{"id":"633df1cb-94b7-4a1f-8b96-a3df3c338891","arxiv_id":"2504.16058","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Autocorr-CLEAN fits clusters of CLEAN components to the residual's autocorrelation function and uses them as adaptive, possibly non-circular basis functions, matching Asp-CLEAN convergence and reaching the noise floor five to ten times faster than classical CLEAN in synthetic tests.","lead":"Radio telescopes rely on the CLEAN algorithm to sharpen images, but CLEAN becomes slow on very large data. This paper introduces Autocorr-CLEAN, which learns the typical shape of source structure from the data's autocorrelation function and cleans many pixels per step, reporting up to tenfold speedups.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Speed-up claim depends on unmeasured l/k ratio: Sec 3.4 predicts factor k/(8l), so a 10x gain needs k/l > 80, while only l < k is argued; no l,k,m values are reported.","rationale":"Autocorr-CLEAN is a plausible and novel algorithmic combination: it replaces Asp-CLEAN's per-iteration nonlinear fit with an autocorrelation-driven cluster update, and the pseudocode avoids explicit convolutions in the inner loops. That part of the argument is internally coherent. The load-bearing weakness is the quantitative bridge from complexity to the headline speed-up. Section 3.4's asymptotic comparison is valid only if l is very small relative to k, but the paper supplies no measurement of l/k, and even the verbal claim in Sec 4.3 (l < k) is too weak for the advertised factor. I do not see an internal contradiction in the derivation, and the duplicate Asp-CLEAN timings in Fig A.5 look like a reporting error rather than the central issue. However, until l/k is measured over the benchmark suite, the factor-of-five-to-ten claim remains a conditional prediction. This aligns with the reader's weakest assumption and with the paper's own statement that the advantage may collapse in low-dynamic-range cases; no verdict change beyond the existing CONDITIONAL is needed.","tokens_in":23116,"tokens_out":5802,"duration_ms":58631,"concrete_test":"Instrument the published LibRA implementation (algorithm in Table 1) to log per exterior iteration: k = number of delta components in the basis function ω, l = number of subminor-loop iterations in Loop 2, and m = total exterior iterations. Run the Cygnus A, Hercules A, and M106 benchmarks at the exact settings of Fig. 5 (gain, f, gamma) and report mean and maximum l/k, plus the observed wall-clock ratio to the reimplemented CLEAN baseline. If mean l/k is not approximately 1/80 or below, or if k/(8l) does not roughly predict the observed speed-up, the Sec 3.4 complexity premise fails and the 'up to a magnitude faster' claim is unsupported. Repeat at the 0.1 Jy flux scaling to probe the low-dynamic-range caveat.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec 3.4 derives Autocorr-CLEAN cost as 8kN + (m/k)(5N + 8Nl) ≈ mN(8l/k) under the assumptions l << k << m and no accumulation of residual errors. Compared with CLEAN's Nm, this predicts a speed-up of k/(8l). A 'magnitude' speed-up therefore requires k/l ≳ 80, but the paper only establishes the qualitative statement in Sec 4.3 that 'fewer iterations are performed in the subminor loop than the number of components in the basis function', i.e. l < k. If l/k is 0.1, the predicted gain is only about 1.25x; if l ≈ k, there is no gain at all. No measured values of l, k, or m are reported for any of the three benchmark sources, so the headline factor-of-five-to-ten speed-up is not connected to the complexity argument. The paper itself concedes in Sec 3.1 and Sec 5 that the advantage may collapse when the dynamic range between large and small scales is low, which makes the absence of l/k measurements load-bearing rather than cosmetic.