Autocorr-CLEAN in Radio Interferometry

• Autocorr-CLEAN is an adaptive deconvolution algorithm that employs the autocorrelation of residual images to construct non-radially symmetric, continuously adapting basis functions.
• It reduces computation by maintaining linear per-iteration costs and decreasing the number of minor-loop iterations by up to an order of magnitude.
• The method integrates seamlessly into existing CLEAN workflows by augmenting traditional point-source subtraction with autocorrelation-based clustering, enhancing fidelity on extended emission.

Searching arXiv for “Autocorr-CLEAN” and closely related radio interferometric CLEAN variants to ground the article in current papers. Autocorr-CLEAN is a name used for multiple distinct methods in the arXiv literature, but in radio interferometric imaging it denotes a CLEAN variant that uses the autocorrelation function of the residual image to accelerate deconvolution while remaining compatible with the standard major/minor loop framework (Müller et al., 22 Apr 2025). In that setting, the method is designed to retain the low per-iteration cost of classical CLEAN while introducing continuously adapting, potentially non-radially symmetric basis functions derived from clusters of CLEAN components fitted to the residual autocorrelation (Müller et al., 22 Apr 2025). The same label has also been applied in unrelated domains to a two-stage correction framework for turn-level statistical inference in LLM conversation analysis (Schessl, 15 Apr 2026) and to a denoising-autoencoder method for cleaning noisy correlation-matrix spectra (Hayou, 2017). This polysemy makes domain specification essential.

1. Terminological scope and disambiguation

In radio interferometry, Autocorr-CLEAN is presented as a deconvolution algorithm motivated by the computational demands of upcoming instruments and by the limitations of point-source minor-loop subtraction on extended emission (Müller et al., 22 Apr 2025). Its stated aim is to combine “continuously adapting basis functions,” “potentially non–radially symmetric shapes,” and “only linear-cost operations per iteration” by using the image-plane autocorrelation of the residual to propose a cluster of $\delta$-components as a single multiscale basis function (Müller et al., 22 Apr 2025).

The same term appears elsewhere with unrelated meanings. In LLM conversation analysis, “Autocorr-CLEAN” denotes a “two-stage correction framework combining Chelton (1983) effective degrees of freedom with conversation-level block bootstrap” for correcting spurious significance inflation caused by serial dependence in turn-level metrics (Schessl, 15 Apr 2026). In quantitative finance and statistical estimation, “Autocorr-CLEAN” is described as a supervised denoising approach for estimating the true eigenvalues of a population correlation matrix from a noisy sample correlation matrix by means of an adjusted autoencoder (Hayou, 2017).

This suggests that “Autocorr-CLEAN” is not a single cross-domain methodology but a reused label whose meaning depends entirely on context. In the interferometric-imaging literature, the term belongs to the family of hybrid CLEAN extensions surveyed alongside CG-CLEAN and PolyCLEAN (Müller, 18 Sep 2025).

2. Position within the CLEAN family

Classical CLEAN iteratively subtracts scaled, shifted copies of the point-spread function at the single brightest residual pixel; it is therefore characterized as “matching pursuit on $\delta$-functions” (Müller et al., 22 Apr 2025). This makes it fast per iteration because it requires only shifts and adds, but it can require many iterations to represent extended structure (Müller et al., 22 Apr 2025). Multiscale CLEAN replaces $\delta$-functions by fixed radial basis functions such as tapered parabolas, Gaussians, or wavelets, which improves convergence at the cost of extra per-iteration work to track all scales (Müller et al., 22 Apr 2025). Asp-CLEAN, by contrast, fits a single Gaussian per iteration via explicit non-linear optimization, improving convergence but introducing computationally heavy, super-linear steps (Müller et al., 22 Apr 2025).

Autocorr-CLEAN is framed as an attempt to combine the advantages of these approaches without inheriting their principal costs. It seeks adaptive basis functions like Asp-CLEAN, non-radial shapes like DoB-CLEAN, and linear-cost operations per iteration like CLEAN, with the autocorrelation of the residual serving as the organizing statistic for grouping components (Müller et al., 22 Apr 2025). A later review places it among “hybrid approaches” that preserve CLEAN’s workflow while incorporating more modern optimization ideas, and argues that such hybrids are especially suitable in the near term for facilities such as the ngVLA and SKA (Müller, 18 Sep 2025).

A common misconception is to treat Autocorr-CLEAN as merely a thresholded post-processing step attached to CLEAN. The radio-imaging description instead presents it as a modified minor-loop strategy in which the basis function used for subtraction is repeatedly re-estimated from second-order residual structure (Müller et al., 22 Apr 2025). The review literature also sketches a related operational picture in which standard point-source subtractions are periodically augmented by an autocorrelation-based “cluster-subtraction” step (Müller, 18 Sep 2025). Because the published descriptions are not identical, it is appropriate to distinguish the algorithm explicitly described in (Müller et al., 22 Apr 2025) from the high-level sketch consolidated in the review (Müller, 18 Sep 2025).

