{"id":"e5db7ec4-7acc-4121-9508-d3c0b0acd2c3","arxiv_id":"2506.11732","paper_version":1,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":0.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A review lecture-note series that surveys classical and data-driven methods for inverse problems, focusing on adversarial regularization and provably convergent plug-and-play denoisers.","lead":"These lecture notes introduce the data-driven paradigm for inverse problems, covering classical regularization, learned iterative schemes, adversarial regularization, and plug-and-play denoisers. They explain how deep learning can be combined with mathematical guarantees, which matters for trustworthy medical imaging.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.2.4's squared-distance maximizer is not 1-Lipschitz on unbounded domains, so the stated theoretical justification for adversarial regularization is incorrect; the intended result should use the unsquared distance function.","rationale":"The reader's verdict UNVERDICTED is appropriate because these are lecture notes without a novel central claim. My stress-test focused on whether the mathematical content supporting the advertised methods is sound. The most load-bearing issue I found is not the linear-denoiser limitation (which the paper openly acknowledges and qualifies), but a concrete error in Theorem 3.2.4: the squared distance to the data manifold is claimed to be a maximizer of the Wasserstein loss over 1-Lipschitz functions, yet it is not 1-Lipschitz on unbounded domains. This is a local correctness problem in the adversarial-regularization section, one of the two in-depth topics named in the abstract. It does not overturn the overall pedagogical value of the notes, and the survey's qualitative message about combining theory with learning still stands, so I do not recommend changing the reader's UNVERDICTED status. However, the theorem should be corrected to the unsquared distance function (or explicitly qualified with bounded-support assumptions), and the notes would benefit from a caveat that the stated version is false without such assumptions. I disagree with the reader's choice of weakest assumption because the linear-denoiser restriction is transparently labeled and not an internal error, whereas the squared-distance theorem is an actual misstatement that affects the credibility of a highlighted theoretical justification.","tokens_in":43423,"tokens_out":12074,"duration_ms":116719,"concrete_test":"Set M={0}⊂R^n, P_u=δ_0, P_n=δ_x with ||x||=2. The supremum in (3.5) over 1-Lipschitz R equals ||x||=2, attained by the unsquared distance R(u)=−||u||. The squared distance R(u)=||u||^2 gives value 4 and has Lipschitz constant at least 4, so it is not in the admissible class; Theorem 3.2.4's claimed maximizer therefore fails in this elementary case. Check the original proof or citation of Theorem 3.2.4 to confirm whether the intended statement is the unsquared distance function.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2.2 motivates adversarial regularization with Theorem 3.2.4, which asserts that the squared distance to the data manifold u ↦ min_{v∈M} ||u−v||^2 maximizes the Wasserstein loss (3.5) over 1-Lipschitz functions R. This is not correct as stated. A 1-Lipschitz function must satisfy |R(u)−R(v)| ≤ ||u−v||, but the squared distance has gradient 2(u−P_M(u)), whose norm is unbounded on R^n. The hypotheses (DMA and LNA) do not restrict the support of the noisy distribution to a set of small diameter, so the squared distance is generally not an admissible candidate in the supremum defining (3.5). The classical and correct statement uses the unsquared distance function d_M(u)=min_{v∈M}||u−v||, which is 1-Lipschitz. Since adversarial regularization is one of the two methods highlighted in the abstract, this misstatement directly weakens the notes' claim that the presented methods are theoretically grounded. The reader's concern about linear plug-and-play denoisers is separate and valid, but the squared-distance issue is a concrete internal mathematical error that should be corrected before the notes are used as a reference.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"These lecture notes provide an introduction to inverse problems and survey both classical and data-driven reconstruction approaches. The first part covers well-posedness, variational regularization, total variation and PDE-based methods, and the main numerical optimization tools. The second part discusses learned iterative schemes, learned variational models with a focus on adversarial regularization, and plug-and-play methods, including a spectral-filtering construction for linear denoisers that yields convergent regularization. The notes argue that combining deep learning with rigorous regularization theory is necessary for reliable and interpretable reconstruction, and they close with perspectives on task adaptation and open problems.","tokens_in":43727,"tokens_out":8080,"duration_ms":75805,"significance":"The notes are a valuable and readable survey of an active area, with the considerable strength that they consistently separate provable results from empirical heuristics. The discussion of learned iterative schemes honestly documents their lack of convergence guarantees and the risk of hallucination in severely ill-posed problems, and the spectral-filtering treatment of linear plug-and-play denoisers gives a concrete, parameter-free mechanism for controlling regularization strength. The authors also give useful pointers