{"id":"b31dcd16-3f8c-4337-a618-5574698c9a3a","arxiv_id":"2412.07718","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper proves novel error bounds showing that a previously proposed closed-form TV proximal approximation can be made arbitrarily accurate by shrinking its scaling parameter.","lead":"This paper analyzes a closed-form approximation of the total variation (TV) proximal operator, proving it is the proximal operator of some convex function, a gradient step on a smoothed version of TV, and an approximate TV proximal whose error can be bounded and controlled by a scaling parameter. The theory is validated on image denoising and sparse-view CT, but the proven error bounds are loose and the practical speedups come at the cost of accuracy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central operator error bounds appear sound, but the explicit claim that Sτ always decreases TV is unsupported: Proposition 2 proves descent only for the Huber-smoothed TV, and the ϵ-subdifferential bound permits an O(τ) increase.","rationale":"I read the paper in good faith. The central theoretical claim, as captured by the Reader's strongest_claim, is the operator-level approximation: Propositions 1–3 give a convex function whose proximal operator is Sτ, an exact gradient step on a Huber-smoothed TV, and quantitative bounds showing Sτ is within O(τ) of the true TV proximal operator. The proofs of Propositions 2 and 3 appear internally consistent, and the dependence of the error constants on n and d is plausible from the W-frame construction. The Reader's primary worry, that operator-level error does not transfer to solution-level accuracy in APGM/ADMM, strikes me as answerable from the paper's own Proposition 3(b): at an APGM fixed point one obtains 0 ∈ ∇g(x) + λ∂_{τϵ2}h(x), and convexity then gives f(x) ≤ f* + λτϵ2, i.e., objective-level control; with strong convexity of g one even gets distance control O(√τ). So I do not treat that as the decisive weakness. The genuinely load-bearing unsupported claim is the monotone TV decrease. It is explicitly asserted twice, and the only argument offered (Proposition 2 plus a Lipschitz bound) proves decrease of the smoothed functional, not of TV. The ϵ-subdifferential inequality leaves an additive O(τ) slack, so the assertion is not a corollary of the paper's results. A computational falsification search can settle whether the claim is actually false; if it is false, that part of the paper must be corrected even though the core approximation theorem would stand. For these reasons the Reader's CONDITIONAL verdict remains appropriate, so I set verdict_should_be to UNCHANGED.","tokens_in":17988,"tokens_out":28141,"duration_ms":265068,"concrete_test":"Run a numerical falsification search: for d = 1, 2 and n ∈ {32, 64, 128}, draw 10^4 random z (Gaussian and piecewise-constant), set τ ∈ logspace(−4, 1, 50), and compute Δ = h(Sτ(z)) − h(z) for both anisotropic and isotropic Sτ with periodic boundary conditions. A single positive Δ falsifies the monotone-decrease claim; if no positive Δ appears in this battery, the claim still lacks the proof the manuscript currently asserts it has.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest load-bearing point is the asserted monotone TV decrease, stated in the Introduction as a contribution and repeated in the Conclusion: \"we demonstrated that the operator consistently reduces the TV function.\" Nowhere is h(Sτ(z)) ≤ h(z) proved. Proposition 2 only establishes Sτ(z) = z − τ∇h̃(z), so with the step size τ = 1/L it guarantees descent of the Huber-smoothed function h̃, not of TV itself. The closest available consequence is Proposition 3(b), which after setting y = z gives h(Sτ(z)) ≤ h(z) − (1/τ)‖z − Sτ(z)‖² + τ4nd²; the positive τ4nd² term prevents a decrease conclusion. Thus the claim is not a corollary of the presented analysis. This does not invalidate Propositions 1–3 or the O(τ) operator-level approximation as τ → 0, so the core approximation claim survives, but the paper's stated contribution and concluding summary overreach. The Reader's separate concern about solution-level transfer is less serious: at an APGM fixed point, Proposition 3(b) yields −∇g(x) ∈ λ∂_{τϵ2}h(x), hence f(x) ≤ f* + λτϵ2, giving objective-level control directly.