3. Mathematical construction from residual autocorrelation

The radio-interferometric formulation begins with the current dirty residual image, denoted $I^{res}$, and defines the autocorrelation operator by

$f \diamond g := f(x,y) \star g(-x,-y),$

where $\star$ is convolution (Müller et al., 22 Apr 2025). The image autocorrelation is then

$II = I^{res} \diamond I^{res}.$

Because $I^{res} = B^D \star I_{\text{true}}$, the corresponding factorization is

$$II = (B^D \diamond B^D) \star (I_{\text{true}} \diamond I_{\text{true}}),$$

with $B^D$ the dirty-beam point-spread function (Müller et al., 22 Apr 2025).

The central idea is to CLEAN the dirty autocorrelation rather than only the residual image itself. In each outer iteration, the method CLEANs $\delta$0 using the PSF autocorrelation

$\delta$1

to obtain a point-component model

$\delta$2

such that

$\delta$3

(Müller et al., 22 Apr 2025). The multiscale basis function is then formed as

$\delta$4

with the exposition noting that typically $\delta$5 so that $\delta$6 (Müller et al., 22 Apr 2025).

This representation avoids selecting an explicit scale radius. Each $\delta$7 has a location and a weight, and the basis shape emerges from the cloud of $\delta$8-components rather than from a pre-specified radial template (Müller et al., 22 Apr 2025). In practice,

$\delta$9

and the basis function actually used in the minor loop is

$\delta$0

(Müller et al., 22 Apr 2025). Because the support points $\delta$1 can be arbitrarily distributed, the resulting basis need not be radially symmetric (Müller et al., 22 Apr 2025).

The consolidated review offers a complementary but less formal formulation in terms of a penalized objective,

$\delta$2

where

$\delta$3

$\delta$4

and

$\delta$5

with $\delta$6 the current residual image (Müller, 18 Sep 2025). That review states that one does not minimize $\delta$7 in one shot; rather, one periodically computes the residual autocorrelation, identifies a dominant lag $\delta$8, and subtracts a cluster of CLEAN components whose spacing matches it (Müller, 18 Sep 2025). This suggests a conceptual connection between the explicit autocorrelation-deconvolution construction in (Müller et al., 22 Apr 2025) and a broader view of autocorrelation suppression as a regularizing principle.

4. Algorithmic workflow

The explicit algorithmic description in the deconvolution paper is organized around an initialization phase, an outer minor loop, and a subminor loop (Müller et al., 22 Apr 2025). Initialization consists of computing the initial dirty residual image $\delta$9, the dirty beam $I^{res}$0, the clean beam $I^{res}$1, the initial dirty autocorrelation $I^{res}$2, and the PSF autocorrelation $I^{res}$3 (Müller et al., 22 Apr 2025). The autocorrelation image is then CLEANed with $I^{res}$4 to obtain the initial $I^{res}$5 and thus $I^{res}$6; helper correlations are precomputed for subsequent updates (Müller et al., 22 Apr 2025).

The outer minor loop then proceeds iteratively. At iteration $I^{res}$7, the search metric

$I^{res}$8

is computed, the peak location $I^{res}$9 and peak value $f \diamond g := f(x,y) \star g(-x,-y),$0 are found in $f \diamond g := f(x,y) \star g(-x,-y),$1, and the shifted component

$f \diamond g := f(x,y) \star g(-x,-y),$2

is subtracted from the residual to obtain $f \diamond g := f(x,y) \star g(-x,-y),$3 (Müller et al., 22 Apr 2025). Helper arrays are updated in $f \diamond g := f(x,y) \star g(-x,-y),$4 without FFTs, after which a subminor loop re-approximates $f \diamond g := f(x,y) \star g(-x,-y),$5 by CLEAN on the updated autocorrelation image to produce a new $f \diamond g := f(x,y) \star g(-x,-y),$6 (Müller et al., 22 Apr 2025). The subminor loop stops when

$f \diamond g := f(x,y) \star g(-x,-y),$7

with $f \diamond g := f(x,y) \star g(-x,-y),$8 (Müller et al., 22 Apr 2025). The final model is

$f \diamond g := f(x,y) \star g(-x,-y),$9

shifted to the corresponding peak locations and convolved with the clean beam $\star$0 (Müller et al., 22 Apr 2025).

The review article describes a more schematic algorithm in which standard CLEAN runs for a fixed number of minor-loop steps, then computes

$\star$1

finds the largest off-zero peak $\star$2, and if that peak exceeds a threshold $\star$3, subtracts a cluster of CLEAN components in one step before resuming ordinary minor-loop subtraction (Müller, 18 Sep 2025). This should be read as a consolidated picture rather than as a verbatim specification of the algorithm in (Müller et al., 22 Apr 2025).

