{"id":"f7067605-37f3-4041-82c7-64d7bbea5a03","arxiv_id":"2411.13276","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"One-step unrolled analysis and synthesis denoisers in forward-backward plug-and-play, with warm restart, solve the same inverse problems as fully converged denoisers.","lead":"This paper proves that plug-and-play inverse problem solvers using just one step of an unrolled analysis or synthesis denoiser, with a warm start, converge to the same solution as versions that run the denoiser to completion. This matters because one-step plug-and-play is far cheaper, and the proof reveals it is equivalent to classical primal-dual or forward-backward algorithms.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.5's step-size condition is insufficient: for A=0.1, y=0, Γ=2, g=λ|·|, λ=1, τ=1, σ=0.4 (both stated inequalities hold), algorithm (2.7) with L=1 fails to converge to the solution x=0, so the central AD equivalence is false as stated.","rationale":"The reader's weakest_assumption is the warm-restart strategy, but that is an explicit modeling choice stated in the abstract and in the theorem statements, so I do not treat it as a hidden flaw. The more serious soft spot is the step-size condition in Theorem 2.5. The reader did flag an inconsistency between Theorem 2.5 and Appendix A, but framed it as a fixable technical issue. The explicit counterexample above shows the theorem as stated is false, not merely under-proved: an admissible choice of (A, Γ, τ, σ) makes the one-step AD iteration fail to converge to the unique minimizer of (2.6). The synthesis theorem (3.8) does not have this problem because its step-size product τζ appears directly as the FB step-size and the equivalence proof is exact. Thus the analysis theorem is the single load-bearing point. Because the algebraic equivalence and the overall framework may survive with a corrected condition (for example σ<1/||ΓΓ*|| or a coupled condition involving τ and σ), I concur with a CONDITIONAL verdict pending major correction; I do not see reason to move to REJECT. No ad hominem is intended; this is an internal mathematical defect in the stated theorem.","tokens_in":28265,"tokens_out":34110,"duration_ms":319288,"concrete_test":"Implement algorithm (2.7) with L=1 exactly as written for the scalar problem A=0.1, y=0, Γ=2, gλ=λ|·| (λ=1), with τ=1, σ=0.4, starting from (x0,u0)=(1,0). Run for 10^4 iterations and record (x_k,u_k); the observed non-convergence (a 2-cycle away from x=0) disproves Theorem 2.5 as stated. Then rerun the same experiment with σ=0.2 (which satisfies σ<1/||Γ||²) and verify convergence to x=0, confirming that the step-size condition, not the algebraic equivalence, is the defect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central analysis-denoiser equivalence (Theorem 2.5) has an insufficient step-size condition. Consider the scalar instance A=0.1, y=0, Γ=2, gλ=λ|·| with λ=1, and choose τ=1, σ=0.4. These satisfy the theorem's assumptions: 0<τ=1<2/||A||²=200 and 0<σ=0.4<2/||Γ||²=0.5. Problem (2.6) becomes min_x 0.005x²+2|x|, whose unique solution is x=0. With L=1, algorithm (2.7) reads v_k=0.99x_k, u_{k+1}=clip(0.792x_k−0.6u_k,−1,1), x_{k+1}=0.99x_k−2u_{k+1}. Starting at (x0,u0)=(1,0), it generates x: 1, −0.594, 1.303, −0.710, 1.297, −0.716, ... and does not converge to 0; u oscillates between ±1. Hence the stated assumptions do not guarantee convergence, and Theorem 2.5 is false as stated. The proof maps to a Loris-Verhoeven scheme but overlooks that the inner dual contraction (σ||Γ||²<2) and the outer FB step must satisfy a coupled stability condition; the paper's own Appendix A gives the stricter condition σ<1/||ΓΓ*||, which would exclude this counterexample. Since Theorem 2.5 is one of the two pillars of the 'one step = infinite steps' claim, this is load-bearing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies forward-backward plug-and-play (FB-PnP) algorithms in which the proximity operator is replaced by a sub-iterative, unrolled Gaussian denoiser built from a fixed analysis or synthesis dictionary. For the analysis denoiser, the paper claims that with one sub-iteration and a warm restart the FB-PnP algorithm is equivalent to the Loris-Verhoeven primal-dual algorithm and converges to a solution of (2.6); for the synthesis denoiser, it claims equivalence with forward-backward iterations on the synthesis