PolyCLEAN: Hybrid Imaging in Radio Interferometry

• PolyCLEAN is a radio interferometric imaging algorithm that extends classic CLEAN by integrating explicit ℓ1 convex optimization for sparse, super-resolution imaging.
• It utilizes a polyatomic Frank-Wolfe approach with restricted LASSO to efficiently update sparse active sets, reducing memory usage and accelerating convergence.
• The method introduces a dual certificate image for uncertainty quantification, localizing the recovered source positions and enhancing imaging reliability.

Searching arXiv for PolyCLEAN and closely related radio interferometric imaging work to ground the article in current literature. arxiv_search query="PolyCLEAN radio interferometry CLEAN LASSO Frank-Wolfe dual certificate" max_results=10

arxiv_search({"query":"PolyCLEAN radio interferometry CLEAN LASSO Frank-Wolfe dual certificate","max_results":10}) Searching arXiv for exact paper title and related review. arxiv_search: PolyCLEAN: Atomic Optimization for Super-Resolution Imaging and Uncertainty Estimation in Radio Interferometry PolyCLEAN is a radio interferometric imaging algorithm that extends and modifies the traditional CLEAN algorithm by integrating explicit convex optimization, specifically an $\ell_1$-based regularization, and was introduced as a method for super-resolution imaging and uncertainty estimation in radio interferometry (Jarret et al., 2024). In the recent interferometric-imaging literature, it is characterized as a hybrid deconvolution method that implements CLEAN within a LASSO framework, preserving the familiar major/minor-loop workflow while replacing heuristic component selection with an explicit sparsity-promoting optimization (Müller, 18 Sep 2025).

1. Position within radio interferometric imaging

Classical radio interferometric imaging is an ill-posed, noisy inverse problem, and the CLEAN family has historically remained dominant because of speed, robustness, and integration with calibration pipelines. That dominance persists even though CLEAN assumes the sky can be modeled as a collection of point sources, a modeling choice that limits recovery of extended emission and hampers automation (Müller, 18 Sep 2025).

PolyCLEAN was developed against this background. Its stated aim is to address two obstacles that have limited the adoption of convex optimization in radio interferometry: a fine-resolution setup that scales naturally in terms of memory usage and reconstruction speed, and a tool to localize a region of uncertainty for the recovered sources (Jarret et al., 2024). In that sense, PolyCLEAN occupies an intermediate position between classical CLEAN and more computationally demanding forward-modeling approaches such as Bayesian inference and AI-driven methods. The review literature places it among the near-term hybrid methods that retain CLEAN’s workflow while incorporating modern optimization (Müller, 18 Sep 2025).

A common misconception is that PolyCLEAN is simply another heuristic CLEAN variant. The central distinction is that it is designed to reproduce the numerical behavior of CLEAN while guaranteeing that it solves the minimization problem of interest. That property places it closer to an explicitly posed inverse-problem solver than to purely greedy deconvolution heuristics (Jarret et al., 2024).

2. Inverse problem and optimization model

PolyCLEAN begins from the discrete measurement equation

$$\mathbf{V} = \bm{\Phi}(\mathbf{I}) + \bm{\varepsilon},$$

where $\mathbf{V}$ denotes the measured visibilities, $\bm{\Phi}$ the Fourier or measurement operator, $\mathbf{I}$ the unknown sky image, and $\bm{\varepsilon}$ i.i.d. complex Gaussian noise (Jarret et al., 2024).

The reconstruction is then formulated as a sparsity-promoting convex optimization problem. Using the classical $\ell_1$ penalty, the inverse problem is written as

$$\underset{\mathbf{I} \in \mathbb{R}^N}{\arg\min}\quad \frac{1}{2}\|\mathbf{V}-\bm{\Phi}\mathbf{I}\|_2^2+\lambda\|\mathbf{I}\|_1,$$

or, in the equivalent imaging-language formulation used in the review literature,

$$\underset{\mathbf{x}}{\mathrm{minimize}}\ \frac{1}{2}\lVert \mathbf{D} - \mathbf{A}\mathbf{x} \rVert_2^2 + \lambda \lVert \mathbf{x} \rVert_1,$$

with $\mathbf{D}$ the dirty image, $$\mathbf{V} = \bm{\Phi}(\mathbf{I}) + \bm{\varepsilon},$$0 the PSF convolution operator, and $$\mathbf{V} = \bm{\Phi}(\mathbf{I}) + \bm{\varepsilon},$$1 the sparse sky model (Jarret et al., 2024).

