical derivation from the structured ambiguity set restricted to acquisition-aligned subsets such as P(Y|X). No equations or claims reduce by construction to fitted parameters, self-defined quantities, or load-bearing self-citations. The framework is presented as an independent modeling choice whose consequences (including reduced conservatism for well-posed problems) follow from the stated assumptions without circular renaming or smuggling. This matches the reader's assessment of minimal circularity concern.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract; the framework invokes strong duality for the general structured DRO formulation and assumes the existence of finite-dimensional dual representations for the chosen perturbation classes. No free parameters or invented entities are explicitly named in the abstract.

axioms (1)
  • standard math Strong duality holds for the structured distributionally robust optimization problem with ambiguity sets restricted to subsets such as P(Y|X).
    Invoked to obtain explicit dual representations; location: abstract statement on establishing strong duality.

pith-pipeline@v0.9.1-grok · 5799 in / 1268 out tokens · 25007 ms · 2026-06-30T05:22:47.619577+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

153 extracted references · 65 canonical work pages · 12 internal anchors

  1. [1]

    Adler, Jonas and Oktem, Ozan , year = 2018, month = jun, journal =. Learned. doi:10.1109/TMI.2018.2799231 , copyright =

    work page doi:10.1109/tmi.2018.2799231 2018

  2. [2]

    Learning to solve inverse problems using Wasserstein loss

    Adler, Jonas and Ringh, Axel and. Learning to Solve Inverse Problems Using. 1710.10898 , archiveprefix =

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Inverse Problems , volume =

    Learning Regularization Parameters of Inverse Problems via Deep Neural Networks , author =. Inverse Problems , volume =

  4. [4]

    Infinite Dimensional Analysis: A Hitchhiker's Guide , author =

  5. [5]

    and Burkinshaw, Owen , year = 1990, month = may, publisher =

    Aliprantis, Charalambos D. and Burkinshaw, Owen , year = 1990, month = may, publisher =. Principles of

    1990

  6. [6]

    Lectures on

    Ambrosio, Luigi and Bru. Lectures on. doi:10.1007/978-3-031-76834-7 , copyright =

    work page doi:10.1007/978-3-031-76834-7

  7. [7]

    Amend, Christian and Carioni, Marcello and Zemas, Konstantinos , year = 2026, month = mar, journal =. Atomic

    2026

  8. [8]

    Distributional

    Aolaritei, Liviu and Lanzetti, Nicolas and Chen, Hongruyu and D. Distributional. IEEE Transactions on Automatic Control , pages =

  9. [9]

    Wasserstein

    Arjovsky, Martin and Chintala, Soumith and Bottou, L. Wasserstein. Proceedings of the 34th

  10. [10]

    Acta Numerica , edition =

    Solving Inverse Problems Using Data-Driven Models , author =. Acta Numerica , edition =

  11. [11]

    ESAIM: Control, Optimisation and Calculus of Variations , volume =

    Regularization for Wasserstein Distributionally Robust Optimization , author =. ESAIM: Control, Optimisation and Calculus of Variations , volume =

  12. [12]

    Barata, Jo. The. Brazilian Journal of Physics , volume =

  13. [13]

    Quantifying

    Barbano, Riccardo and Zhang, Chen and Arridge, Simon and Jin, Bangti , year = 2021, month = jan, pages =. Quantifying. 2020 25th

    2021

  14. [14]

    Generalized Inverses: Theory and Applications , author =

  15. [15]

    Acta Numerica , volume =

    Modern Regularization Methods for Inverse Problems , author =. Acta Numerica , volume =. doi:10.1017/S0962492918000016 , langid =

    work page doi:10.1017/s0962492918000016

  16. [16]

    Biguri, Ander , year = 2026, journal =

    2026

  17. [17]

    Birrell, Jeremiah and Dupuis, Paul and Katsoulakis, Markos A. and. Distributional. doi:10.48550/arXiv.1911.09580 , archiveprefix =. 1911.09580 , primaryclass =

    work page doi:10.48550/arxiv.1911.09580 1911

  18. [18]

    Distributionally

    Blanchet, Jose and Li, Jiajin and Lin, Sirui and Zhang, Xuhui , year = 2025, month = aug, journal =. Distributionally

    2025

  19. [19]

    Quantifying

    Blanchet, Jose and Murthy, Karthyek , year = 2019, journal =. Quantifying

    2019

  20. [20]

    Blanchet, Jose and Kang, Yang and Murthy, Karthyek , year = 2019, month = sep, journal =. Robust. doi:10.1017/jpr.2019.49 , langid =

    work page doi:10.1017/jpr.2019.49 2019

  21. [21]