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Autocorr-CLEAN, a multiscale deconvolution algorithm for radio interferometry. The algorithm fits a cloud of CLEAN components to the autocorrelation function of the residual in a subminor loop and uses the powered cloud as a basis function for a CLEAN minor-loop step, thereby avoiding explicit convolutions. The authors claim that all inner-loop operations have linear complexity, that the convergence speed in terms of iterations matches Asp-CLEAN, and that the algorithm is up to an order of magnitude faster than classical Högbom CLEAN. The method is implemented in the LibRA/CASA environment and tested on synthetic VLA observations of Cygnus A, Hercules A, and M106.","tokens_in":23349,"tokens_out":5186,"duration_ms":48080,"significance":"The core idea is attractive and timely: deriving a continuously adapting, potentially non-radially symmetric basis from the residual autocorrelation while keeping every inner-loop operation at linear complexity is a genuine contribution to the CLEAN family. If the speed and fidelity claims hold, the algorithm could be integrated into existing CASA-like pipelines with modest effort. The complexity bookkeeping in Sec. 3.4 is transparent, and the choice of CLEAN and Asp-CLEAN as baselines is well motivated. However, the central speed-up claim is currently not supported by measured complexity parameters, and one figure contains duplicated timing data, so the empirical case needs strengthening.","major_comments":[{"comment":"The headline speed-up is not connected to the complexity analysis. The cost estimate in §3.4 gives Autocorr-CLEAN ≈ mN(8l/k) against CLEAN's Nm, so the predicted speed-up is k/(8l). A ten-fold speed-up therefore requires k/l ≳ 80, but the paper only argues that \"fewer iterations are performed in the subminor loop than the number of components in the basis function ω\" (§4.3), i.e. l < k. No measured values of l, k, or m are reported for any of the three benchmark sources. Because §3.1 and §5 concede that the advantage may collapse at low dynamic range, this omission is load-bearing: without l/k data the abstract's \"up to a magnitude faster\" claim is an extrapolation rather than a demonstrated result. Please report l, k, and m for each test case and verify the k/l ratio, or revise the claimed speed-up accordingly.","section":"§3.4, §4.3, §5"},{"comment":"The timing rows for Asp-CLEAN at gain 0.5 are identical to those for Autocorr-CLEAN at gain 0.5 (21.0, 38.0, 54.0, 74.0, 87.0, 93.0, 99.0, 104.0, 110.0 s). This is implausible because the two algorithms have very different per-iteration costs and indicates a copy-paste error. This affects the gain-robustness comparison discussed in §4.4. Please regenerate the Asp-CLEAN timing data for gain 0.5 or remove that row from the figure.","section":"Fig. A.5"},{"comment":"The empirical support rests on single-run convergence curves without error bars, and the control parameters γ, f, and gain are evaluated on the same synthetic data sets used for the headline benchmarks. Since the central claim is a quantitative speed-up, please provide either multiple noise realisations or an explicit statement that the reported curves are single realisations and that the final parameter choices (e.g., γ = 2, f = 0.1, gain = 0.3) were selected on these same data. Without this information the reader cannot assess how much of the reported speed-up is due to parameter tuning.","section":"§4.3–4.4"}],"minor_comments":[{"comment":"The caption lists \"Cygnus A, Hydra A, Hercules A and M106\", but the text in §4.1 only describes Cygnus A, Hercules A, and M106; please clarify whether Hydra A is a fourth test case or a typo.","section":"Fig. 2 caption"},{"comment":"The update equation for M̃MBB_{j+1} appears to have a self-referential right-hand side; it should presumably read M̃MBB_{j+1} = M̃MBB_j + ... . Please check and correct.","section":"Table 1, Block 6"},{"comment":"The text states \"theoretically several orders of magnitude of speed-up over CLEAN\", while the Abstract and §5 claim \"up to a magnitude\"; please make these statements consistent.","section":"§3.1"},{"comment":"The name \"Högböm\" is a typo for \"Högbom\".","section":"§3.3"}],"recommendation":"major_revision","confidential_remarks":"The duplicated Asp-CLEAN timing row in Fig. A.5 looks like a copy-paste error rather than misconduct, but it must be corrected before the paper can be considered further. The proof-of-concept scope is acceptable for the journal, but the missing l/k measurements are essential for the main speed-up claim; without them the central quantitative conclusion is not yet supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real new algorithm, not a repackaging. Autocorr-CLEAN uses the residual autocorrelation to build a cluster of CLEAN components as a continuously adapting, possibly non-radial basis function, updates it in a subminor loop, and keeps every inner step at linear complexity. That specific combination—autocorrelation-informed clustering plus linear-complexity updates—is new relative to Offringa & Smirnov, Asp-CLEAN, and DoB-CLEAN. The paper is candid that this is proof-of-concept: image-plane only, no interaction with gridding, calibration, or flagging, and no claim of super-resolution.