{"id":"07419058-b7e6-4bca-b85a-179b1024f98a","arxiv_id":"2607.10864","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Preconditioning plus multilevel V-cycles accelerates both exact and inexact proximal gradient methods for regularized image deblurring while preserving reconstruction quality.","lead":"The paper combines multilevel coarse-grid corrections with preconditioned proximal gradient methods (ITTA/PNPD) for image deblurring. This yields faster convergence than either technique alone on standard blur-and-noise tests, with an automatic Armijo-based schedule for coarse iterations.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The central claim is an engineering combination of known preconditioners (P=AᵀA+\nu I) with a known multilevel V-cycle framework, plus an Armijo safeguard and an automatic coarse-budget heuristic. The mathematics is standard convex analysis; the only place the argument could fail is if the coarse model loses fidelity so that prolonged corrections cease to be useful descent directions. That is exactly the reader's weakest assumption, and the paper already supplies the theoretical guardrails (first-order coherence, Armijo, boundedness) plus empirical evidence on four representative blurs. Because those guardrails hold for the convolution operators under study, the concern does not land hard enough to move the verdict. The missing public code and J>2 experiments remain practical limitations already noted by the reader, so the CONDITIONAL verdict stays appropriate.","tokens_in":25707,"tokens_out":422,"duration_ms":5065,"concrete_test":"Re-run Example 2 (51\times51 Gaussian) and Example 4 (shake) with J=3 levels instead of J=2, keeping the same m=8 / adaptive budget and \nu=0.1; if the RRE curves of MITTA/MPNPD remain at least as steep as the two-level curves through the first 20 fine-equivalent iterations, the multilevel claim is robust beyond the reported setting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption (Galerkin coarse operator + first-order coherence + Moreau smoothing remaining accurate enough for prolonged coarse directions to stay useful) is the natural soft spot, but the paper already bounds the risk: Lemmas 6, 9–10 and the Armijo line-search (28) guarantee a controlled fine-level decrease whenever Assumption 2 holds, and the four deblurring experiments (Gaussian, defocus, shake PSFs of varying severity) show the expected early RRE drop without quality loss. No internal inconsistency or hidden failure mode for the claimed convolution setting is evident.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper combines left preconditioning of the form P = AᵀA + νI (as in ITTA/PNPD) with a multilevel V-cycle framework (adapted from Lauga et al.) for both exact proximal gradient methods (FISTA/ITTA with framelet ℓ1) and inexact nested primal-dual methods (NPD/PNPD with TV) applied to regularized convex image deblurring. Coarse models are built via Galerkin restriction of the blur, Moreau-envelope smoothing of the nonsmooth term, and a first-order coherence correction; an Armijo line search controls the prolonged coarse correction, and an adaptive rule automatically selects the number of coarse iterations and V-cycles. Convergence is reduced to the type-2 approximation theory of the underlying multilevel framework via a dual-proximal argument (Proposition 11). Four IRtools deblurring examples (Gaussian, defocus, shake PSFs) with cost-normalized RRE/PSNR plots show that the multilevel-preconditioned variants (MITTA, MPNPD and their auto versions) produce substantially faster early error decrease than the corresponding non-multilevel or non-preconditioned schemes while preserving reconstruction quality.","tokens_in":25885,"tokens_out":870,"duration_ms":11830,"significance":"If the claims hold, the work supplies a practical, largely automated acceleration recipe that unifies two previously separate lines of research (preconditioned proximal methods and multilevel forward-backward schemes) for a standard class of imaging problems. The reduction of convergence to existing type-2 theory is clean, the computational-cost accounting (Appendix B) is transparent, and the adaptive Armijo-based selection of m and p removes two free parameters that would otherwise require manual tuning. The numerical evidence on standard test problems is concrete and reproducible with IRtools. The contribution is incremental rather than foundational, but it is useful for large-scale deblurring and opens a clear path to other linear inverse problems once suitable coarse operators are available.","major_comments":[{"comment":"The numerical claims rest on four carefully chosen convolution