    Statistical

    Blanchet, Jose and Murthy, Karthyek and Nguyen, Viet Anh , year = 2021, month = aug, eprint =. Statistical. doi:10.48550/arXiv.2108.02120 , archiveprefix =

    work page doi:10.48550/arxiv.2108.02120 2021

  22. [22]

    Unifying

    Blanchet, Jose and Kuhn, Daniel and Li, Jiajin and Taskesen, Bahar , year = 2025, month = dec, eprint =. Unifying. doi:10.48550/arXiv.2308.05414 , archiveprefix =

    work page doi:10.48550/arxiv.2308.05414 2025

  23. [23]

    Bonnans, J. Fr. Perturbation

  24. [24]

    Boull. A. doi:10.1016/bs.hna.2024.05.003 , archiveprefix =. 2312.14688 , primaryclass =

    work page doi:10.1016/bs.hna.2024.05.003 2024

  25. [25]

    On the Extremal Points of the Ball of the

    Bredies, Kristian and Carioni, Marcello and Fanzon, Silvio and Romero, Francisco , year = 2019, month = jul, journal =. On the Extremal Points of the Ball of the. doi:10.1112/blms.12509 , howpublished =

    work page doi:10.1112/blms.12509 2019

  26. [26]

    Foundations of Computational Mathematics , volume =

    A Generalized Conditional Gradient Method for Dynamic Inverse Problems with Optimal Transport Regularization , author =. Foundations of Computational Mathematics , volume =. doi:10.1007/s10208-022-09561-z , archiveprefix =. 2012.11706 , primaryclass =

    work page doi:10.1007/s10208-022-09561-z 2012

  27. [27]

    Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges

    Bronstein, Michael M. and Bruna, Joan and Cohen, Taco and Veli. Geometric. doi:10.48550/arXiv.2104.13478 , archiveprefix =. 2104.13478 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2104.13478

  28. [28]

    doi:10.48550/arXiv.2501.03096 , archiveprefix =

    Analysis of Mean-Field Models Arising from Self-Attention Dynamics in Transformer Architectures with Layer Normalization , author =. doi:10.48550/arXiv.2501.03096 , archiveprefix =. 2501.03096 , primaryclass =

    work page doi:10.48550/arxiv.2501.03096

  29. [29]

    Symplectic

    Canizares, Priscilla and Murari, Davide and Sch. Symplectic. doi:10.48550/arXiv.2412.16787 , archiveprefix =. 2412.16787 , primaryclass =

    work page doi:10.48550/arxiv.2412.16787

  30. [30]

    Optimal Transport and Applications to

    Carioni, Marcello , langid =. Optimal Transport and Applications to

  31. [31]

    Unsupervised Approaches Based on Optimal Transport and Convex Analysis for Inverse Problems in Imaging , booktitle =

    Carioni, Marcello and Mukherjee, Subhadip and Ye Tan, Hong and Tang, Junqi , year = 2025, pages =. Unsupervised Approaches Based on Optimal Transport and Convex Analysis for Inverse Problems in Imaging , booktitle =

    2025

  32. [32]

    and Wang, Xiran and Elgendy, Omar A

    Chan, Stanley H. and Wang, Xiran and Elgendy, Omar A. , year = 2017, month = mar, journal =. Plug-and-. doi:10.1109/TCI.2016.2629286 , copyright =

    work page doi:10.1109/tci.2016.2629286 2017

  33. [33]

    Conditional

    Chemseddine, Jannis and Hagemann, Paul and Steidl, Gabriele and Wald, Christian , year = 2025, journal =. Conditional

    2025

  34. [34]

    , year = 2020, journal =

    Chen, Ruidi and Paschalidis, Ioannis Ch. , year = 2020, journal =. Distributionally

    2020

  35. [35]

    Chenreddy, Abhilash Reddy and Bandi, Nymisha and Delage, Erick , year = 2022, month = dec, journal =. Data-

    2022

  36. [36]

    Chizat, Lenaic and Bach, Francis , year = 2018, month = oct, eprint =. On the. doi:10.48550/arXiv.1805.09545 , archiveprefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1805.09545 2018

  37. [37]

    Regularization of

    Clason, Christian , year = 2021, month = feb, eprint =. Regularization of. doi:10.48550/arXiv.2001.00617 , archiveprefix =

    work page doi:10.48550/arxiv.2001.00617 2021

  38. [38]

    Sinkhorn Distances: Lightspeed Computation of Optimal Transport , booktitle =

    Cuturi, Marco , year = 2013, pages =. Sinkhorn Distances: Lightspeed Computation of Optimal Transport , booktitle =

    2013

  39. [39]