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes the closed-form operator Sτ(z) = W^T T_{τ2√d}(Wz), previously proposed as an approximation of the total-variation proximal operator, and provides three theoretical results: Proposition 1 shows that Sτ is the proximal operator of some proper closed convex function; Proposition 2 identifies Sτ as one gradient descent step on a Huber-smoothed version of TV; Proposition 3 gives quantitative approximate-proximal bounds, namely Sτ(z) = prox_{τh}(z+δ) with ||δ||₂ ≤ τ ε1 and z − Sτ(z) ∈ τ ∂_{τ ε2} h(Sτ(z)), with explicit constants ε1 = 4d√n and ε2 = 4nd² for both anisotropic and isotropic TV. The authors then validate the operator numerically by embedding it in APGM and ADMM for image denoising and limited-angle CT reconstruction, comparing against the exact TV proximal computed by FPG.","tokens_in":18241,"tokens_out":11031,"duration_ms":107006,"significance":"If the theoretical results stand, the paper provides a useful quantitative justification for a computationally cheap O(nd) TV-proximal approximation, with explicit error constants that can be driven to zero by reducing the proximal scaling parameter. The proofs in the appendix are largely self-contained, cover both anisotropic and isotropic TV, and connect the operator to the Huber smoothing and to standard notions of inexact proximal operators. The numerical validation is appropriate in scope, comparing against an external benchmark (FPG-based exact TV proximal). The main weakness is that one of the paper's stated contributions, the claim that Sτ always decreases the TV function, is not proven and is not a corollary of the presented results; this overstatement appears in the Introduction and Conclusion but does not invalidate the operator-level bounds in Propositions 1–3.","major_comments":[{"comment":"The claim that Sτ 'always decreases the TV function' is stated as a contribution in the Introduction and repeated in the Conclusion ('we demonstrated that the operator consistently reduces the TV function'), but it is not proven anywhere. Proposition 2 establishes Sτ(z) = z − τ∇h̃(z), where h̃ is the Huber-smoothed TV, so the standard descent argument applies only to h̃, not to h itself. The closest consequence derivable from Proposition 3(b) is, after setting y = z, h(Sτ(z)) ≤ h(z) − (1/τ)||z − Sτ(z)||² + τ4nd², which still permits an O(τ) increase. The authors should either prove the TV-descent claim or remove it from the list of contributions and from the Conclusion.","section":"§1, §5"}],"minor_comments":[{"comment":"The proposition statement lists only parts (a) and (b), but the discussion immediately after it refers to 'Proposition 3(b)' and 'Proposition 3(c)'; the proof also ends Part (a) with 'this establishes the desired result of part (b)' and Part (b) with 'Part (c)'. The labels should be made consistent.","section":"§3.2, Proposition 3"},{"comment":"In the isotropic part of the proof, the operator is written as Sτ(z) = z − τD^T∇ϕ_{τ4d}(Dz), but the function being differentiated is ψ from Eq. (12), not ϕ; this should be corrected to ∇ψ_{τ4d}.","section":"§6, Proof of Proposition 2"},{"comment":"The 'exact' TV proximal operator is computed with 50 iterations of FPG, but the manuscript does not report the tolerance achieved by these sub-iterations; since the reference solution may itself be inexact, a brief statement of the resulting accuracy is needed to interpret the small relative errors in Table 1.","section":"§4"},{"comment":"The statement that Proposition 1 guarantees convergence of proximal-based algorithms should be qualified: convergence is to a minimizer of the implied convex function bh, not necessarily to a solution of the TV-regularized problem. The closeness of the final iterates to the exact TV solution is supported only empirically, and the Conclusion should not imply a theoretical solution-level guarantee.","section":"§4.3"}],"recommendation":"minor_revision","confidential_remarks":"The paper is within the scope of the journal and the central operator-level analysis appears sound. The main requested change is local: either prove or retract the unsupported monotone TV-descent claim. I do not see a need for new experiments or a re-derivation of the core bounds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick version. The new content is Proposition 3's explicit error characterization of the closed-form operator Sτ: the perturbed-input and ε-subdifferential bounds with constants ε1=4d√n and ε2=4nd². The proofs of Propositions 1–3 are coherent and the paper is self-contained; the operator comes from the authors' own prior work, and the new analysis does not assume the conclusion. That is real value for people who want to use Sτ inside proximal algorithms and know how far they are from exact TV. The paper also proves Sτ is the proximal operator of some convex function, and identifies Sτ as one gradient step on a Huber-smoothed TV with step size at the Lipschitz limit. Those are clean, useful facts, and the numerical work uses the public SCICO package, which helps reproducibility.