{"id":"0d6b003e-02a5-4aa9-abe0-1f7dc658daf2","arxiv_id":"2607.23347","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Permutation-symmetrized positive-weight aggregation creates trainable denoisers that are globally nonexpansive and make plug-and-play reconstruction a contraction.","lead":"This paper builds a trainable image denoiser that is guaranteed not to stretch differences between images, by symmetrizing learned pixel weights across permutations; plugging it into plug-and-play reconstruction makes the whole iteration contractive. The payoff is certified convergence for learned priors, where earlier Lipschitz-controlled denoisers only enforced stability on training samples.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unstated boundary convention in T[7]: if translations are not circular, π is not a permutation and Theorem 3.1 does not apply to the implementation.","rationale":"The reader's weakest assumption targeted the ICC shortcut, and that is indeed where theory and code must align. However, reading Algorithm 1 closely, the shortcut defines the effective weight for π^{-1} as π^{-1}·w; for any weight vector w and any actual permutation π, the two accumulated terms form a symmetric pair, so the symmetry of K follows without requiring the network itself to satisfy ICC. The more fragile premise is that π is an actual permutation of the finite lattice. T[7] consists of translations; unless the code uses periodic shifts, each translation maps some boundary pixels out of the domain and introduces zeros, destroying the group action, invertibility, transitivity, and the symmetry argument. The paper never states the boundary convention. This is a single, checkable condition on which the central contraction claim depends. If it fails, the method may still work empirically, but it is not the provably contractive algorithm advertised. The reader's CONDITIONAL verdict is therefore unchanged, with this additional condition to verify.","tokens_in":15425,"tokens_out":26480,"duration_ms":240286,"concrete_test":"Inspect the released code for how the permutation π is applied to x and ξ. Specifically, check whether T[7] uses circular shifts (e.g., torch.roll or numpy.roll) or zero-padded shifts (e.g., F.pad with value 0). Then run a small numerical check: on a 16×16 random image x and positive weight vector w, compute the matrix of K_impl by evaluating Algorithm 1 on basis vectors, and verify K_impl(i,j) = K_impl(j,i) for all i,j. If zero padding is used, boundary entries will violate symmetry, demonstrating that Theorem 3.1 does not hold for the implementation as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that Algorithm 2 is contractive and globally convergent — rests on assumptions (A1)–(A5) about G being a group of permutations of the finite lattice Ω. In Section 4, G is instantiated as T[7], a set of translations of radius 7. The paper never specifies how these translations act on the boundaries of a finite image. A translation is a permutation of Ω only under periodic (circular) boundary conditions. If the code uses zero-padded or replicated shifts — common in imaging libraries — each π is not a bijection on Ω, the pullback action (7) is not invertible, and the inverse-consistency shortcut in Algorithm 1 (winv = π^{-1}·w) produces zeroed boundary entries. Lemma 3.3's symmetry proof and Assumption (A4)'s transitivity both fail, so the contraction bound ∥DsymE∥_2 < 1 of Theorem 3.1 does not apply to the evaluated operator. The abstract's 'provably contractive and therefore globally convergent' is thus conditional on an unstated boundary convention.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a plug-and-play reconstruction framework in which the denoiser is a normalized, permutation-aggregated linear operator whose weights are produced by a small CNN. By imposing an inverse-consistency condition (ICC) on the CNN outputs, the unnormalized aggregation operator becomes symmetric; a subsequent symmetrization step yields a symmetric, entrywise nonnegative, stochastic operator Dsym. Under assumptions of positivity, transitivity of the permutation set, and a nonannihilating forward operator, the paper claims that Dsym composed with the proximal operator of the quadratic data-fidelity term is a strict contraction, so the fixed-point iteration converges from any initialization. Experiments report competitive PSNR on deblurring and superresolution benchmarks compared with recently proposed convergent PnP methods.","tokens_in":15686,"tokens_out":13100,"duration_ms":128550,"significance":"If the claims hold, the paper offers a useful structural route to globally nonexpansive trainable denoisers: the Lipschitz control is by construction and does not depend on the training distribution. This is in contrast to sample-based spectral-regularization