    A Survey on Diffusion Models for Inverse Problems

    Daras, Giannis and Chung, Hyungjin and Lai, Chieh-Hsin and Mitsufuji, Yuki and Ye, Jong Chul and Milanfar, Peyman and Dimakis, Alexandros G. and Delbracio, Mauricio , year = 2024, month = sep, eprint =. A. doi:10.48550/arXiv.2410.00083 , archiveprefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2410.00083 2024

  40. [40]

    , year = 2017, pages =

    Dashti, Masoumeh and Stuart, Andrew M. , year = 2017, pages =. The. Handbook of. doi:10.1007/978-3-319-12385-1_7 , isbn =

    work page doi:10.1007/978-3-319-12385-1_7 2017

  41. [41]

    Deng, Li , year = 2012, month = nov, journal =. The

    2012

  42. [42]

    Drazin, M. P. , year = 1958, journal =. Pseudo-

    1958

  43. [43]

    Learning

    Duchi, John and Namkoong, Hongseok , year = 2020, month = jul, journal =. Learning

    2020

  44. [44]

    Regularising

    Duff, M A G and Campbell, N D F and Ehrhardt, M J , year = 2024, journal =. Regularising

    2024

  45. [45]

    and Lang, Lukas F

    Ehrhardt, Matthias J. and Lang, Lukas F. , year = 2018, journal =. Inverse

    2018

  46. [46]

    , year = 1996, month = jul, publisher =

    Engl, Heinz Werner and Hanke, Martin and Neubauer, A. , year = 1996, month = jul, publisher =. Regularization of

    1996

  47. [47]

    and Pichler, Alois and Sprungk, Bj

    Ernst, Oliver G. and Pichler, Alois and Sprungk, Bj. Wasserstein. SIAM/ASA Journal on Uncertainty Quantification , volume =

  48. [48]

    Nathan and Brunton, Bingni W

    Fasel, Urban and Kutz, J. Nathan and Brunton, Bingni W. and Brunton, Steven L. , year = 2022, month = apr, journal =. Ensemble-. doi:10.1098/rspa.2021.0904 , archiveprefix =. 2111.10992 , primaryclass =

    work page doi:10.1098/rspa.2021.0904 2022

  49. [49]

    Distributionally

    Gao, Rui and Kleywegt, Anton , year = 2023, month = may, journal =. Distributionally

    2023

  50. [50]

    Journal of Computational Physics , volume =

    Wasserstein Generative Adversarial Uncertainty Quantification in Physics-Informed Neural Networks , author =. Journal of Computational Physics , volume =

  51. [51]

    Wasserstein

    Gao, Rui and Chen, Xi and Kleywegt, Anton J , year = 2022, month = nov, journal =. Wasserstein

    2022

  52. [52]

    Gao, Yiming , year = 2023, month = nov, journal =. A

    2023

  53. [53]

    Learning

    Genevay, Aude and Peyre, Gabriel and Cuturi, Marco , year = 2018, month = mar, pages =. Learning. Proceedings of the

    2018

  54. [54]

    A Physics-Informed Variational

    Goswami, Somdatta and Yin, Minglang and Yu, Yue and Karniadakis, George , year = 2022, month = mar, journal =. A Physics-Informed Variational. doi:10.1016/j.cma.2022.114587 , archiveprefix =. 2108.06905 , primaryclass =

    work page doi:10.1016/j.cma.2022.114587 2022

  55. [55]

    Goujon, Alexis and Neumayer, Sebastian and Bohra, Pakshal and Ducotterd, Stanislas and Unser, Michael , year = 2023, journal =. A

    2023

  56. [56]

    Guo, Mengwu , publisher =

  57. [57]

    GAMM-Mitteilungen , volume =

    Neural-Network-Based Regularization Methods for Inverse Problems in Imaging , author =. GAMM-Mitteilungen , volume =. doi:10.1002/gamm.202470004 , copyright =

    work page doi:10.1002/gamm.202470004

  58. [58]

    Multilevel

    Hagemann, Paul and Mildenberger, Sophie and Ruthotto, Lars and Steidl, Gabriele and Yang, Nicole Tianjiao , year = 2025, journal =. Multilevel

    2025

  59. [59]

    Regularization of

    Haltmeier, Markus and Nguyen, Linh , year = 2023, pages =. Regularization of. Handbook of. doi:10.1007/978-3-030-98661-2_81 , isbn =

    work page doi:10.1007/978-3-030-98661-2_81 2023

  60. [60]

    Mathematical Programming , volume =

    A Distributionally Robust Perspective on Uncertainty Quantification and Chance Constrained Programming , author =. Mathematical Programming , volume =. doi:10.1007/s10107-015-0896-z , langid =

    work page doi:10.1007/s10107-015-0896-z

  61. [61]