\n\nNow the soft spots. The Introduction and Conclusion state that Sτ 'always decreases the TV function' and call that a contribution. The appendix does not prove that. Proposition 2 guarantees descent only for the Huber-smoothed TV; the ε-subdifferential bound in Proposition 3(b) has a positive O(τ) term that prevents a decrease conclusion. So that claim is overreach and should be softened or removed. The core approximation result survives; it is the summary that overshoots.\n\nThe reader worried about transferring operator-level error to solution-level accuracy in APGM/ADMM. That is a legitimate gap in the paper, but it is less severe than it looks: at an APGM fixed point, Proposition 3(b) gives an objective bound, so the transfer can be made. The authors just do not make it, and a referee should ask them to.\n\nThe experiments show the expected trend: smaller γ (hence smaller τ) gives smaller relative error and higher PSNR relative to exact TV. No error bars, and the acceleration numbers are favorable to the method, but the qualitative validation of the τ-dependence is fine. The bounds grow with image size, so they are loose in practice, but that is not a hidden flaw.\n\nThis paper is for people working with proximal algorithms for imaging who want a theoretically grounded closed-form TV proximal substitute. It deserves a serious referee; I would send it out and ask for the TV-decrease claim to be fixed or removed.","headline":"The real contribution is the explicit error bounds for a known closed-form TV-prox substitute, but the paper's 'always decreases TV' claim is not proven and should be softened.","tokens_in":18782,"tokens_out":2648,"would_cite":true,"duration_ms":23922,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C25","65K05","68U10","94A08"],"pacs":[],"model":"deepseek-v4-flash","headline":"Closed-form operator $S_\\tau$ is proved to be a quantitatively controlled approximation of the total-variation proximal operator, the proximal map of some convex function, and one gradient step on a Huber-smoothed TV.","keywords":["total variation","proximal operator","closed-form approximation","Huber smoothing","image reconstruction","computed tomography","convex optimization","proximal algorithms"],"falsifier":"Take a small random image $z$, compute the exact TV proximal operator $\\mathrm{prox}_{\\tau h}(z)$ by an iterative method, and compare it with $S_\\tau(z)$ over a range of $\\tau$: if $\\|S_\\tau(z)-\\mathrm{prox}_{\\tau h}(z)\\|_2>\\tau(4d\\sqrt n)$ for any $z$ and $\\tau$, or if the $\\epsilon$-subdifferential inclusion in Proposition 3(b) fails, the main theorem is false. A second check targets the transfer assumption: run an accelerated proximal gradient method on a denoising problem at fixed $\\tau$ while increasing image size $n$; if the distance from the final iterate to the exact TV solution grows with $n$ rather than tracking the operator error, the algorithmic premise is unsupported.","tokens_in":17814,"feed_emoji":"🖼️","tokens_out":13825,"duration_ms":118305,"temperature":0.7,"pith_summary":"Total variation (TV) regularization encourages piecewise-constant solutions to imaging inverse problems, but its proximal operator, the routine that iterative solvers call repeatedly, has no closed form and normally requires internal iterations. This paper takes a previously proposed closed-form operator $S_\\tau$ and proves that it is the proximal operator of some convex function, that it is exactly one gradient-descent step on a Huber-smoothed version of TV, and that it is an approximate TV proximal operator whose error can be bounded and driven to zero by shrinking $\\tau$. The same operator is tested inside accelerated proximal gradient and alternating-direction-multiplier algorithms on image denoising and limited-angle computed tomography. If the proofs are correct, TV-regularized imaging can be accelerated by replacing the iterative proximal step with a single closed-form operation whose accuracy is set by one scalar.","feed_headline":"Closed-form TV proximal step now comes with error bounds","feed_subtitle":"Swap the iterative TV prox step for one formula; a smaller scale parameter gives closer reconstructions.","key_machinery":"The load-bearing object is the operator $S_\\tau(z)=W^T T_{\\tau 2\\sqrt d}(Wz)$, a