examples with fixed ν = 0.1 and hand-tuned λ. While the early RRE drop is clear, the paper does not report results for non-convolution operators (e.g., tomography) or for a systematic sweep of ν and noise levels. Because the abstract and conclusion present the method as a “robust … acceleration framework,” at least one additional experiment outside pure deblurring, or a short sensitivity study of ν, is needed to support that breadth.","section":null},{"comment":"Section 5 and Proposition 11 correctly identify the dual sequence as a type-2 approximation, yet the practical implementation freezes the number of inner dual iterations at one. The text asserts that more inner steps bring no improvement, but no supporting table or residual plot is given. A brief verification (even for a single example) that the type-2 residual remains controlled under this fixed budget would strengthen the link between the theory and the reported numerics.","section":null}],"minor_comments":[{"comment":"In Algorithm 8 the gradient of the fine-level objective appears without the Moreau smoothing parameter that is present in Algorithm 5; the notation should be made consistent.","section":null},{"comment":"Figure 1 and the surrounding text describe a four-level scheme, yet all experiments use only two levels; a short remark clarifying that the theory extends while the reported timings are two-level would avoid confusion.","section":null},{"comment":"Several typographical slips remain (e.g., “In subria”, “Techno logy”, missing spaces after commas in the author list). A careful proof-reading pass is needed.","section":null},{"comment":"The choice γ = 1.1 for the Moreau parameter is stated without justification or sensitivity check; a one-sentence remark would be helpful.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is a solid, well-executed combination of the authors’ own PNPD work and the multilevel framework of Lauga et al. Novelty is incremental but genuine; the adaptive Armijo strategy and the consistent use of the same preconditioner on all levels are useful practical contributions. Fit for a numerical-analysis / imaging journal is good. I see no ethical or citation issues."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean, practical marriage of two acceleration ideas that already exist separately: left preconditioning of the form P = AᵀA + \nu I (their own ITTA/PNPD line) with the multilevel proximal framework of Lauga et al. What is new is the integrated package—Armijo line search instead of the earlier decrease condition, an automatic coarse-budget heuristic that halves m until the Armijo test passes, and the consistent use of the same preconditioned solver on every level. The four deblurring examples (Gaussian, defocus, shake) show the expected early RRE drop without quality loss, and Appendix B’s cost accounting is careful enough that the iteration plots are fair.\n\nThe math is standard convex analysis. They reduce convergence to the type-2 approximation theory already proved in [28] via a short dual-argument (Prop. 11); Lemmas 6 and 9–10 plus the Armijo safeguard keep the prolonged coarse directions under control. No internal contradiction appears for the convolution setting they claim.\n\nSoft spots are real but proportional. Everything is two-level only; J > 2 is left as “straightforward.” Free parameters (\nu, \nu_H = \nu/4, \tau-backtracking, eta = 1/8, \rho_k) still need hand-tuning even with the auto heuristic. Code is not released. The Galerkin coarse operator + Moreau smoothing could lose fidelity for non-convolution operators, but the paper never claims those and the Armijo test already bounds the risk on the problems they run.\n\nThis is for people who already solve large deblurring problems with proximal methods and want a faster, more automatic first-order scheme. It is not a new theory of multilevel optimization. I would send it to peer review; a referee can ask for a three-level experiment and public code, but the core claim is already solid enough to deserve that time. Worth reading if you work in the area; I would cite the auto multilevel + preconditioner combination when I next need a fast baseline.","headline":"Solid engineering combination of existing preconditioned proximal schemes with multilevel V-cycles; clear early speed-up on deblurring, incremental but clean and usable.","tokens_in":26459,"tokens_out":504,"would_cite":true,"duration_ms":6562,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65K10","65F22","90C25","68U10"],"pacs":[],"model":"grok-4.5","headline":"Combining left preconditioning with multilevel V-cycle corrections accelerates both exact and inexact proximal