sparse-coding problem. The paper also analyzes a Moreau-envelope-smoothed version of the problem under a sufficient-decrease condition, and presents numerical experiments on compressive sensing and deep dictionary learning image restoration.","tokens_in":28739,"tokens_out":13345,"duration_ms":128760,"significance":"If correct, the paper would provide a clean and practically valuable statement: a single inner sub-iteration, warm-started from the previous outer iterate, can yield the same asymptotic solution as fully converged denoisers in FB-PnP. This would substantially simplify training and analysis of shallow unrolled PnP networks. The synthesis part of the paper is solid: Proposition 3.1 and Theorem 3.2 give a correct equivalence between the synthesis denoising problem and a proximity operator, and Theorem 3.8 correctly identifies the L=1 synthesis update as an FB iteration in the sparse-code domain. The algebraic proofs are direct and do not rely on fitted parameters, which is a strength. However, the central analysis-denoiser result, Theorem 2.5, has an insufficient step-size condition and is false as stated, and Theorem 4.3 has an unproven coercivity assertion. These issues affect load-bearing claims, so the paper cannot be accepted in its current form.","major_comments":[{"comment":"The stated convergence condition 0 < σ < 2||Γ||_S^{-2} is insufficient for the claimed result. Take A = 0.1, y = 0, Γ = 2, g_λ = λ|·| with λ = 1, τ = 1, and σ = 0.4. These values satisfy both conditions in Theorem 2.5 (τ = 1 < 200 and σ = 0.4 < 0.5), but algorithm (2.7) with L = 1 generates the iterates x_{k+1} = 0.99 x_k − 2 clip(0.792 x_k − 0.6 u_k, −1, 1), u_{k+1} = clip(0.792 x_k − 0.6 u_k, −1, 1). Starting from (x_0, u_0) = (1, 0) this yields x_1 = −0.594, x_2 = 1.303, x_3 = −0.710, x_4 = 1.297, ..., which does not converge to the unique minimizer x = 0 of 0.005 x^2 + 2|x|. The proof invokes the Loris-Verhoeven theorem, but the condition for that theorem, as correctly stated in the paper's own Appendix A, is σ < 1/||ΓΓ*|| (which equals 0.25 in this example), and it excludes the counterexample. Therefore Theorem 2.5 is false as stated; the step-size condition must be corrected to match the Loris-Verhoeven condition, and Table 2 must be updated accordingly. This is load-bearing because Theorem 2.5 is the basis for the one-step/infinite-step equivalence for the analysis denoiser.","section":"Section 2.2, Theorem 2.5 and Table 2"},{"comment":"In the proof of Theorem 4.3, after deriving ||∇h(x_k)|| → 0, the paper states 'Since h is convex, coercive (see [4, Cor. 11.16 & 11.17])' and concludes that (x_k) is bounded and converges to a minimizer. However, no coercivity assumption on f or g_λ appears in the theorem statement, and the Moreau envelope of a non-coercive function need not be coercive (for instance, if f = 0 and g_λ is an affine function on a subspace). Without coercivity, ∇h(x_k) → 0 does not imply boundedness of (x_k), so the cluster-point argument is unsupported. The authors should either add an explicit coercivity assumption on the objective in (4.4) or prove that the Lyapunov descent inequality implies boundedness of the iterates. This gap affects the convergence result for the smoothed problem, which is one of the main contributions.","section":"Section 4, Theorem 4.3"}],"minor_comments":[{"comment":"The statement after (3.26) says convergence holds 'if 0 < τ ζ < 2||AD||_S^2' but the exponent should be negative, i.e., 0 < τ ζ < 2||AD||_S^{-2}, matching the theorem statement and Table 2.","section":"Section 3.3, proof of Theorem 3.8"},{"comment":"The text claims that 'the trajectory per iteration is the same independently of the value of L' for the analysis denoiser and 'fairly the same' for the synthesis denoiser. The theorems only guarantee identical limit sets, not identical iterates for different L; this phrasing should be clarified to avoid overstating the experimental evidence.","section":"Section 5.1, Figures 1 and 2"},{"comment":"In the displayed equation (3.15), the set F^S_λ(D) is defined with a superscript S and the text after the equation refers to 'F S λ (D)'; the subscripts and superscripts should be made consistent for readability.","section":"Section 3.1, Example 3.3"},{"comment":"The Loris-Verhoeven algorithm in Appendix A states the condition σ < 1/||Γ*Γ||, but the main text Theorem 2.5 uses the different bound 0 < σ < 2||Γ||_S^{-2}. These discrepancies should be reconciled in the revised manuscript, since the attached proof of Theorem 2.5 relies on the appendix condition.