shrink-then-project map: the linear map $W=\\frac{1}{2\\sqrt d}\\begin{bmatrix}M\\\\D\\end{bmatrix}$ stacks averaging and finite-difference operators so that $W^TW=I$, soft-thresholding $T_{\\tau 2\\sqrt d}$ acts componentwise for anisotropic TV and per-pixel group for isotropic TV, and $W^T$ returns the result to image space. The proofs work by recognizing $S_\\tau$ as the composition of the proximal map of the scaled analysis norm $\\bar h(u)=2\\sqrt d\\,\\|u^{\\mathrm{dif}}\\|_{p,1}$ with the projection $WW^T$ onto the subspace generated by $W$, then bounding the projection distortion $\\beta=WW^TT_{\\tau 2\\sqrt d}(Wz)-T_{\\tau 2\\sqrt d}(Wz)$. The same construction makes $S_\\tau$ exactly a $\\tau$-step of gradient descent on the Huber-smoothed TV function, because soft-thresholding equals subtracting the gradient of a Huber penalty.","core_discovery":"The paper's central claim is that the closed-form operator $S_\\tau(z)=W^T T_{\\tau 2\\sqrt d}(Wz)$, where $W$ is a normalized union of averaging and difference operators with $W^TW=I$ and $T$ applies soft-thresholding with parameter $\\tau 2\\sqrt d$, is a quantitatively controllable approximation of the exact TV proximal operator. Proposition 3 states that for every $z$, $S_\\tau(z)=\\mathrm{prox}_{\\tau h}(z+\\delta)$ with $\\|\\delta\\|_2\\le \\tau\\epsilon_1$ and $z-S_\\tau(z)\\in\\tau\\partial_{\\tau\\epsilon_2}h(S_\\tau(z))$, where $\\epsilon_1=4d\\sqrt n$ and $\\epsilon_2=4nd^2$, for both anisotropic and isotropic TV. Proposition 2 shows the same operator equals one gradient-descent step on a Huber-smoothed TV function with step size $\\tau$, and Proposition 1 establishes that a proper, closed, convex function $\\hat h$ exists whose proximal operator is $S_\\tau$, even though $\\hat h$ is not given in closed form. Together these results make the approximation error shrink to zero as $\\tau\\to 0$ and underpin the use of $S_\\tau$ inside proximal algorithms.","pith_inferences":["The error constants $\\epsilon_1=4d\\sqrt n$ and $\\epsilon_2=4nd^2$ grow with image size, so for a fixed $\\tau$ the operator-level guarantee weakens as images get larger; keeping a uniform absolute error would require shrinking $\\tau$ with $n$, a consequence the paper does not develop.","Since $S_\\tau$ is a proximal operator of an unknown convex function, one cannot read off which regularizer is actually minimized; one could estimate $\\hat h$ numerically from its Moreau envelope and compare it with TV, a test the paper does not run.","The proof strategy of shrinking coefficients in an analysis frame and bounding the projection error looks transferable to other analysis-sparsity regularizers whose analysis operator satisfies $W^TW=I$, with dimension-dependent constants appearing in the same way.","The Introduction and Conclusion assert that $S_\\tau$ always decreases the TV function, while the appendix proves the decrease only for the smoothed Huber version; checking whether $h(S_\\tau(z))\\le h(z)$ fails on some inputs would settle this asserted monotonicity."],"forward_implications":["TV-regularized denoising and CT reconstruction can skip the inner proximal iterations: replace $\\mathrm{prox}_{\\tau h}$ with $S_\\tau$ and control accuracy through the step size or penalty parameter $\\gamma$, since $\\tau=\\gamma\\lambda$.","Smaller $\\tau$ drives the approximation toward the exact TV proximal operator, giving one operator a continuum from fast rough reconstructions to near-exact ones.","Because $S_\\tau$ is itself the proximal operator of some convex function, proximal algorithms that use it retain convergence guarantees even though the underlying function is not TV itself.","The equivalence to a single gradient step on Huber-smoothed TV connects the operator to smooth optimization theory; the gradient's Lipschitz constant is $1/\\tau$, so the chosen step size sits at the boundary of the safe range.","The experiments show the practical payoff: in the tested APGM CT settings, the closed-form operator reaches the exact TV reconstruction quality while replacing 50 FPG sub-iterations per outer step with one closed-form evaluation."],"supporting_citations":[{"why":"Introduces the closed-form approximation of the anisotropic TV proximal operator that this paper analyzes.","marker":"[9]"},{"why":"Introduces the corresponding closed-form approximation for isotropic TV.","marker":"[10]"},{"why":"Supplies the proximal-average viewpoint used to motivate the approximation.","marker":"[11]"},{"why":"Provides the iterative FPG algorithm used to compute