baselines, and the paper is honest that its guarantee covers the symmetrized operator and the frozen-reference phase. The theoretical development is explicit, with proofs in the appendix and a code link, and the experimental comparisons are against relevant convergent baselines. The main value is the combination of a simple permutation-based architecture with a parameter-independent contractivity guarantee, which would be a meaningful contribution to convergent deep image reconstruction.","major_comments":[{"comment":"The paper states that the translation set T[7] 'satisfies the permutation assumptions' because translations act transitively on the lattice. This is only true if Ω is treated as a torus with periodic boundary conditions. A finite-lattice shift is a bijection only under circular wrap-around; with zero-padded or replicated shifts, each π is not a permutation of Ω, the pullback (7) is not a group action, and the inverse steps in Lemma 3.3 and the transitivity condition (A4) fail. The manuscript must state the boundary convention explicitly and confirm that the implementation uses circular shifts. Otherwise Theorem 3.1 does not apply to the reported experiments.","section":"Section 4, Assumption (A4), Theorem 3.1"},{"comment":"The theoretical operator K in Eq. (12) is defined in terms of the CNN Nθ satisfying the ICC (A3). In Algorithm 1, ICC is instead imposed by defining winv = π^{-1}·w for the inverse permutation and never querying Nθ for π^{-1}. This creates an effective operator that coincides with (12) only if the CNN itself satisfies (15). Since a generic trained CNN will not satisfy ICC as a function on X×X, the proof of Lemma 3.3 applies to the algorithm's effective map, not to the map (12) with an arbitrary Nθ. The paper should define the effective network output, e.g., Ñθ, and state that all theoretical statements are for that effective map, so that the theorem and the implementation refer to the same object.","section":"Section 3, Eq. (15), Algorithm 1"},{"comment":"The contraction guarantee is established for Dsym = φ(D), not for the denoiser D in Algorithm 1. The text itself notes that the normalization in D can destroy symmetry and that ∥D(·;ξ)∥2 can exceed 1. Nevertheless, the standalone denoising results in Figure 2 and Table 4 are reported for Algorithm 1's D, not for Dsym. In addition, the warm-up phase of Algorithm 2 updates the reference ξ, so the contraction theorem does not cover the initial N_warm iterations; the paper acknowledges this but the abstract and contribution statements should consistently scope 'globally nonexpansive denoiser' and 'globally convergent reconstruction' to the symmetrized, frozen-reference operator to avoid overclaiming.","section":"Sections 2, 3, and Algorithm 2"}],"minor_comments":[{"comment":"The claim that T[R] contains no nontrivial order-2 elements depends on the lattice dimensions being larger than 2R. On a torus of side length ≤ 2R, a shift by half the side length is an involution. Since Algorithm 1 skips all involutions, the effective permutation set is smaller than G; please state the size condition and note that the analysis applies to the reduced set.","section":"Section 4"},{"comment":"G is repeatedly called a 'group of permutations,' but T[R] is only inversion-closed and is not closed under composition. The assumptions actually use a set of permutations with inversion closure and transitivity of the generated action. The terminology should be corrected or clarified.","section":"Section 2"},{"comment":"In the proof, the reference to Eq. (29) is ambiguous; the displayed expression for K_{ij} should be numbered or cited accurately. This is a minor presentation issue but helps the reader verify the irreducibility argument.","section":"Appendix, Theorem 3.1 proof"},{"comment":"The notation C^{1/2}DC^{-1/2}(x) is dense and appears before its componentwise convention is fully explained. A short example or an explicit matrix expression would improve readability.","section":"Equation (17)"}],"recommendation":"major_revision","confidential_remarks":"The authors should be asked to confirm, ideally in the text and in the released code, that all translations are circular (periodic). If the implementation uses zero-padded shifts, the main theorem does not cover the evaluated operator. The remaining issues are rigor and scope of the stated guarantees, which should be fixable within a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The construction is the real thing: permutation-paired aggregation plus the inverse-consistency condition gives a trainable denoiser that is globally nonexpansive for all parameter values, not just on training data. That closes a real gap in the PnP convergence literature, where prior learned denoisers only enforced Lipschitz control empirically. The proof outline is coherent: ICC makes K symmetric, the symmetrization recipe makes it a symmetric stochastic operator, and Perron-Frobenius plus A5 gives strict contraction. The warm-up/freeze-reference split in Algorithm 2 is an honest acknowledgment that the guarantee applies only after the reference is frozen.