to their own convergence results for weakly convex regularizers. However, the theoretical justification of adversarial regularization contains a genuine mathematical error in Theorem 3.2.4, and this must be corrected before the notes can serve as a reliable reference.","major_comments":[{"comment":"The theorem states that, under the Data Manifold Assumption and the Low Noise Assumption, the squared distance function u ↦ min_{v∈M} ‖u−v‖² is a maximizer of the Wasserstein loss (3.5) over all 1-Lipschitz functions R. This is not correct as stated: on Rⁿ the squared distance has gradient 2(u−P_M(u)), whose norm is unbounded when the domain is unbounded, so it is not 1-Lipschitz and is therefore not admissible in the supremum defining (3.5). The assumptions DMA and LNA do not restrict the support of the noisy distribution P_n to a bounded set, so the classical remedy of restricting to a compact domain does not apply. The correct and standard statement uses the unsquared distance d_M(u)=min_{v∈M} ‖u−v‖, which is 1-Lipschitz. Because Theorem 3.2.4 is the explicit theoretical justification for adversarial regularization advertised in the abstract, the statement must be corrected or replaced with the correct theorem, with a precise citation to the original source.","section":"§3.2.2, Theorem 3.2.4"}],"minor_comments":[{"comment":"The definition says 'R_α y → A†y = u† for all f ∈ dom(A†)'; the quantifier should be 'for all y ∈ dom(A†)', not 'for all f'.","section":"§1.2.1, Definition 1.2.1"},{"comment":"The last line of the theorem statement says 'generalizing Theorem 1.2.1', but there is no Theorem 1.2.1; the intended cross-reference is likely Definition 1.2.1. Also, the proof is only referenced to Mukherjee et al. [2024], which is acceptable for lecture notes, but the reference should be made explicit in the statement.","section":"§1.2.2, Theorem 1.2.2"},{"comment":"The dimensions are used inconsistently: the first bullet writes 'n < m' for A: Rⁿ → ran(A) ⊂ Rᵐ, while the second bullet writes 'n > m and A: Rᵈ → Rᵐ' with an undefined symbol d; this should be A: Rⁿ → Rᵐ.","section":"§1.1, Example 1.1.2"},{"comment":"The text refers to 'Theorem 2.4.1' and Example 2.4.1 refers to 'Theorem 2.3.1', but neither theorem exists; these should be cross-references to the named ROF problem or to the appropriate example/definition numbers.","section":"§2.4, Examples 2.4.1 and surrounding text"},{"comment":"The text defines the spectral filter as g_τ(λ)=λ/(τ−λ(τ−1)), but the vertical axis of Figure 3.8a is labeled 'h_τ(λ)'; the notation should be made consistent.","section":"§3.3.2, Figure 3.8"},{"comment":"The assumptions that the linear denoiser is symmetric, positive semi-definite, non-expansive, and has bounded inverse are stated in prose; since they are the crucial hypotheses for the convergent-regularization result, they should be displayed as a formal assumption and followed by an explicit remark that most learned deep denoisers do not satisfy them, so the result does not directly apply in the general nonlinear setting.","section":"§3.3.2"}],"recommendation":"major_revision","confidential_remarks":"The squared-distance error in Theorem 3.2.4 is load-bearing but readily fixable by replacing it with the unsquared distance function or by citing the correct original result. No other load-bearing mathematical errors were found in the derivations included in the notes. I therefore recommend major revision rather than rejection: the survey is otherwise sound and useful, but the stated theoretical foundation for adversarial regularization must be corrected before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"These are lecture notes, not a research contribution, and they should be reviewed as such. What they do well: they give a clear, mathematically literate tour of classical regularization, variational models, and the main data-driven paradigms, with honest discussion of limitations and open problems. The spectral filtering derivation for linear plug-and-play denoisers is a nice self-contained presentation of Hauptmann et al., and the notes are refreshingly candid about the gap between empirical success and theoretical guarantees.\n\nThe soft spots are real but localized. Theorem 3.2.4 states that the squared distance to the data manifold maximizes the 1-Lipschitz Wasserstein loss. That cannot be right: the squared distance is not 1-Lipschitz on an unbounded domain, so it is not an admissible candidate in the supremum defining the loss. The intended statement should use the unsquared distance, which is 1-Lipschitz. This matters because adversarial regularization is one of the two methods highlighted in the abstract, and the error undermines the claimed theoretical grounding of that section. Definition 1.2.1 also has a typo (\"for all f in dom(A†)\" should be \"for all y\"), and the proof of Theorem 1.2.2 is only referenced, which is fine for lecture notes but should be flagged. The linear plug-and-play convergence result assumes symmetric positive semi-definite denoisers; most learned denoisers do not satisfy this, so the convergent regularization result has narrow direct applicability. The notes do state the assumption, so this is a limitation rather than an error.\n\nWho gets value from this? Graduate students and researchers entering the field will find a competent, well-organized overview. As a review it has no new theorems or data, but it is not worthless—it could serve as course material or a survey chapter. The mathematical error in Theorem 3.2.4, plus the typo, should be corrected before the notes are used as a reference. A serious referee would catch these issues.