methods for image deblurring without losing reconstruction quality.","keywords":["image deblurring","proximal gradient methods","multilevel optimization","preconditioning","forward-backward algorithms","total variation","framelets","ill-posed problems"],"falsifier":"Run the same deblurring experiments with a non-convolutional forward operator (e.g., limited-angle tomography) or with a severe non-Gaussian blur for which the Galerkin coarse operator AH ceases to approximate the fine-level Hessian; if the multilevel-preconditioned methods no longer reduce relative reconstruction error faster than the single-level preconditioned baselines, the central claim fails.","tokens_in":26602,"feed_emoji":"🖼️","tokens_out":701,"duration_ms":6721,"temperature":0.7,"pith_summary":"Proximal gradient methods solve regularized image deblurring problems but converge slowly when the blur is severe or the proximal operator must be approximated. The paper shows that a left preconditioner of the form P = AᵀA + νI can be used at every level of a multilevel hierarchy, so that inexpensive coarse-grid corrections become more accurate and can be transferred back to the fine grid without an unstable extrapolation step. The resulting multilevel-preconditioned schemes (MITTA for framelet ℓ1 and MPNPD for total variation) reduce the relative reconstruction error far more quickly than their non-preconditioned or single-level counterparts while producing images of comparable quality. An automatic Armijo-based rule further removes the need to hand-tune the number of coarse iterations and V-cycles. The practical payoff is that high-quality restorations appear after only a few coarse corrections, opening the door to large-scale deblurring.","feed_headline":"Preconditioning plus multilevel cuts deblurring iterations","feed_subtitle":"Coarse-grid corrections with a left preconditioner restore sharp images faster than standard proximal methods","key_machinery":"The multilevel-preconditioned V-cycle (Algorithms 5–8): a Galerkin coarse operator AH = I_H^h A_h I_h^H together with the linear correction term that enforces first-order coherence, so that a bounded coarse descent direction remains a useful fine-level descent direction after prolongation and an Armijo line search.","core_discovery":"A left preconditioner P = AᵀA + νI, applied consistently at every level of a multilevel V-cycle hierarchy that also enforces first-order coherence via Moreau-envelope smoothing, yields a convergent and substantially faster acceleration of both exact (ITTA) and inexact (PNPD) forward-backward schemes for regularized convex image-deblurring problems.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Multilevel preconditioning accelerates convex image deblurring","Left preconditioner on multilevel cycles speeds proximal deblur","Coarse-grid P plus Moreau smoothing cuts deblurring iterations","Consistent multilevel preconditioning for exact and inexact schemes","V-cycle hierarchy with AᵀA preconditioner restores images faster"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The coarse model built by Galerkin restriction and Moreau smoothing must stay faithful enough that a coarse descent direction is still a useful fine-level correction after it is prolonged; if that fidelity is lost the multilevel speed-up disappears.","fun_headline_variants_meta":{"raw":{"variants":["Multilevel preconditioning accelerates convex image deblurring","Left preconditioner on multilevel cycles speeds proximal deblur","Coarse-grid P plus Moreau smoothing cuts deblurring iterations","Consistent multilevel preconditioning for exact and inexact schemes","V-cycle hierarchy with AᵀA preconditioner restores images faster"]},"model":"grok-4.5","effort":"low","cost_usd":0.003116,"raw_usage":{"total_tokens":990,"prompt_tokens":658,"num_sources_used":0,"completion_tokens":85,"cost_in_usd_ticks":31160000,"prompt_tokens_details":{"text_tokens":658,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":247,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":658,"tokens_out":85,"duration_ms":4282,"temperature":1.0,"reasoning_tokens":247,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T08:41:23.081996+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same deblurring experiments with a non-convolutional forward operator (e.g., limited-angle tomography) or with a severe non-Gaussian blur for which the Galerkin coarse operator AH ceases to approximate the fine-level Hessian; if the multilevel-preconditioned methods no longer reduce relative reconstruction error faster than the single-level preconditioned baselines, the central claim fails.","supporting_citations":[],"review_version":1}