The work, authored by Shenghan Mei and Yifei Lou of the University of North Carolina at Chapel Hill and Chengyu Ke and Miju Ahn of Southern Methodist University, targets composite problems in which the objective is the sum of two parts: a smooth convex function, typically measuring how well a model fits the data, and a weakly convex term that may be non-smooth, usually encoding desired structure such as sparsity or robustness. Weak convexity is a relaxed notion: a function is weakly convex if adding a suitable quadratic term makes it convex. The smallest such curvature correction is called the weak convexity modulus, and it plays a central role throughout the new theory.

This seemingly modest relaxation covers an impressive range of practical models. Popular non-convex regularizers used in statistics and signal processing, including the smoothly clipped absolute deviation penalty, the minimax concave penalty, logarithmic regularization, and the transformed l1 penalty, all fall into the weakly convex category. Because these penalties can shrink small coefficients all the way to zero without the heavy bias of the classic l1 penalty, they are prized for feature selection in high-dimensional settings such as gene expression analysis and large-vocabulary text classification. Yet their non-convexity means standard convex toolboxes no longer apply, and researchers must carefully design algorithms whose behavior can still be guaranteed.

The team’s first contribution is to apply the alternating direction method of multipliers, or ADMM, to an unconstrained version of this composite problem. ADMM is a splitting technique with roots in the 1970s: it introduces an auxiliary variable to split the problem, then alternates between minimizing over each block while a dual variable enforces agreement between them. The elegance of the approach is that each subproblem can become convex even though the original problem is not. The authors prove a key proposition showing that if the penalty parameter rho is at least as large as the weak convexity modulus, both the non-smooth subproblem and the smooth subproblem are individually convex and therefore solvable by standard methods such as gradient descent or Newton’s method.

Why prefer ADMM over simpler first-order alternatives such as proximal gradient, majorization-minimization, or difference-of-convex algorithms? The authors point out that those methods take a single gradient step on the smooth term, with a step size capped by its Lipschitz constant, which can be conservative when that constant is large or hard to estimate. ADMM instead solves the smooth-term subproblem to high accuracy at each outer iteration, which in practice can produce faster progress. However, this accuracy comes at a price: the smooth subproblem often has no closed-form solution, forcing the algorithm to run an expensive inner iterative solver at every outer iteration, and these nested loops can dominate the total runtime.

To eliminate the bottleneck, the researchers developed a linearized ADMM variant, LADMM. The idea is to replace the smooth function in the subproblem with its first-order approximation, meaning a linearization around the current iterate. This preserves the splitting structure and leaves the other subproblem and dual update untouched, while turning the expensive subproblem into a single closed-form gradient step. The result is an algorithm with dramatically cheaper iterations. The trade-off is theoretical: linearization introduces a lagged-gradient effect, because the update relies on a gradient evaluated at the previous point rather than the current one, and this effect must be controlled carefully in any convergence proof.

The convergence analysis is where the paper makes its most distinctive technical contribution. Prior non-convex ADMM analyses, such as the influential work of Hong and colleagues, typically require the non-convex term to be differentiable with a Lipschitz-continuous gradient, an assumption that non-smooth weakly convex regularizers violate. Others rely on the Kurdyka-Lojasiewicz property or semi-algebraic structure. The new work instead develops a sufficient-decrease argument that directly leverages the weak convexity modulus, yielding explicit and verifiable parameter conditions: for ADMM, rho must exceed both the modulus and the Lipschitz constant of the smooth gradient, while LADMM needs the stronger condition rho larger than the maximum of the modulus and two expressions involving the Lipschitz constant. For LADMM, the authors also introduce a modified augmented Lagrangian, augmented with an auxiliary quadratic correction sequence that absorbs the lagged-gradient effect and restores a monotone decrease property.

With these tools, the team proves that both algorithms generate iterates with subsequences converging to directional stationary points of the original problem. A directional stationary point is one where the directional derivative of the objective is non-negative in every feasible direction, a notion often stronger than ordinary critical points in general non-convex optimization. A key structural insight of the paper is that for this weakly convex composite class, directional stationarity, critical-point stationarity, and Clarke stationarity all coincide, because the objective admits a difference-of-convex decomposition with a differentiable concave component. This equivalence means that proving convergence to one type of stationary point automatically secures the other two, giving the guarantees unusually broad scope for a non-smooth non-convex setting.