\n\nThe soft spots are real but addressable. The load-bearing one is the boundary convention. Section 4 instantiates G as T[7], translations of radius 7 on a finite lattice. A translation is a permutation only under circular boundary conditions. The paper never says that. If the code uses zero-padded or reflected shifts, the pullback is not invertible, the ICC shortcut winv = π^{-1}·w produces zeroed boundary entries, and Lemma 3.3 no longer holds. The contraction theorem then does not apply to the evaluated operator. This is not a mathematical error in the theorem; it is an implementation-theory alignment gap, but it is central. The fix is simple: state that translations are circular, or use a genuine permutation group such as a cyclic group.\n\nSecond, the abstract says 'provably contractive and therefore globally convergent' without mentioning the warm-up. The paper does admit in Section 5 that the guarantee applies only after the reference is frozen, so the abstract overstates. Third, the empirical section is thin: 10 images, no variance, and the symmetrization step in (17) is computationally expensive and not fully detailed. Also, the code is not auditable from the text.\n\nNone of these are fatal. The central construction is new and the proof likely holds once the boundary condition is nailed down. This deserves a serious referee, not a desk reject. I would ask the authors to clarify the boundary handling, make the code available for inspection, and qualify the convergence claim.","headline":"Genuinely new mechanism for global nonexpansivity, but the boundary convention in the implementation must be clarified before the contraction guarantee is believable.","tokens_in":16179,"tokens_out":2265,"would_cite":true,"duration_ms":20524,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that a permutation-based weighted-aggregation denoiser can be made globally nonexpansive (Lipschitz constant at most 1) by construction, and that the resulting plug-and-play reconstruction operator is contractive, guarantee","keywords":["nonexpansive denoiser","plug-and-play reconstruction","contractive fixed-point iteration","permutation-based aggregation","inverse-consistency condition","image deblurring","superresolution","convergence guarantee"],"falsifier":"For a fixed trained network, forward operator A, and reference ξ, numerically compute the matrix norm ∥Dsym(·;ξ)(I+ρAᵀA)⁻¹∥₂; the paper predicts it is strictly less than 1, so finding any ξ or network instance with norm ≥ 1—or observing two different initializations converge to different fixed points—refutes the contractivity claim. A code-level check is simpler: replace the permuted-copy weight with a genuine evaluation of Nθ(ξ, π⁻¹·ξ) and repeat the reconstruction; any change in the limiting image indicates Assumption (A3) was not actually satisfied.","tokens_in":15306,"feed_emoji":"🖼️","tokens_out":6775,"duration_ms":64028,"temperature":0.7,"pith_summary":"Plug-and-play image reconstruction treats a denoiser as an implicit regularizer, but the convergence of the fixed-point iteration depends on the denoiser's Lipschitz properties, which existing deep denoisers enforce only on training samples, not globally. This paper constructs a denoiser whose global Lipschitz constant is at most one for every parameter value by keeping all nonlinearity in the prediction of positive weights and making the actual denoising step a linear, normalized aggregation over permuted copies of the image. The authors prove that, under mild conditions on the permutation set, the network's outputs, and the forward operator, plugging this denoiser into half-quadratic splitting makes the reconstruction operator a contraction. Contractivity implies the algorithm converges to a unique fixed point regardless of initialization, and experiments on deblurring and superresolution show performance competitive with softly constrained baselines that offer no such global guarantee.","feed_headline":"Permutation consistency makes denoisers globally nonexpansive","feed_subtitle":"Plug-and-play reconstruction becomes provably contractive, so images converge from any start.","key_machinery":"The central object is the weighted aggregation operator K(x;ξ) = Σ_{π∈G} Nθ(ξ, π·ξ) ⊙ (π·x), normalized element-wise by C(ξ) = Σ_{π∈G} Nθ(ξ, π·ξ), where G is an inversion-closed, transitive set of lattice permutations (in practice, local translations T[7]). The load-bearing identity is the inverse-consistency condition Nθ(ξ, π·ξ) = π·Nθ(ξ, π⁻¹·ξ), which makes K symmetric. The symmetrization