    Truncated

    Hansen, Per Christian , year = 1990, journal =. Truncated

    1990

  62. [62]

    Convergent

    Hauptmann, Andreas and Mukherjee, Subhadip and Sch. Convergent. Foundations of Computational Mathematics , volume =. doi:10.1007/s10208-024-09654-x , langid =

    work page doi:10.1007/s10208-024-09654-x

  63. [63]

    Wasserstein-

    Heaton, Howard and Fung, Samy Wu and Lin, Alex Tong and Osher, Stanley and Yin, Wotao , year = 2022, month = jun, journal =. Wasserstein-

    2022

  64. [64]

    Tomosipo:

    Hendriksen, Allard and Schut, Dirk and Palenstijn, Willem Jan and Vigan. Tomosipo:. Optics Express , publisher =

  65. [65]

    Learning

    Hertrich, Johannes and Wong, Hok Shing and Denker, Alexander and Ducotterd, Stanislas and Fang, Zhenghan and Haltmeier, Markus and Kereta, Zeljko and Kobler, Erich and Leong, Oscar and Salehi, Mohammad Sadegh and Sch. Learning. Medical

  66. [66]

    , year = 2012, publisher =

    Hu, Zhaolin and Hong, J. , year = 2012, publisher =. Kullback-

    2012

  67. [67]

    Wasserstein

    Jang, Minhyuk and Hakobyan, Astghik and Yang, Insoon , year = 2024, month = jun, eprint =. Wasserstein. doi:10.48550/arXiv.2406.01723 , archiveprefix =

    work page doi:10.48550/arxiv.2406.01723 2024

  68. [68]

    and Froustey, Emmanuel and Unser, Michael , year = 2017, month = sep, journal =

    Jin, Kyong Hwan and McCann, Michael T. and Froustey, Emmanuel and Unser, Michael , year = 2017, month = sep, journal =. Deep

    2017

  69. [69]

    Deep Learning of Normal Form Autoencoders for Universal Parameter-Dependent Dynamics , author =

  70. [70]

    doi:10.48550/arXiv.2106.05102 , archiveprefix =

    Learning Normal Form Autoencoders for Data-Driven Discovery of Universal,Parameter-Dependent Governing Equations , author =. doi:10.48550/arXiv.2106.05102 , archiveprefix =. 2106.05102 , primaryclass =

    work page doi:10.48550/arxiv.2106.05102

  71. [71]

    Generalized

    Karlsson, Johan and Ringh, Axel , year = 2017, journal =. Generalized

    2017

  72. [72]

    Manifold-

    Keegan, Katherine and Ruthotto, Lars , year = 2026, month = jan, eprint =. Manifold-. doi:10.48550/arXiv.2601.23151 , archiveprefix =

    work page doi:10.48550/arxiv.2601.23151 2026

  73. [73]

    and Coban, Sophia B

    Kiss, Maximilian B. and Coban, Sophia B. and Batenburg, K. Joost and. Scientific Data , volume =

  74. [74]

    Kobler, Erich and Effland, Alexander and Kunisch, Karl and Pock, Thomas , year = 2020, pages =. Total. Proceedings of the

    2020

  75. [75]

    and Lanthaler, Samuel and Stuart, Andrew M

    Kovachki, Nikola B. and Lanthaler, Samuel and Stuart, Andrew M. , year = 2024, month = feb, eprint =. Operator. doi:10.48550/arXiv.2402.15715 , archiveprefix =

    work page doi:10.48550/arxiv.2402.15715 2024

  76. [76]

    Learning

    Krizhevsky, Alex and Hinton, Geoffrey , year = 2009, address =. Learning

    2009

  77. [77]

    Operations Research & Management Science in the Age of Analytics , author =

    Wasserstein. Operations Research & Management Science in the Age of Analytics , author =

  78. [78]

    Operator Learning with

    Lanthaler, Samuel , year = 2023, journal =. Operator Learning with

    2023

  79. [79]

    doi:10.48550/arXiv.2509.03910 , archiveprefix =

    An Invertible Generative Model for Forward and Inverse Problems , author =. doi:10.48550/arXiv.2509.03910 , archiveprefix =. 2509.03910 , primaryclass =

    work page doi:10.48550/arxiv.2509.03910

  80. [80]

    doi:10.1088/1361-6420/ab6d57 , langid =

    Li, Housen and Schwab, Johannes and Antholzer, Stephan and Haltmeier, Markus , year = 2020, month = jun, journal =. doi:10.1088/1361-6420/ab6d57 , langid =

    work page doi:10.1088/1361-6420/ab6d57 2020

Showing first 80 references.