the exact TV proximal operator in the numerical comparisons.","marker":"[5]"},{"why":"Defines the notions of inexact proximal operators whose type-2 and type-3 precision are used to state Proposition 3.","marker":"[49]"},{"why":"Gives the characterization (Corollary 10.c) used to prove that $S_\\tau$ is the proximal operator of some convex function in Proposition 1.","marker":"[62]"},{"why":"Introduces the $\\epsilon$-subdifferential used in Proposition 3(b).","marker":"[67]"},{"why":"Provides the subdifferential sum rule used in Lemma 1 to bound subgradients of the analysis norm $\\bar h$.","marker":"[70]"}],"fun_headline_variants":["TV proximal operator: closed-form approximation with error control","One formula replaces iterative TV prox step, with error bounds","Closed-form TV prox: error provably shrinks with scale","TV denoising gets a fast prox step with rigorous error check","Closed-form approximation of TV prox: accuracy provable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument's load-bearing assumption is that a proximal step that is close to the exact TV one, with error bounds that grow with image size, will still yield final reconstructions close to the true TV solution when used inside an iterative solver; the paper proves the step-level closeness and tests the transfer only numerically.","fun_headline_variants_meta":{"raw":{"variants":["TV proximal operator: closed-form approximation with error control","One formula replaces iterative TV prox step, with error bounds","Closed-form TV prox: error provably shrinks with scale","TV denoising gets a fast prox step with rigorous error check","Closed-form approximation of TV prox: accuracy provable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1252,"prompt_tokens":951,"completion_tokens":301,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":219}},"tokens_in":567,"tokens_out":301,"duration_ms":3526,"temperature":1.0,"reasoning_tokens":219,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:36:03.334117+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small random image $z$, compute the exact TV proximal operator $\\mathrm{prox}_{\\tau h}(z)$ by an iterative method, and compare it with $S_\\tau(z)$ over a range of $\\tau$: if $\\|S_\\tau(z)-\\mathrm{prox}_{\\tau h}(z)\\|_2>\\tau(4d\\sqrt n)$ for any $z$ and $\\tau$, or if the $\\epsilon$-subdifferential inclusion in Proposition 3(b) fails, the main theorem is false. A second check targets the transfer assumption: run an accelerated proximal gradient method on a denoising problem at fixed $\\tau$ while increasing image size $n$; if the distance from the final iterate to the exact TV solution grows with $n$ rather than tracking the operator error, the algorithmic premise is unsupported.","supporting_citations":[{"cited_title":"A parallel proximal algorithm for anisotropic total variation minimization,","cited_arxiv_id":null,"evidence_quote":"Introduces the closed-form approximation of the anisotropic TV proximal operator that this paper analyzes."},{"cited_title":"Minimizing isotropic total variation without subiterations,","cited_arxiv_id":null,"evidence_quote":"Introduces the corresponding closed-form approximation for isotropic TV."},{"cited_title":"Better approximation and faster algorithm using the proximal average,","cited_arxiv_id":null,"evidence_quote":"Supplies the proximal-average viewpoint used to motivate the approximation."},{"cited_title":"Fast gradient-based algorithm for constrained total variation image denoising and deblurring problems,","cited_arxiv_id":null,"evidence_quote":"Provides the iterative FPG algorithm used to compute the exact TV proximal operator in the numerical comparisons."},{"cited_title":"Inexact and accelerated proximal point algorithms,","cited_arxiv_id":null,"evidence_quote":"Defines the notions of inexact proximal operators whose type-2 and type-3 precision are used to state Proposition 3."},{"cited_title":"Proximité et dualité dans un espace hilbertien,","cited_arxiv_id":null,"evidence_quote":"Gives the characterization (Corollary 10.c) used to prove that $S_\\tau$ is the proximal operator of some convex function in Proposition 1."},{"cited_title":"On the subdifferentiability of convex functions,","cited_arxiv_id":null,"evidence_quote":"Introduces the $\\epsilon$-subdifferential used in Proposition 3(b)."},{"cited_title":"Bauschke and Patrick L","cited_arxiv_id":null,"evidence_quote":"Provides the subdifferential sum rule used in Lemma 1 to bound subgradients of the analysis norm $\\bar h$."}],"review_version":1}