\n\nRecommendation: send to peer review with a request for correction of Theorem 3.2.4 and the typo. The notes are solid enough to be worth the referee time, provided the review focuses on accuracy of exposition rather than novelty.","headline":"Useful lecture notes on data-driven inverse problems, but Theorem 3.2.4 misstates the distance function and should be corrected before the notes are used as a reference.","tokens_in":44218,"tokens_out":1845,"would_cite":false,"duration_ms":19217,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65J20","65K10","68T07","94A08"],"pacs":[],"model":"deepseek-v4-flash","headline":"These lecture notes argue that data-driven solvers for inverse problems can be both highly accurate and mathematically trustworthy when learned components are embedded in classical variational regularization, through mechanisms such as…","keywords":["inverse problems","deep learning","regularization theory","plug-and-play denoisers","adversarial regularization","convergent regularization","variational methods","medical imaging"],"falsifier":"Take a trained deep denoiser and compute its Jacobian at several natural images: if the Jacobian is not symmetric with eigenvalues in $[0,1]$, the denoiser cannot be the proximal operator of a convex functional, and the spectral-filtering convergence theorem in the notes does not cover it. Alternatively, run the filtered plug-and-play iteration on a known ground-truth case with noise levels $\\delta\\to 0$: if reconstruction error does not approach zero under the prescribed parameter rule, the claimed convergent-regularization property fails.","tokens_in":43262,"feed_emoji":"🩻","tokens_out":12531,"duration_ms":115904,"temperature":0.7,"pith_summary":"Inverse problems ask for an unknown image from indirect, noisy measurements, and they are ill-posed: small errors in data can become large errors in the answer. Classical regularization makes them stable by adding handcrafted priors, but those priors capture only part of real-image structure. The notes claim that data-driven deep-learning solvers can keep their accuracy and speed while regaining the guarantees of classical regularization, provided their learned components are embedded in a variational or proximal framework. The two worked mechanisms are adversarial regularization, which learns a prior by separating good images from bad ones, and convergent linear plug-and-play denoising, which tunes regularization by spectral filtering so that reconstructions provably approach the true solution as noise vanishes. The payoff is a route to learned reconstruction that is accurate, efficient, and reliable enough for safety-critical use such as medical imaging.","feed_headline":"Learned imaging gains provable reliability when paired with math","feed_subtitle":"Adversarial regularization and spectral filtering extend classical stability guarantees to learned imaging.","key_machinery":"The load-bearing object is the proximal operator of convex analysis, $\\mathrm{prox}_J=(I+\\partial J)^{-1}$, because it is the bridge between variational regularization and denoisers: a plug-and-play scheme replaces the proximal step of a regularizer by a denoiser. For a linear denoiser $D_\\sigma$, the notes use the equivalence $J(x)=\\frac{1}{2}\\langle x,(D_\\sigma^{-1}-I)x\\rangle$ and derive that scaling $J$ by $\\tau$ corresponds to the spectral filter $g_\\tau(\\lambda)=\\lambda/(\\tau-\\lambda(\\tau-1))$ applied to the eigenvalues of $D_\\sigma$, so $\\mathrm{prox}_{\\tau J}=g_\\tau(D_\\sigma)$. That identity converts a fixed denoiser into a one-parameter family of regularization operators and makes the convergence-to-truth result possible. On the adversarial side, the key object is a regularizer $R_\\Theta(u)=\\Psi_\\Theta(u)+\\rho_0|u|^2/2$ trained through a Wasserstein-1 loss with a gradient penalty; its role is to encode the geometry of the clean-image distribution without paired supervision, with the distance to the data manifold serving as the ideal regularizer under the stated assumptions.","core_discovery":"On the paper's terms, the central discovery is that the dichotomy between mathematical reconstruction and deep learning is false: learned components can be grafted onto the variational formulation $\\min_u D(Au,y)+\\alpha R(u)$ and still inherit its stability and convergence. Adversarial regularization trains $R$ as a 1-Lipschitz function that assigns low values to clean images and high values to corrupted ones, using a Wasserstein-1 loss that does not require paired examples; under a data-manifold and low-noise assumption, the distance to the data manifold is a maximizer of that loss, and the gradient flow of the trained regularizer decreases the Wasserstein distance at the fastest possible rate. For plug-and-play methods, the notes show that when a linear denoiser is symmetric and positive semi-definite with eigenvalues in $[0,1]$, it is the proximal operator of a convex functional $J$, and scaling $J$ by $\\tau$ is implemented by filtering the denoiser's eigenvalues through $g_\\tau(\\lambda)=\\lambda/(\\tau-\\lambda(\\tau-1))$; with a suitable parameter rule this gives a convergent regularization, a guarantee absent from generic learned iterative schemes. The notes also argue that unconstrained fully learned methods can hallucinate structures in severely ill-posed problems even when their pixel metrics improve, so mathematical structure is not an optional extra