The theory was stress-tested in three numerical case studies, all run in MATLAB on a modest laptop. The first two involve classic low-dimensional benchmark functions, the Three-Hump Camel and the Six-Hump Camel, compared against gradient descent, heavy ball, and Nesterov’s accelerated gradient. LADMM was the fastest method in both tests: on the Three-Hump Camel it averaged 0.0086 seconds versus roughly 0.10 seconds for gradient descent, while ADMM, despite requiring nearly twice the total iterations of gradient descent owing to inner solves, still converged several times faster. Contour plots of the trajectories revealed why: ADMM’s alternating updates of two variables produce widely spaced early steps that descend more effectively than single-variable gradient methods, even when the total iteration count is higher.

The third and most consequential experiment tackled a high-dimensional sparse logistic regression model with logarithmic regularization, a weakly convex penalty well suited to classification tasks where features outnumber samples. Here the proposed methods competed against a majorization-minimization scheme, a proximal linearized ADMM, the proximal alternating linearized minimization method known as PALM, and a standard DC algorithm. Over 30 random instances, LADMM achieved the lowest average computation time at about 0.46 seconds, roughly 3.7 times faster than ADMM, while matching the best baselines in accuracy, precision, and recall, all around 0.97, and attaining the smallest proximal-gradient mapping residual, a measure of closeness to directional stationarity. A scaling study from 50 to 300 features showed LADMM’s advantage widening dramatically, with a majorization-minimization baseline needing nearly 670 seconds at the largest size versus half a second for LADMM.

The authors are candid about the limits of their results. The convergence guarantees are subsequential rather than full-sequence results, extending them would require additional assumptions such as the Kurdyka-Lojasiewicz property, and the theory assumes subproblems are solved exactly, leaving inexact updates as important future work. They also note that the theoretical lower bounds on the penalty parameter can be conservative in practice, since stable convergence was observed even below the threshold, and they call for adaptive strategies for tuning rho. One experimental variant that linearizes the weakly convex term itself is not yet covered by the analysis, another open thread. Even so, the combination of explicit, checkable parameter conditions, three coinciding notions of stationarity, and striking empirical speedups makes this framework a compelling blueprint for the next generation of algorithms operating in the vast, structured territory between convex and fully non-convex optimization.

Subject of Research: Alternating direction methods of multipliers for weakly convex composite optimization

Article Title: ADMM and linearized ADMM for weakly convex minimization

Article References: Mei, S., Ke, C., Lou, Y., & Ahn, M. (2026). ADMM and linearized ADMM for weakly convex minimization. Frontiers in Applied Mathematics and Statistics, 12, Article 1898659. https://doi.org/10.3389/fams.2026.1898659

Image Credits: AI Generated

DOI: 10.3389/fams.2026.1898659

Keywords: ADMM, linearized ADMM, weakly convex optimization, convergence analysis, directional stationary point, non-convex optimization, splitting methods, sparse logistic regression, regularization, nonsmooth optimization, machine learning, optimization algorithms

The research team, led by Mohammad Akram of the Islamic University of Madinah together with Umar Ishtiaq, Khaleel Ahmad, and Ioan-Lucian Popa, proves that convergence survives under a far more permissive condition called enriched nonexpansivity. A map in this class is allowed to be expansive, to stretch distances, as long as it becomes nonexpansive once it is averaged with the identity operator. Formally, an operator T is b-enriched when a specific inequality holds that bounds how much T can expand points, with the amount of permissible expansion absorbed by averaging. The enriched class strictly contains the nonexpansive one, so every previously certifiable denoiser still qualifies, but an entire new territory of expressive networks opens up alongside it. Crucially, the enrichment constant b is not merely an abstraction; it becomes a computable number that a program can extract from a trained network.