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The counterexample to Theorem 2.5 is decisive and the theorem must be corrected before publication. The fix is local—the step-size condition needs to match the Loris-Verhoeven condition—so I do not recommend rejection. The synthesis section and the algebraic-equivalence viewpoint are valuable. The gap in Theorem 4.3 about coercivity should also be addressed; it may be fixable with a modest additional assumption. Overall the paper has a strong core idea but currently overstates the guarantees."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper shows that a single sub-iteration of an unrolled analysis or synthesis denoiser, with warm restart, lands on a known primal-dual or FB iteration. That is a genuinely useful observation, and the algebraic derivations are correct. But Theorem 2.5, one of the two pillars of the one-step-equals-infinite-steps claim, states a step-size condition that is insufficient. I checked the stress-test counterexample: A=0.1, y=0, Gamma=2, g=abs, lambda=1, tau=1, sigma=0.4. Both stated inequalities hold (tau<2/||A||^2 and sigma<2/||Gamma||^2), yet the iterates oscillate and do not converge to the unique solution x=0. The proof maps the algorithm to Loris-Verhoeven but then applies the wrong convergence condition. The paper's own Appendix A gives the correct condition, sigma<1/||Gamma Gamma^*||, which excludes the counterexample. Since Theorem 2.5 is load-bearing, this is a real flaw. It is fixable by restating the condition to match Appendix A and the LV theorem. What is genuinely new: the L=1 equivalences for both analysis and synthesis (Theorems 2.5 and 3.8) and the Moreau-envelope result (Theorem 4.3) are new statements, even though some intermediate propositions are known. The synthesis equivalence (Theorem 3.8) checks out; the derivation is clean and the condition tau zeta < 2/||AD||^2 is correct. The paper is honest about what is known, and the numerical experiments are reproducible (code on GitHub). Soft spots, in proportion: Theorem 2.5 is the big one. Theorem 4.3's proof asserts coercivity of h by citing [4, Cor. 11.16 & 11.17], but the assumptions do not guarantee it; for example, if g~=0, the Moreau envelope is not coercive. That is a smaller gap, fixable by adding a coercivity assumption on f~ or on the smoothed objective. The numerical section uses convergence loosely for intermediate L, a minor wording issue. Citation pattern is fine; self-citations are contextual and the relevant unrolling/PnP literature is covered. Who it is for: researchers working on PnP convergence, unrolled optimization, or efficient dictionary-based denoisers. The paper deserves a serious referee, but not as is. I would send it to review with a clear request to correct the step-size condition in Theorem 2.5 and patch the coercivity step in Theorem 4.3. The core idea is sound and likely right after those fixes.","headline":"Useful idea, but the central analysis theorem has a step-size condition that is too weak; a valid counterexample exists and the paper's own appendix supplies the fix.","tokens_in":781,"tokens_out":892,"would_cite":false,"duration_ms":88640,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C59","65K10","68T07","68U10","94A08"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single inner denoising step provably converges to the full PnP solution","keywords":["Plug-and-Play","forward-backward algorithm","unrolling","analysis denoiser","synthesis denoiser","primal-dual algorithm","warm restart","dictionary learning"],"falsifier":"On the compressive-sensing toy example of Section 5.1, run the analysis PnP algorithm (2.7) with $L=1$ and the same fixed dictionary, once with warm restart and once with $u_k$ reset to zero at every $k$, then compare both limits to the reference solution obtained with $L=10^4$. If the zero-restart version converges to the same reference solution, the claimed dependence on warm restart is not necessary; if the warm-restart version does not converge to that reference, the theorem's conclusion is false.","tokens_in":28037,"feed_emoji":"🧩","tokens_out":6001,"duration_ms":56171,"temperature":0.7,"pith_summary":"This paper asks