φ (defined in Eq. 17) converts the normalized denoiser into Dsym, a symmetric, stochastic matrix with spectral norm 1. Theorem 3.1 then proves that composing Dsym with the proximal operator of the quadratic data-fidelity term yields a contraction, combining a Perron–Frobenius argument for the simplicity o","core_discovery":"For a fixed reference image ξ, the proposed denoiser D(·;ξ) is linear in the input image, while remaining nonlinear in ξ through the weight-predicting network Nθ. By imposing an inverse-consistency condition (ICC) — computing the weight for permutation π⁻¹ as the permuted copy of the weight for π — the aggregation operator K(·;ξ) becomes symmetric for every ξ. A symmetrization step then yields Dsym, a symmetric, entrywise nonnegative, stochastic operator with spectral norm exactly 1, i.e., globally nonexpansive. The main theorem (Theorem 3.1) shows that the reconstruction operator Tξ = Dsym ∘ prox_{ρf} is contractive, meaning ∥Dsym (I+ρAᵀA)⁻¹∥₂ < 1, under assumptions that include transitivit","pith_inferences":["The inverse-consistency construction could be ported to any input-dependent linear aggregation scheme (e.g., non-local means or attention) to obtain certified nonexpansive variants, since the symmetry argument only requires an inversion-closed, transitive permutation structure and positive weights.","A natural testable extension is an adaptive-reference schedule: re-freeze the reference only when the iterate change is small, which might improve adaptation while retaining contraction in practice, although the paper's theorem does not cover the adaptive phase.","The sharpest stress test of the proof is to remove the ICC shortcut and evaluate the network on inverse pairs directly; if the reconstructed images differ materially, the implementation–theory alignment is carrying the guarantee rather than the network itself."],"forward_implications":["With the reference image fixed after warm-up, the reconstruction iteration converges to the same unique fixed point from any initialization, including all-zeros, random noise, the observed image, or the true image.","Because the nonexpansive property is architectural rather than penalty-based, the Lipschitz guarantee holds for all inputs and all parameter values, ruling out divergence on out-of-distribution data where empirically constrained baselines can fail.","The denoiser can be reused across different forward operators without retraining; contractivity requires only that the forward operator be non-annihilating (A e ≠ 0), which holds for standard deblurring and superresolution operators.","On standard CBSD10 benchmarks, the method's PSNR falls within roughly 0.3–0.8 dB of the strongest convergent baselines while providing a strictly stronger, global convergence guarantee."],"fun_headline_variants":["Permutation symmetry yields provably contractive reconstruction","Global Lipschitz bound for trainable denoisers via permutation","Nonexpansive denoiser enables globally convergent image reconstruction","Permutation-constrained denoiser yields provably contractive reconstruction"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof treats the weight for π⁻¹ as a permuted copy of the weight for π via a computational shortcut, an identity that is not intrinsic to a generic trained CNN—if inverse pairs are evaluated separately, or order-2 permutations are included, the symmetry of K and the contractivity conclusion can fail.","fun_headline_variants_meta":{"raw":{"variants":["Permutation symmetry yields provably contractive reconstruction","Global Lipschitz bound for trainable denoisers via permutation","Nonexpansive denoiser enables globally convergent image reconstruction","Permutation-constrained denoiser yields provably contractive reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000618,"raw_usage":{"total_tokens":2682,"prompt_tokens":699,"completion_tokens":1983,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":443,"completion_tokens_details":{"reasoning_tokens":1913}},"tokens_in":443,"tokens_out":1983,"duration_ms":13319,"temperature":1.0,"reasoning_tokens":1913,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T23:41:18.045868+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a fixed trained network, forward operator A, and reference ξ, numerically compute the matrix norm ∥Dsym(·;ξ)(I+ρAᵀA)⁻¹∥₂; the paper predicts it is strictly less than 1, so finding any ξ or network instance with norm ≥ 1—or observing two different initializations converge to different fixed points—refutes the contractivity claim. A code-level check is simpler: replace the permuted-copy weight with a genuine evaluation of Nθ(ξ, π⁻¹·ξ) and repeat the reconstruction; any change in the limiting image indicates Assumption (A3) was not actually satisfied.","supporting_citations":[],"review_version":1}