\n\nWhat it does well: the pseudocode in Table 1 is specific enough to reimplement; the complexity bookkeeping in Sec. 3.4 is transparent; the synthetic benchmarks cover several source morphologies and compare against a purpose-built C++ CLEAN baseline in the same library, which is more careful than most algorithm papers. The convergence plots show Autocorr-CLEAN reaching CLEAN-level residuals in roughly a fifth to a tenth of the wall time, and it handles faint diffuse emission well in some examples. The citation pattern is appropriate: the earlier subminor-loop, adaptive-scale, and non-radial-basis work is credited.\n\nWhere it is soft: (1) The complexity argument in Sec. 3.4 predicts a speed-up of order k/(8l), but no measured values of k, l, or m are reported. The paper argues l < k qualitatively; that alone gives no quantitative bound, and the text concedes the advantage can collapse at low dynamic range. The measured wall-clock speed-up is evidence that the ratio is favorable in these examples, but the theory is not yet tied to the data. (2) The headline benchmarks are single runs with no error bars, and gamma, f, and the gain were tuned at least partly on the same synthetic sets used for the headline numbers. Acceptable for a proof-of-concept, but not for a magnitude claim. (3) Fig. A.5 appears to have a copy-paste error: Asp-CLEAN at gain 0.5 shows exactly the same cumulative timings as Autocorr-CLEAN at gain 0.5. That needs fixing. (4) The CLEAN baseline is a custom reimplementation, not production CASA; the fairness case is plausible, but a comparison against an optimized reference would strengthen it.\n\nNet: this deserves a serious referee. I would send it to review and ask for measured k, l, and m per source, repeated runs or run-to-run variability, and a corrected Fig. A.5. If those land, it is a useful pipeline-relevant contribution for the ngVLA/SKA era.","headline":"A genuinely new autocorrelation-driven CLEAN variant with a plausible large speed-up; the claim needs measured k/l ratios and a fix to a duplicated timing figure before it is established.","tokens_in":23892,"tokens_out":3626,"would_cite":true,"duration_ms":34773,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Autocorr-CLEAN aims to deconvolve radio interferometer images up to an order of magnitude faster than classical CLEAN by using the residual's autocorrelation to fit clusters of components in each iteration.","keywords":["radio interferometry","deconvolution","CLEAN","multiscale imaging","autocorrelation function","adaptive basis functions","computational complexity","aperture synthesis"],"falsifier":"A concrete test is to run Autocorr-CLEAN on a source with low dynamic range between bright compact emission and faint extended emission and measure $l$ (subminor-loop iterations), $k$ (components in the basis), and $m$ (iterations needed by classical CLEAN) with $N$ the number of pixels: if $l$ is not much smaller than $k$, or $k$ not much smaller than $m$, the predicted cost $mN(8l/k)$ exceeds CLEAN's $Nm$ and the wall-clock advantage disappears.","tokens_in":22851,"feed_emoji":"📡","tokens_out":12051,"duration_ms":102072,"temperature":0.7,"pith_summary":"Autocorr-CLEAN is a new multiscale variant of the standard CLEAN deconvolution algorithm for radio interferometry. The paper claims it reaches the reconstruction quality and iteration-level convergence of modern multiscale CLEAN methods, such as Asp-CLEAN, while keeping every inner-loop operation as cheap as classical CLEAN: shifts, rescales, and subtractions only, with no convolutions or Fourier evaluations inside the loops. The mechanism is to fit a basis function from the autocorrelation function of the current residual, then subtract many point-like components at once. On synthetic aperture-synthesis observations of three bright radio sources, the authors report reaching the same residual level as CLEAN in five to ten times less wall-clock time, and continuing to recover fainter diffuse emission. This matters because deconvolution is the innermost loop of radio data reduction; a faster minor loop translates directly into handling the data volumes of upcoming arrays.","feed_headline":"Autocorr-CLEAN aims for 10x