The heart of the paper is a reduction, and it has a satisfying geometric flavor. The authors show that when the composite Plug-and-Play operator is enriched, an inertial Mann or Ishikawa iteration, classical fixed-point schemes that move only part of the way toward the operator’s output and add a momentum-like correction, becomes term by term identical to an ordinary iteration on a hidden nonexpansive operator called the generator. The inertial correction turns out to be exactly the quantity that cancels the overshoot introduced by the expansion. In other words, the rigidity is not imposed on the network at all; it is moved off the network and onto the iteration, where averaging repairs the loss of structure automatically. The fixed points of the expansive composite and its tame generator are the same set, so the convergent iteration lands on a genuine Plug-and-Play solution.

Turning this into usable software required a certificate, a way of checking for any given trained denoiser whether the enrichment hypothesis actually holds. The team’s main characterization theorem states that a smooth operator is b-enriched precisely when its displacement, the difference between the identity and the operator, is infinitesimally cocoercive with a modulus tied to b. The smallest admissible enrichment constant is one less than the supremum of a Rayleigh-type quotient of that displacement, a ratio that measures how far the displacement moves a direction relative to how much of that motion is aligned with the direction itself. This quotient is exact rather than merely sufficient, it is attained on non-scalar operators, it remains bounded near the identity where coarser bounds degenerate, and it can be evaluated through Jacobian-vector products that automatic differentiation supplies at the cost of roughly one network evaluation each.

The numbers that emerge are striking. On a linear surrogate problem in 120 dimensions, constructed so that every constant is available in closed form, the exact certificate widened the admissible step-size window by about 31 percent compared with the standard field-of-values bound, and the computed values matched the ground truth. The surrogate itself made the point vividly: an expansive operator with a spectral norm above one, admitted by no nonexpansive condition, was certified with a finite enrichment constant, and averaging it yielded a generator whose norm measured just below one. On real image restoration tasks with trained denoisers, the team measured the certified constant for each composite operator individually, using Lanczos iterations and bisection built on the same differentiable primitives.

The headline experimental result concerns Gaussian deblurring. A denoiser trained with a relaxed Lipschitz target produced a composite operator with a Lipschitz constant of 2.08, a value that no nonexpansive, firmly nonexpansive, or contractive-residual condition in the existing literature would accept. Yet its symmetric-part maximum stayed below one, and the exact characterization certified it with an enrichment constant of just 0.60, more than fifteen times tighter than the field-of-values relaxation for the same operator. The gap is a direct measurement of non-normality: the deep network expands strongly in directions poorly aligned with the residual, something the pair of conventional spectral quantities cannot see. Most remarkably, this expansive operator restored images within four hundredths of a decibel of a certified nonexpansive composite of identical architecture, meaning the larger, more expressive class was reached at no measured cost in quality.

Certification alone would be a curiosity without control over the iteration itself, so the authors built a safeguarded step-size controller on top of the theory. Any adaptive policy that stays inside the certified window preserves the convergence guarantee, whatever the policy has learned, so speed can be trained while safety is decided by the certificate. Within that architecture they derived a closed-form greedy relaxation, a Barzilai-Borwein-type step for the fixed-point residual, which roughly halved the iteration count of the constant-step scheme in both the surrogate and the image experiments. An amortized policy trained by unrolled differentiation through the solver matched the greedy rule without ever violating its guarantee, and the team verified empirically that stepping beyond the certified window by a factor of two caused divergence, confirming that the certificates mark a real stability boundary rather than an overly cautious one.

The framework also proved robust in ways that matter for practice. Convergence rates degrade gracefully, with the enrichment constant entering the residual bound as a simple known factor while the order of convergence is untouched. The theory tolerates denoiser evaluation errors of summable norm, covering finite-precision arithmetic and early-stopped inner solvers, which is exactly the situation encountered in real pipelines. A splitting-agnostic extension carries the entire construction beyond forward-backward splitting to Plug-and-Play ADMM and other schemes, since all that is ever required is a finite enrichment constant for the fixed-point operator of the scheme, checkable through the same eigenvalue computations.

The authors are careful about scope. The certificate is first order and measured on a finite sample, so the reported guarantees hold on the region actually certified rather than the whole image domain without further work. On the linear surrogate, an unrelaxed Picard iteration that carries no guarantee happened to be the fastest method, a reminder that acceleration claims here are comparisons among safeguarded schemes. The restoration study used small residual networks on modest image crops, so the reconstruction numbers are not a bid for state-of-the-art performance. What the results establish is arguably more consequential: the enlarged operator class is genuinely occupied by networks the existing guarantees reject, and using it costs nothing measurable in quality.