Aspect	Explicit deconvolution description	Review sketch
Residual statistic	$\star$4	$\star$5
Basis update	CLEAN $\star$6 with $\star$7 to obtain $\star$8	Find dominant lag $\star$9 and subtract a matching cluster
Loop structure	Outer minor loop plus subminor loop	Standard minor-loop interleaved with autocorr checks

5. Computational complexity and empirical behavior

The computational claim advanced for Autocorr-CLEAN is that it can deliver multiscale-style convergence improvements without super-linear work in the minor loops (Müller et al., 22 Apr 2025). If $II = I^{res} \diamond I^{res}.$0 is the number of pixels, $II = I^{res} \diamond I^{res}.$1 the number of outer iterations, $II = I^{res} \diamond I^{res}.$2 the number of $II = I^{res} \diamond I^{res}.$3-components per $II = I^{res} \diamond I^{res}.$4, and $II = I^{res} \diamond I^{res}.$5 the number of subminor iterations, then classical CLEAN is described as approximately $II = I^{res} \diamond I^{res}.$6 (Müller et al., 22 Apr 2025). Autocorr-CLEAN requires $II = I^{res} \diamond I^{res}.$7 to compute the search metric and perform subtraction, $II = I^{res} \diamond I^{res}.$8 to update helper correlations, and $II = I^{res} \diamond I^{res}.$9 to run $I^{res} = B^D \star I_{\text{true}}$0 autocorrelation-CLEAN steps, yielding approximately $I^{res} = B^D \star I_{\text{true}}$1 per outer iteration and total cost

$I^{res} = B^D \star I_{\text{true}}$2

(Müller et al., 22 Apr 2025). Because $I^{res} = B^D \star I_{\text{true}}$3 and $I^{res} = B^D \star I_{\text{true}}$4, and because each outer step fits $I^{res} = B^D \star I_{\text{true}}$5 $I^{res} = B^D \star I_{\text{true}}$6-components at once, the runtime is summarized as

$I^{res} = B^D \star I_{\text{true}}$7

(Müller et al., 22 Apr 2025).

The empirical evaluation reported in the deconvolution paper uses VLA-A synthetic tests of Cygnus A, Hercules A, M106, and Hydra A (Müller et al., 22 Apr 2025). On these tests, classical CLEAN needed approximately $I^{res} = B^D \star I_{\text{true}}$8–$I^{res} = B^D \star I_{\text{true}}$9 more time to reach the same residual level; Asp-CLEAN converged in few iterations but was approximately $$II = (B^D \diamond B^D) \star (I_{\text{true}} \diamond I_{\text{true}}),$$0–$$II = (B^D \diamond B^D) \star (I_{\text{true}} \diamond I_{\text{true}}),$$1 slower per iteration because of FFT calls; and Autocorr-CLEAN attained CLEAN-level residuals in $$II = (B^D \diamond B^D) \star (I_{\text{true}} \diamond I_{\text{true}}),$$2–$$II = (B^D \diamond B^D) \star (I_{\text{true}} \diamond I_{\text{true}}),$$3 less wall time than classical CLEAN and in $$II = (B^D \diamond B^D) \star (I_{\text{true}} \diamond I_{\text{true}}),$$4–$$II = (B^D \diamond B^D) \star (I_{\text{true}} \diamond I_{\text{true}}),$$5 less wall time than Asp-CLEAN (Müller et al., 22 Apr 2025). Reconstruction fidelity, measured as $$II = (B^D \diamond B^D) \star (I_{\text{true}} \diamond I_{\text{true}}),$$6, matched Asp-CLEAN and surpassed plain CLEAN, especially on extended, diffuse emission (Müller et al., 22 Apr 2025).

The ngVLA-era review gives a broader synthesis. It states that classical CLEAN’s point-source assumption limits recovery of extended emission and hampers automation, whereas hybrid methods such as Autocorr-CLEAN preserve the familiar workflow while improving optimization behavior (Müller, 18 Sep 2025). It further reports that Autocorr-CLEAN reduces the number of minor-loop iterations “by up to an order of magnitude” and that, at fixed wall-clock time or fixed number of minor iterations, residuals fall approximately $$II = (B^D \diamond B^D) \star (I_{\text{true}} \diamond I_{\text{true}}),$$7–$$II = (B^D \diamond B^D) \star (I_{\text{true}} \diamond I_{\text{true}}),$$8 faster than CLEAN and comparably to Asp-CLEAN, while final dynamic range can be “up to 20% better” (Müller, 18 Sep 2025). Since the review explicitly consolidates material from Mueller (2021) and an in-preparation follow-up, these figures are best interpreted as a secondary summary rather than a substitute for the primary deconvolution paper.