faster radio-image deconvolution","feed_subtitle":"By fitting a cloud of components from the residual's autocorrelation, each iteration stays as cheap as classical CLEAN.","key_machinery":"The load-bearing object is the autocorrelation function of the residual, $\\Pi = I_D \\diamond I_D$, treated as a second deconvolution problem that can be solved quickly with CLEAN. From its cleaned components the algorithm builds a basis function $\\omega = \\sum_i (\\delta^{\\omega_i})^\\gamma$, a cloud of point components that encodes the global spatial correlation structure of the emission and is point-symmetric by construction; this cloud is the multiscale \"component\" subtracted in the exterior minor loop. Keeping the loops cheap requires a set of precomputed correlation products, including $M_B = \\omega \\star B_D$ and $M_I = \\omega \\star I_D$, which are updated by second-order formulas such as $\\Pi \\leftarrow \\Pi - I_D \\diamond (M \\star \\delta_k) - (M \\star \\delta_k) \\diamond I_D + M \\diamond M$ whenever a component is subtracted. The complexity analysis summarizes the cost as $8kN + (m/k)(5N + 8Nl) \\approx mN \\cdot 8l/k$, where $m$ is the number of classical CLEAN iterations, $k$ the number of components in the basis, and $l$ the number of subminor-loop iterations; the claimed speed-up rests on the assumption $l \\ll k \\ll m$.","core_discovery":"The central claim, stated on the paper's own terms, is that a multiscale CLEAN can be made both adaptive and fast if the shape of the model component is learned from the autocorrelation function of the residual rather than from a fixed dictionary or a local optimization. The authors write the dirty image as $I_D = B_D \\star I$ and its autocorrelation as $\\Pi = I_D \\diamond I_D$, which obeys $\\Pi = (B_D \\diamond B_D) \\star (I \\diamond I)$; so the autocorrelation problem can itself be solved with CLEAN. A subminor loop CLEANs this autocorrelation with the point-spread-function autocorrelation and collects the resulting delta components into a basis function $\\omega = \\sum_i (\\delta^{\\omega_i})^\\gamma$, with $\\gamma = 2$ as a Gaussian-motivated default. Each exterior minor-loop iteration then performs one MS-CLEAN step with this continuously changing, possibly non-radially symmetric basis function, subtracts many components at once, and updates all autocorrelation products analytically to second order, avoiding explicit convolutions. The paper reports that the number of iterations needed matches the adaptive-scale Asp-CLEAN algorithm, while every iteration stays at linear complexity; on the synthetic image-plane benchmarks the residual reaches CLEAN's level in five to ten times less time, and the reconstruction quality matches or improves on the comparison multiscale methods for very diffuse emission.","pith_inferences":["The recursive second-order update of the autocorrelation products is essentially a low-rank update of a residual-covariance estimate; the same update trick could accelerate other matching-pursuit or sparse-imaging methods that subtract localized components while keeping a global correlation model.","The authors treat $\\gamma$ as a fixed scale bias; an extension they do not explore is to make $\\gamma$ adaptive to the measured convergence rate per scale, which could reduce sensitivity to control-parameter choices at low dynamic range.","Because the benchmarks are image-plane only, a natural next test is end-to-end runs that include gridding, self-calibration, and flagging; the relative speed-up may shrink if the minor loop is no longer the dominant cost, or grow if fewer major loops are needed.","The idea is not tied to radio interferometry: any deconvolution problem with a known point-spread function and a residual whose autocorrelation changes slowly could use the same component-clustering heuristic."],"forward_implications":["If the speed-up holds on real data, the deconvolution minor loop will no longer dominate pipeline runtime: the same residual level is reached five to ten times faster than with classical CLEAN, and fainter diffuse emission is recovered.","Iteration counts match Asp-CLEAN without its per-iteration nonlinear optimization, so modern multiscale reconstruction quality becomes available at near-classical per-iteration cost.","Because the basis function is re-derived from the residual autocorrelation at every step, it adapts continuously and can be non-radially symmetric, which