The deeper significance of the work may lie beyond imaging. Wherever a learned operator is iterated to a fixed point, in deep equilibrium models, learned optimizers, or unrolled reconstruction networks, the same tension between expressivity and guaranteed convergence reappears. The transfer-return mechanism offers a uniform recipe: certify the composite as enriched, transfer to the nonexpansive generator, iterate with the inertial correction, and transfer back. Whether that recipe extends to those settings remains an open conjecture, but the authors have made it checkable, reducing the entire question to eigenvalue computations that automatic differentiation already supports. In a field where convergence proofs have traditionally been bought by constraining networks into submission, the idea that the mathematics can instead bend around the network is a quiet revolution with loud implications.

Subject of Research: Convergence theory and certification for enriched-nonexpansive Plug-and-Play image restoration with learned denoisers

Article Title: Enriched-nonexpansive Plug-and-Play image restoration

Article References: Akram, M., Ishtiaq, U., Ahmad, K., & Popa, I.-L. (2026). Enriched-nonexpansive Plug-and-Play image restoration. Frontiers in Applied Mathematics and Statistics, 12, Article 1945678. https://doi.org/10.3389/fams.2026.1945678

Image Credits: AI Generated

DOI: 10.3389/fams.2026.1945678

Keywords: Plug-and-Play restoration, enriched nonexpansive mappings, fixed-point iteration, inertial Mann iteration, denoisers, image deblurring, super-resolution, convergence guarantees, cocoercivity, automatic differentiation, spectral certification, ADMM

The core problem the researchers tackle is a classic inverse problem in imaging physics. A CT scanner measures how X-rays attenuate as they pass through tissue, and from those projection measurements a reconstruction algorithm must infer the internal structure of the body. When the dose is lowered, the measurements become noisier and the mathematical problem becomes more ill-posed, meaning many different images could plausibly explain the same data. Traditional iterative reconstruction methods handle this by imposing explicit constraints, such as data consistency with the measured projections and regularization assumptions about what natural images look like. These methods are mathematically principled and interpretable, but they are computationally expensive and their hand-crafted priors often fail to capture the full statistical richness of real anatomical images.

Deep learning has swept through this field in recent years. Networks trained on pairs of low-dose and normal-dose images can learn to denoise and reconstruct with remarkable speed and quality, often in a single forward pass. Yet these end-to-end approaches carry a serious liability in a clinical setting: they are opaque. When a network produces a reconstruction, no one can say with confidence which parts of the output are trustworthy measurements and which are hallucinated details invented by the network’s training data. In medicine, where a fabricated structure could lead to a misdiagnosis, this lack of transparency is not a philosophical inconvenience but a practical barrier to adoption. Plug-and-play methods, which insert a learned denoiser into an iterative optimization loop, offer a partial compromise, but they too leave the relationship between the learned component and the optimization objective largely unexamined.

U-NetL2O takes a different route, one rooted in the Learning-to-Optimize paradigm. Instead of training a network to map noisy images directly to clean ones, the authors embed a U-Net-based learned update operator inside an unrolled Linearized ADMM scheme. ADMM, the Alternating Direction Method of Multipliers, is a workhorse of large-scale optimization that breaks a difficult problem into simpler subproblems solved alternately until convergence. In the unrolled version, a fixed number of these iterations are laid out as layers of a network, and the update steps themselves are parameterized by learnable components. The U-Net, a convolutional architecture famous for its success in biomedical image segmentation, serves as the learned update operator that refines the image estimate at each iteration, injecting data-driven prior knowledge about what realistic CT images look like.

The crucial design decision is that the overall optimization structure is preserved. The data-fidelity term, which ties the reconstruction to the actual scanner measurements, remains explicit and intact. The learned network does not replace the physics; it augments it. This means that every intermediate image produced during reconstruction is a meaningful point in the optimization trajectory, not an arbitrary feature activation. If the algorithm converges, it converges toward a solution that satisfies both the measured data and the learned prior. If it fails, the failure occurs within a structured mathematical framework where it can be located and diagnosed. This is precisely the property that end-to-end networks lack, and it is what the authors mean when they describe their method as explainable by construction.