whether a Plug-and-Play forward-backward (FB-PnP) algorithm still converges to the right solution when the denoiser inside it is computed with only a single unrolled sub-iteration instead of being solved to high accuracy. For two dictionary-based denoisers—an analysis denoiser built from dual forward-backward steps and a synthesis denoiser built from forward-backward steps—the authors show that with a warm-restart strategy the one-sub-iteration algorithm is algebraically identical to known primal-dual or forward-backward schemes. Consequently, one step and infinitely many steps of the inner denoiser lead to the same asymptotic solution of the outer inverse problem. The paper also proves that the synthesis denoising problem is itself a proximity operator, and gives convergence results for a smoothed Moreau-Yosida version with any finite number of inner steps. If correct, this removes the need to run many inner denoising iterations to guarantee Plug-and-Play convergence.","feed_headline":"One denoising step can match a fully iterated PnP solver","feed_subtitle":"With a warm start, one inner denoising step provably converges to the same solution as the full solver.","key_machinery":"The central object is the unrolled denoiser. The analysis denoiser is $G^A_{L,\\lambda,v}(u_0)=v-\\Gamma^*\\widetilde{G}^A_{L,\\lambda,v}(u_0)$, where $\\widetilde{G}^A$ is $L$ compositions of the dual-FB layer $T^A_{\\lambda,v}(u)=\\mathrm{prox}_{\\sigma g^*_\\lambda}(u-\\sigma\\Gamma^*(\\Gamma u-v))$. The synthesis denoiser is $G^S_{L,\\lambda,v}(z_0)=D\\widetilde{G}^S_{L,\\lambda,v}(z_0)$, where $\\widetilde{G}^S$ composes $L$ FB layers $T^S_{\\lambda,v}(z)=\\mathrm{prox}_{\\zeta g_\\lambda}(z-\\zeta D^*(Dz-v))$. The proof identifies the case $L=1$ with a warm-started inner variable and shows, by direct algebra, that the outer iteration reduces to either the scaled Loris-Verhoeven primal-dual algorithm or to forward-backward iterations in the sparse-code variable. The surrounding arguments use Fenchel-Rockafellar duality and the infimal post-composition identity $D\\triangleright g_\\lambda=(g^*_\\lambda\\circ D^*)^*$.","core_discovery":"For a fixed dictionary, running FB-PnP with exactly one sub-iteration of an unrolled analysis denoiser, with the dual variable carried across outer iterations, is equivalent to the Loris-Verhoeven primal-dual algorithm, and converges to a minimizer of $\\frac{1}{2}\\|Ax-y\\|^2 + g_\\lambda(\\Gamma x)$. With one sub-iteration of an unrolled synthesis denoiser, the iteration is equivalent to forward-backward on the sparse-code problem $\\frac{1}{2}\\|ADz-y\\|^2 + g_\\lambda(z)$, and its image under $D$ converges to a minimizer of the analysis-form problem with $D\\triangleright g_\\lambda$. Hence, for these dictionary-based denoisers, one-step PnP and fully iterated PnP have the same asymptotic solution.","pith_inferences":["The equivalence suggests that a deep-unrolled PnP network with one layer per outer step can be viewed as a convex optimization algorithm, so its fixed point, not just its trained output, is interpretable; this reading goes beyond the paper's explicit statements.","The same warm-restart algebra may extend to other proximal algorithms, such as Douglas-Rachford or ADMM, when the inner solver is unrolled, which would be a natural testable extension.","The result suggests a practical protocol: instead of tuning the number of inner iterations for accuracy, tune it for speed and rely on warm restart to preserve the solution; this could be checked on larger inverse problems with learned dictionaries.","If the equivalence holds for learned dictionaries as well as fixed ones, then stability of the PnP iteration may be assessed through the convex problem it implicitly solves, which would connect these unrolled networks to fixed-point analysis of deep equilibrium models."],"forward_implications":["One inner sub-iteration suffices for asymptotic correctness, so PnP implementations can be much cheaper per outer iteration without changing the target problem.","The analysis PnP with one step is exactly a scaled primal-dual algorithm, so convergence-rate and step-size results for Loris-Verhoeven apply verbatim.","The synthesis PnP with one step is exactly forward-backward on the sparse-code variable $z$, so the outer fixed point inherits the FB convergence theory and sparsity structure.","For the