This formulation is the principal theoretical shift from CLEAN to PolyCLEAN. CLEAN enforces sparsity through heuristic elements such as gain, thresholding, and stopping rules. PolyCLEAN replaces that heuristic regularization with an explicit $$\mathbf{V} = \bm{\Phi}(\mathbf{I}) + \bm{\varepsilon},$$2 penalty. The review literature describes this as bringing convex optimization into the CLEAN minor loop and replacing the strictly greedy point-source selection with a LASSO-based component-selection step (Müller, 18 Sep 2025).

The model is particularly natural for point-source-dominated fields, but the underlying convex framework is also described as flexible enough to handle diffuse emission through appropriate transforms or dictionaries. This is important because one of the stated limitations of traditional CLEAN is its inefficiency in representing faint or extended structure (Jarret et al., 2024).

3. Atomic and polyatomic algorithmic structure

The algorithmic core of PolyCLEAN is a polyatomic variant of the Frank-Wolfe algorithm. Standard Frank-Wolfe is projection-free and constructs solutions as sparse convex combinations of atoms; in this setting the atoms are delta-function pixels or point sources. PolyCLEAN augments that scheme with a polyatomic refinement, meaning that several atoms are added at each iteration rather than only one (Jarret et al., 2024).

The paper’s algorithmic description proceeds through sparse residual-driven exploration and active-set reweighting. Starting from $$\mathbf{V} = \bm{\Phi}(\mathbf{I}) + \bm{\varepsilon},$$3, PolyCLEAN computes a residual image,

$$\mathbf{V} = \bm{\Phi}(\mathbf{I}) + \bm{\varepsilon},$$4

selects a set of pixels above a threshold governed by the quality parameter $$\mathbf{V} = \bm{\Phi}(\mathbf{I}) + \bm{\varepsilon},$$5, solves a restricted LASSO over the active support, and prunes coefficients that return to zero (Jarret et al., 2024). In the review description, this corresponds to a minor loop that performs LASSO-based component selection while the major loop compares model visibilities to the actual data (Müller, 18 Sep 2025).

This architecture explains why PolyCLEAN is repeatedly described as both familiar and non-heuristic. It retains the major/minor-loop organization known from CLEAN-based software, yet the actual component update is no longer a peak-picking subtraction rule. Instead, it is a convex optimization over the active set. The result is a method that naturally produces sparse images at each iteration and leverages that sparsity to reduce memory and computational requirements (Jarret et al., 2024).

The term “polyatomic” is therefore not merely terminological. It denotes a convergence-acceleration device that also makes the method resemble the practical behavior of CLEAN minor cycles more closely than a one-atom-at-a-time solver would.

4. Super-resolution, sparse operators, and scalability

A central claim of PolyCLEAN is that it addresses fine-resolution imaging regimes without the dense-iterate overhead typical of many proximal $$\mathbf{V} = \bm{\Phi}(\mathbf{I}) + \bm{\varepsilon},$$6 solvers. Classical dense solvers such as FISTA or APGD operate over full images, which becomes memory-intensive at very fine pixel scales. PolyCLEAN avoids that by storing and updating only sparse active sets, so that cost scales with the active atoms rather than the total number of pixels (Jarret et al., 2024).

The implementation uses a sparsity-aware non-uniform FFT operator through the HVOX package, described as accepting sparse input and output directly and avoiding conversions to dense FFT representations in every step. This design is explicitly tied to the claim that PolyCLEAN demonstrates good scalability performance, in particular for fine-resolution grids (Jarret et al., 2024).