6. Parameters, failure modes, and pipeline integration

The deconvolution paper identifies three principal tuning parameters: the gain, the power $$II = (B^D \diamond B^D) \star (I_{\text{true}} \diamond I_{\text{true}}),$$9, and the subminor stopping fraction $B^D$0 (Müller et al., 22 Apr 2025). High gains around $B^D$1 are reported to speed convergence, whereas very high gains above $B^D$2 may accumulate errors; the default gain is described as approximately $B^D$3 (Müller et al., 22 Apr 2025). Recommended values for $B^D$4 are approximately $B^D$5–$B^D$6, motivated by “Gaussian-like square-root behavior,” and the method is said to work robustly even if $B^D$7 is off by $B^D$8 (Müller et al., 22 Apr 2025). For the subminor stopping fraction, the default is approximately $B^D$9; smaller $\delta$00 yields a deeper autocorrelation fit at extra cost, while larger $\delta$01 risks a crude $\delta$02 and slows residual CLEAN (Müller et al., 22 Apr 2025).

A practical safeguard is the “Hogbom switch”: if $\delta$03 becomes too $\delta$04-like, indicating small-scale domination, the algorithm can switch back to standard CLEAN permanently (Müller et al., 22 Apr 2025). The paper also states that Autocorr-CLEAN is less dependent on manual CLEAN windows because $\delta$05 encodes sidelobe suppression (Müller et al., 22 Apr 2025). In terms of software architecture, the method “fits into the usual major/minor self-cal loop hierarchy” and requires replacement only of the minor-loop code (Müller et al., 22 Apr 2025).

The review sketch provides compatible but somewhat different operational defaults: loop-gain $\delta$06–$\delta$07, $\delta$08–$\delta$09, $\delta$10–$\delta$11, and an autocorrelation detection threshold set to a few times the expected noise-autocorrelation level (Müller, 18 Sep 2025). It recommends stopping when the residual peak is at most $\delta$12 or when peak autocorrelation drops below the threshold, and emphasizes that the method can be inserted into an existing CASA minor-loop imaging task because only the FFT-based autocorrelation check and cluster-subtraction need to be added (Müller, 18 Sep 2025). These values should be understood as part of a review-level implementation sketch rather than the exact operational defaults of (Müller et al., 22 Apr 2025).

The documented limitations are specific. In the review, clustering via a single lag $\delta$13 is said to work best when the sky emission breaks into quasi-regularly spaced clumps; very diffuse or highly fractal emission may not produce a dominant autocorrelation peak (Müller, 18 Sep 2025). In the deconvolution paper, the need for a Hogbom fallback when $\delta$14 becomes too $\delta$15-like indicates that the adaptive clustered representation is not uniformly advantageous across all residual morphologies (Müller et al., 22 Apr 2025). A plausible implication is that the method is most effective when extended or grouped emission generates stable second-order structure that can be exploited repeatedly across minor-loop updates.

7. Broader significance and relation to other “Autocorr-CLEAN” usages

Within radio interferometry, Autocorr-CLEAN occupies a specific niche: it is neither a wholesale departure from CLEAN nor a purely asymptotic inverse-problem formalism. The ngVLA-era survey argues that hybrid methods of this kind are the most practical near-term route because they retain CLEAN’s speed, robustness, and accumulated pipeline integration while improving performance on extended emission and automation-sensitive workloads (Müller, 18 Sep 2025). In that sense, Autocorr-CLEAN belongs to a broader movement toward methods that preserve legacy workflow but replace pointwise subtraction heuristics with statistics-informed or optimization-informed updates.

The existence of unrelated methods bearing the same name is notable. In LLM evaluation, Autocorr-CLEAN refers to a statistical-inference pipeline that screens findings with pooled tests and false-discovery-rate correction, then confirms them using Chelton effective degrees of freedom and conversation-level block bootstrap; its central quantity is the inflation rate

$\delta$16

which measures how many pooled-significant findings are revoked by cluster-robust correction (Schessl, 15 Apr 2026). In correlation-matrix denoising, Autocorr-CLEAN refers to an adjusted autoencoder that takes the sorted sample eigenvalues together with the noise-level parameter $\delta$17 and learns to estimate the true eigenvalue spectrum, outperforming the Rotational Invariant Estimator over $\delta$18 in the reported experiments (Hayou, 2017).

These uses share an abstract concern with autocorrelation or correlation structure and with some notion of “cleaning,” but they are otherwise methodologically independent. For encyclopedia purposes, therefore, the most precise interpretation of Autocorr-CLEAN is domain-qualified: in radio interferometric imaging it is an autocorrelation-guided CLEAN variant for deconvolution (Müller et al., 22 Apr 2025, Müller, 18 Sep 2025), while in other fields the same label denotes unrelated procedures (Schessl, 15 Apr 2026, Hayou, 2017).