Interpretability, however, is not just an architectural property; it needs to be visible to the people who use the system. To that end, the researchers introduce an optimization-aware explainable AI approach based on heatmap visualizations. Borrowing ideas from gradient-based localization techniques such as Grad-CAM, which highlight which regions of an input most influence a network’s output, their method illustrates how structural information is progressively refined across the optimization iterations. A clinician or researcher can watch, iteration by iteration, which anatomical regions the algorithm is focusing on and how the learned update operator reshapes them. Early iterations may establish coarse structures, while later iterations sharpen edges and suppress noise in targeted areas. The heatmaps turn the reconstruction from a single inscrutable output into a traceable process, revealing the internal dynamics of the learned optimization in a form that humans can inspect and audit.

The framework goes one step further with empirical diagnostic indicators that monitor in-distribution consistency. The idea is statistical: a reconstruction algorithm trained on one class of images should behave predictably when applied to similar data. When the input departs too far from the training distribution, for example because of unusual scanner settings, extreme noise levels, or rare anatomy, the learned components may behave unreliably. The diagnostic indicators provide warning signals that flag heavily degraded reconstructions, alerting users that the output should be treated with suspicion. This kind of built-in self-monitoring echoes broader efforts in the machine learning community, such as model cards and certification frameworks for trustworthy AI, and it addresses a genuine clinical need: radiologists must know not only what an algorithm produces but when that production is likely to be wrong.

Experimentally, the authors report that U-NetL2O achieves competitive reconstruction performance on low-dose CT benchmarks, including the LoDoPaB-CT dataset, a widely used benchmark built from the Lung Image Database Consortium collection that has become a standard proving ground for reconstruction algorithms. Competitive performance is itself notable, because interpretability is often assumed to come at a steep cost in accuracy. Here, the structured optimization framework does not sacrifice image quality to gain transparency. The combination suggests that the dichotomy between principled classical methods and powerful learned methods is not a true trade-off but a design choice, and that hybrid architectures can capture the best of both worlds when the learned components are placed carefully within the mathematical structure.

The significance of this work extends beyond CT reconstruction. Algorithm unrolling, the technique of converting iterative optimization algorithms into trainable network architectures, has emerged as a major theme in signal and image processing precisely because it retains interpretability while gaining representational power. By demonstrating an explainable L2O framework with optimization-aware visualizations and statistical diagnostics, the Vietnamese team provides a template that could be adapted to other inverse problems: magnetic resonance imaging, positron emission tomography, seismic imaging, and astronomical deconvolution all share the same fundamental structure of noisy measurements and ill-posed inference. In each of these domains, the question of whether an algorithm’s output can be trusted is as important as the output itself, and frameworks that make trust verifiable are likely to shape how regulators, clinicians, and scientists adopt machine learning.

The authors have also made their work reproducible, releasing both code and datasets through a public GitHub repository, which allows other researchers to scrutinize, extend, and stress-test the framework. Supported by funding from Vietnam’s National Foundation for Science and Technology Development, the study reflects a growing international effort to make medical AI accountable rather than merely accurate. As low-dose imaging protocols spread and deep learning penetrates deeper into clinical workflows, the demand for systems that can explain themselves will only intensify. U-NetL2O does not claim to have solved medical AI transparency in one stroke, but it demonstrates concretely that a reconstruction algorithm can be simultaneously learned, competitive, and legible, showing its reasoning in heatmaps and flagging its own failures in statistics. In a field where the stakes are measured in diagnoses and radiation doses, that combination may prove to be the most important result of all.

Subject of Research: Explainable learning-to-optimize deep learning for low-dose CT image reconstruction

Article Title: U-NetL2O: An explainable model for low-dose CT image reconstruction

Article References: Pham, S., Nguyen, N.-T., & Le, B. (2026). U-NetL2O: An explainable model for low-dose CT image reconstruction. Applied Intelligence, 56(15), Article 432. https://doi.org/10.1007/s10489-026-07480-y

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07480-y

Keywords: low-dose CT, image reconstruction, explainable AI, learning to optimize, U-Net, ADMM, algorithm unrolling, plug-and-play priors, heatmap visualization, diagnostic indicators, medical imaging, deep learning