The empirical characterization reported in the paper is that the Python-based implementation is competitive with the fast numerically-optimized CLEAN solver. The paper further states that this acceleration does not affect image reconstruction quality: PolyCLEAN images are consistent with CLEAN-obtained ones for both point sources and diffuse emission recovery, and the method is also demonstrated on observed radio measurements (Jarret et al., 2024).

In the broader review literature, this computational profile is treated as especially relevant for the ngVLA, SKA, and ALMA’s Wideband Sensitivity Upgrade, where data volumes and dynamic ranges are expected to exceed the scalability of traditional methods. Hybrid approaches such as PolyCLEAN are accordingly presented as a near-term route to lower residuals in fewer iterations while preserving pipeline compatibility (Müller, 18 Sep 2025).

5. Dual certificate image and uncertainty localization

PolyCLEAN’s most distinctive addition to the CLEAN lineage is the introduction of the dual certificate image. This object appears as a numerical byproduct of the Frank-Wolfe algorithm and is proposed as a tool for uncertainty quantification on the location of the recovered sources (Jarret et al., 2024).

In the LASSO setting, the dual certificate is written as

$$\mathbf{V} = \bm{\Phi}(\mathbf{I}) + \bm{\varepsilon},$$7

and is described as unique for a given data and measurement set, independent of which specific LASSO-optimal solution is obtained (Jarret et al., 2024). Its saturated pixels, where $$\mathbf{V} = \bm{\Phi}(\mathbf{I}) + \bm{\varepsilon},$$8, define a saturation set that necessarily contains the support of all possible LASSO solutions. The paper interprets those saturated regions as localizing all positions consistent with any minimizer, so that the size of the regions provides an uncertainty or ambiguity estimate for source positions (Jarret et al., 2024).

The same framework is also used to propose a data-driven, a posteriori super-resolution beam built from the autocorrelation of the saturated dual certificate. The paper describes this as sharper than the traditional CLEAN beam and reflective of the actual resolving power of the data and model (Jarret et al., 2024).

At the same time, the 2025 review still contrasts PolyCLEAN with Bayesian and AI-based methods as not providing the broader uncertainty quantification associated with forward-modeling frameworks (Müller, 18 Sep 2025). This suggests that PolyCLEAN’s uncertainty machinery is localized and optimization-specific rather than a full posterior characterization.

6. Comparative assessment and role in pipelines

PolyCLEAN is best understood comparatively, because its stated significance lies in combining the operational habits of CLEAN with the formal structure of convex optimization.

Aspect	CLEAN	PolyCLEAN
Inverse problem	Greedy matching pursuit	Explicit convex LASSO minimization
Regularization	Heuristic	Explicit convex $$\mathbf{V} = \bm{\Phi}(\mathbf{I}) + \bm{\varepsilon},$$9
Iteration structure	Major/minor loops	Major/minor loops with LASSO-based minor loop
Fine-grid scaling	Limited by image size assumptions	Good scalability performance for fine-resolution grids
Uncertainty tool	None	Dual certificate image

The review literature attributes several consequences to this design: improved regularization, better control and reduction of spurious components, superior recovery of extended emission, faster convergence, and greater support for automation because the method is less reliant on user input such as masks and gains (Müller, 18 Sep 2025). The paper introducing PolyCLEAN is more restrained in formulation, emphasizing instead that the method is competitive with optimized CLEAN, consistent with CLEAN reconstructions for both point and diffuse emission, and suitable as an alternative to CLEAN in the radio interferometric imaging pipeline (Jarret et al., 2024).

That difference in emphasis is itself informative. The core paper establishes PolyCLEAN as a scalable atomic optimization method with super-resolution and uncertainty-localization capabilities. The subsequent review places it within a larger algorithmic transition in which hybrid methods provide an evolutionary path from robust legacy pipelines toward Bayesian and AI-based imaging frameworks (Müller, 18 Sep 2025). A plausible implication is that PolyCLEAN’s principal historical importance lies less in displacing CLEAN outright than in demonstrating that Bayesian priors and explicit convex optimization can be inserted into CLEAN-like workflows without sacrificing scalability or numerical performance.