lets the algorithm represent flat, elliptical diffuse structures that Gaussian-based multiscale CLEAN misses.","The two-loop structure fits the standard major-loop/minor-loop architecture, so the method can be dropped into existing pipelines with modest changes; the authors report implementing it in the same code library used for the comparisons.","The recommended aggressive control parameters for the gain and the autocorrelation-fitting fraction are a stated risk in poorly calibrated or low-dynamic-range data, where the authors note the underlying assumptions may be violated."],"supporting_citations":[{"why":"Defines the CLEAN algorithm that Autocorr-CLEAN accelerates and against which the speed-up is measured.","marker":"Högbom (1974)"},{"why":"Introduces the precomputed multiscale convolution strategy that Autocorr-CLEAN adapts to its autocorrelation products.","marker":"Cornwell (2008)"},{"why":"Proposes Asp-CLEAN with continuous scale adaptation, the iteration-count benchmark Autocorr-CLEAN claims to match.","marker":"Bhatnagar & Cornwell (2004)"},{"why":"Supplies the subminor-loop component-clustering idea and the multiscale-mask refinement framework.","marker":"Offringa & Smirnov (2017)"},{"why":"Motivates non-radially symmetric basis functions in multiscale CLEAN.","marker":"Müller & Lobanov (2023b)"},{"why":"Provides the efficient adaptive-scale implementation and the switching-to-classical-CLEAN criterion used in Autocorr-CLEAN.","marker":"Hsieh & Bhatnagar (2021)"},{"why":"Offers the approximate PSF-based fitting that makes adaptive-scale CLEAN practical and informs the cost comparison.","marker":"Zhang et al. (2016)"},{"why":"Describes the software environment in which the algorithm was implemented and benchmarked.","marker":"McMullin et al. (2007)"}],"fun_headline_variants":["Autocorr-CLEAN: 10x faster deconvolution, linear-cost iterations","Speed up CLEAN 10x by learning components from the residual's autocorrelation","Adaptive multiscale CLEAN: same per-iteration cost, 10x faster convergence","Autocorrelation-fitted basis: 10x faster CLEAN without costly substeps","Reshape CLEAN components per iteration via autocorrelation: 10x speedup"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The speed advantage collapses unless refreshing the autocorrelation model takes far fewer inner iterations than the number of components subtracted at once, and far fewer components than the thousands of single-component steps classical CLEAN would need.","fun_headline_variants_meta":{"raw":{"variants":["Autocorr-CLEAN: 10x faster deconvolution, linear-cost iterations","Speed up CLEAN 10x by learning components from the residual's autocorrelation","Adaptive multiscale CLEAN: same per-iteration cost, 10x faster convergence","Autocorrelation-fitted basis: 10x faster CLEAN without costly substeps","Reshape CLEAN components per iteration via autocorrelation: 10x speedup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000743,"raw_usage":{"total_tokens":3433,"prompt_tokens":1180,"completion_tokens":2253,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":796,"completion_tokens_details":{"reasoning_tokens":2140}},"tokens_in":796,"tokens_out":2253,"duration_ms":15320,"temperature":1.0,"reasoning_tokens":2140,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:11:35.350570+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test is to run Autocorr-CLEAN on a source with low dynamic range between bright compact emission and faint extended emission and measure $l$ (subminor-loop iterations), $k$ (components in the basis), and $m$ (iterations needed by classical CLEAN) with $N$ the number of pixels: if $l$ is not much smaller than $k$, or $k$ not much smaller than $m$, the predicted cost $mN(8l/k)$ exceeds CLEAN's $Nm$ and the wall-clock advantage disappears.","supporting_citations":[{"cited_title":"& Cornwell , T","cited_arxiv_id":null,"evidence_quote":"Proposes Asp-CLEAN with continuous scale adaptation, the iteration-count benchmark Autocorr-CLEAN claims to match."},{"cited_title":"& Bhatnagar, S","cited_arxiv_id":null,"evidence_quote":"Provides the efficient adaptive-scale implementation and the switching-to-classical-CLEAN criterion used in Autocorr-CLEAN."},{"cited_title":"2016, , 592, A128","cited_arxiv_id":null,"evidence_quote":"Offers the approximate PSF-based fitting that makes adaptive-scale CLEAN practical and informs the cost comparison."}],"review_version":1}