but the condition for reliability.","pith_inferences":["Editorial inference: if the spectral-filtering construction can be extended to nonlinear denoisers, for instance through the Tweedie-scaling ideas the notes cite, then pretrained deep denoisers could be endowed with convergent-regularization guarantees without being retrained under restrictive linearity constraints.","Editorial inference: the data-manifold characterization of adversarial regularizers suggests a natural out-of-distribution detector, since images with unusually high regularizer values sit far from the learned clean distribution, which could support uncertainty quantification in clinical deployment.","Editorial inference: the same eigenvalue-filtering logic gives a principled calibration rule for regularization strength in other operator-splitting schemes beyond plug-and-play, potentially replacing heuristics such as denoiser scaling with a parameter rule that has a convergence guarantee.","Editorial inference: task-adapted reconstruction implies that evaluation benchmarks for imaging should weight downstream diagnostic utility at least as heavily as pixel-level metrics, since a reconstruction can be better for segmentation without having better PSNR."],"forward_implications":["Learned regularizers trained with a Wasserstein separation loss can be inserted into variational problems and inherit existence, uniqueness, stability, and convergence guarantees from the classical theory.","A linear plug-and-play denoiser whose eigenvalues are filtered by $g_\\tau$ defines a convergent regularization: as the noise level $\\delta$ tends to zero and $\\tau$ is chosen by a suitable rule, the reconstruction converges to the underlying solution.","Deep equilibrium networks that constrain the learned update to be a contraction converge to a fixed point even beyond the number of training steps, avoiding the divergence artifacts seen in unconstrained unrolled networks.","Jointly training reconstruction and a downstream task such as segmentation can improve the downstream task relative to isolated training, because the extra degrees of freedom of an ill-posed problem can be steered toward task-relevant reconstructions.","Purely learned methods can report better PSNR and SSIM while hallucinating anatomical structures in severely ill-posed problems, so reliability requires grounding learned components in mathematical guarantees."],"supporting_citations":[{"why":"Supplies the spectral-filter construction and the convergent-regularization theorem for linear plug-and-play denoisers.","marker":"[Hauptmann et al., 2024]"},{"why":"Defines adversarial regularizers and the Wasserstein training loss, including the gradient-flow optimality result.","marker":"[Lunz et al., 2018]"},{"why":"Introduces plug-and-play priors, the framework that the convergence analysis extends.","marker":"[Venkatakrishnan et al., 2013]"},{"why":"Gives the contraction condition and convergence theorem for deep equilibrium networks, plus the artifact example.","marker":"[Gilton et al., 2021a]"},{"why":"Provides the proximity-operator characterization that lets a linear symmetric positive semi-definite denoiser be treated as a proximal map.","marker":"[Moreau, 1965]"},{"why":"Gives the modern characterization of proximity operators used to identify the implicit regularizer.","marker":"[Gribonval and Nikolova, 2020]"},{"why":"Proposes RED, an explicit denoiser-based regularizer whose symmetry restriction motivates the linear-case analysis.","marker":"[Romano et al., 2017]"},{"why":"Supplies the joint reconstruction-segmentation loss and the empirical task-adaptation results in the final chapter.","marker":"[Adler et al., 2022]"}],"fun_headline_variants":["Deep learning meets math for reliable inverse problems","Provable convergence for learned imaging methods","Data-driven inverse problems get rigorous guarantees","Learned imaging inherits classical stability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The convergent plug-and-play guarantee rests on the assumption that the denoiser is linear, symmetric, and positive semi-definite, meaning it is exactly the proximal map of some convex penalty; because trained deep denoisers are neither linear nor symmetric, the guarantee does not directly cover them.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning meets math for reliable inverse problems","Provable convergence for learned imaging methods","Data-driven inverse problems get rigorous guarantees","Learned imaging inherits classical stability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1378,"prompt_tokens":1077,"completion_tokens":301,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":693,"completion_tokens_details":{"reasoning_tokens":249}},"tokens_in":693,"tokens_out":301,"duration_ms":3358,"temperature":1.0,"reasoning_tokens":249,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:02:58.067097+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained deep denoiser and compute its Jacobian at several natural images: if the Jacobian is not symmetric with eigenvalues in $[0,1]$, the denoiser cannot be the proximal operator of a convex functional, and the spectral-filtering convergence theorem in the notes does not cover it. Alternatively, run the filtered plug-and-play iteration on a known ground-truth case with noise levels $\\delta\\to 0$: if reconstruction error does not approach zero under the prescribed parameter rule, the claimed convergent-regularization property fails.","supporting_citations":[],"review_version":1}