smoothed Moreau-Yosida objective, any finite number $L$ of inner steps converges, provided the approximate proximity operator contracts the previous error by a factor $\\alpha_L < 1/\\sqrt{2}$.","In the deep-dictionary experiments, training the analysis denoiser with only one sub-iteration gave reconstruction quality comparable to training with twenty, suggesting single-layer training is sufficient in practice."],"supporting_citations":[{"why":"Defines the scaled Loris-Verhoeven primal-dual algorithm to which the one-sub-iteration analysis PnP is shown equivalent, and supplies its convergence theorem.","marker":"[38]"},{"why":"Provides the forward-backward convergence results used to prove convergence when the inner denoiser is solved exactly, and for the synthesis one-step equivalence.","marker":"[20]"},{"why":"Gives the dual forward-backward algorithm whose unrolled layers form the analysis denoiser.","marker":"[17]"},{"why":"Supplies the convex-analysis facts, including Moreau's identity, infimal post-composition, and Fenchel duality, used throughout the proofs.","marker":"[4]"},{"why":"Noted that the synthesis denoising problem is a proximity operator, supporting Proposition 3.1.","marker":"[12]"},{"why":"Reformulates composite convex optimization through infimal post-composition, supporting the synthesis proximity and problem-equivalence statements.","marker":"[8]"}],"fun_headline_variants":["One denoiser step matches full PnP solver","Warm start makes one-step PnP exact","Single sub-iteration PnP provably converges","Analysis and synthesis need just one pass","One inner step: same PnP solution as infinite"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central equivalences hold only when the inner variable, $u_k$ for analysis or $z_k$ for synthesis, is carried from one outer iteration to the next as the warm start for the single inner step; if the denoiser were re-initialized from zero at each outer step, the algebraic identities that make one-step PnP match the full solver would break.","fun_headline_variants_meta":{"raw":{"variants":["One denoiser step matches full PnP solver","Warm start makes one-step PnP exact","Single sub-iteration PnP provably converges","Analysis and synthesis need just one pass","One inner step: same PnP solution as infinite"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000273,"raw_usage":{"total_tokens":1649,"prompt_tokens":970,"completion_tokens":679,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":604}},"tokens_in":586,"tokens_out":679,"duration_ms":6752,"temperature":1.0,"reasoning_tokens":604,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:39:07.891468+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the compressive-sensing toy example of Section 5.1, run the analysis PnP algorithm (2.7) with $L=1$ and the same fixed dictionary, once with warm restart and once with $u_k$ reset to zero at every $k$, then compare both limits to the reference solution obtained with $L=10^4$. If the zero-restart version converges to the same reference solution, the claimed dependence on warm restart is not necessary; if the warm-restart version does not converge to that reference, the theorem's conclusion is false.","supporting_citations":[{"cited_title":"L ORIS AND C","cited_arxiv_id":null,"evidence_quote":"Defines the scaled Loris-Verhoeven primal-dual algorithm to which the one-sub-iteration analysis PnP is shown equivalent, and supplies its convergence theorem."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the forward-backward convergence results used to prove convergence when the inner denoiser is solved exactly, and for the synthesis one-step equivalence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the dual forward-backward algorithm whose unrolled layers form the analysis denoiser."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the convex-analysis facts, including Moreau's identity, infimal post-composition, and Fenchel duality, used throughout the proofs."},{"cited_title":"C HAMBOLLE AND T","cited_arxiv_id":null,"evidence_quote":"Noted that the synthesis denoising problem is a proximity operator, supporting Proposition 3.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reformulates composite convex optimization through infimal post-composition, supporting the synthesis proximity and problem-equivalence statements."}],"review_version":1}