{"id":"ed28402a-1915-408b-84ef-1b5ba03b797f","arxiv_id":"2412.06204","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"KAN-PnP plugs Kolmogorov-Arnold Networks into Plug-and-Play ADMM as single-instance denoisers, claiming convergence and improved super-resolution and joint restoration results.","lead":"This paper proposes KAN-PnP, a Plug-and-Play optimization framework that uses Kolmogorov-Arnold Networks as denoisers trained on a single noisy image, and claims convergence guarantees. The approach is motivated by limited-data imaging, but the main convergence proof has a serious gap and the experimental evidence is small-scale and uncontrolled.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2 is false as stated: Lipschitz continuity plus μ>L_H does not imply contraction; the linear case H(x)=2x, f≡0 diverges under all stated assumptions.","rationale":"The paper's empirical framing is coherent: KANs as single-instance priors and comparison with INR-based PnP are reasonable, and Theorem 1's Lipschitz bound for B-spline KANs is plausible though conservative. However, the central advertised theoretical contribution is the convergence guarantee of Theorem 2, which is the load-bearing support for the abstract's claim that KAN-PnP converges under mild conditions. The reader's weakest_assumption focused on the unverified boundedness of the spline coefficients and the non-existence of R; my analysis goes further and shows that even if those assumptions were satisfied, Theorem 2's proof is invalid and the statement is false, as demonstrated by a linear counterexample satisfying all stated hypotheses. Because the central theorem fails, the paper's correctness risk is high, and the experimental comparisons (which lack code, error bars, and statistical tests, and often show only marginal gains over INR) cannot compensate for the loss of the theoretical guarantee. No change to the reader's REJECT verdict is needed.","tokens_in":15600,"tokens_out":7494,"duration_ms":80422,"concrete_test":"Run the exact ADMM iteration (17) with H(x)=2x, f(x)≡0, μ=3, starting from z^0=1, u^0=0. Since prox_{0/3}(v)=v, the iterates are u^{k+1}=0, x^{k+1}=2z^k, z^{k+1}=x^{k+1}, so x^k=2^k, which diverges while satisfying all assumptions of Theorem 2. This one-line counterexample settles that the theorem is false as stated. For the paper's actual algorithm, additionally compute a certified upper bound on the trained KAN's Lipschitz constant and check whether the chosen μ exceeds it; no such verification currently exists.","verdict_should_be":"REJECT","load_bearing_attack":"The single most load-bearing problem is that Theorem 2's contraction argument is not valid. The proof asserts that composing the Lipschitz operator H with the firmly nonexpansive proximal operator yields an averaged operator with contraction factor strictly less than one whenever μ > L_H. Lipschitz continuity alone with L_H ≥ 1 does not imply nonexpansiveness, and the condition μ > L_H does not make H contractive; no standard fixed-point theorem gives the claimed conclusion. The condition is insufficient even in simple linear cases: take H(x)=2x (L_H=2), f(x)≡0, and μ=3>L_H. The ADMM updates (17) reduce to u^{k+1}=0 and x^{k+1}=2z^k, so every nonzero initialization diverges while all hypotheses (i)-(iii) hold. Thus Theorem 2 is false as stated, not merely unverified. The paper's own Section 2.2 concedes that a regularizer R satisfying (16) need not exist, so the boundedness assumption invoked in the abstract cannot repair the proof. The empirical claim of convergence in 5 iterations therefore has no theoretical backing from the only convergence theorem offered.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes KAN-PnP, a Plug-and-Play ADMM scheme in which a Kolmogorov-Arnold Network trained on a single noisy observation is used as the denoising prior. The authors state two theoretical results: Theorem 1 asserts that the KAN denoiser is Lipschitz continuous when built from B-splines with bounded coefficients and derivatives, and Theorem 2 asserts that PnP-ADMM with a Lipschitz denoiser and penalty parameter μ > L_H converges to a fixed point. The experimental section compares KAN-PnP with INR, FFDNet, UNet, BM3D, and TV priors on super-resolution and joint deconvolution/demosaicing, reporting small PSNR/SSIM improvements over the closest baseline.","tokens_in":15829,"tokens_out":5010,"duration_ms":47973,"significance":"If the claimed guarantees were valid, KAN-PnP would be a relevant contribution to single-instance PnP: it would provide a denoiser with an explicit Lipschitz bound and a fixed-point convergence guarantee under minimal data. The paper also provides an experimental comparison against several single-instance baselines, which is a useful step. However, the central theoretical claim is not sound: Theorem 2 is false as stated, and Theorem 1's conclusion depends on boundedness assumptions that are neither enforced nor verified. The manuscript therefore does not currently support its main advertised contribution.","major_comments":[{"comment":"The proof of Theorem 2 is invalid and the theorem is false as stated. Take f(x)=0 (proper, closed, and convex), H(x)=2x (2-Lipschitz), and μ=3>2. The updates in (17) reduce to x^{k+1}=2(z^k-u^k), z^{k+1}=x^{k+1}+u^k, and u^{k+1}=0, so x^{k+1}=z^{k+1}=2z^k. For any nonzero z^0 the iterates diverge while all hypotheses (i)-(iii) hold, contradicting the claimed convergence to a fixed point. The proof's key assertion that composing a Lipschitz operator with a firmly nonexpansive proximal operator yields an averaged contraction whenever μ>L_H is unsupported; Lipschitz continuity with L_H≥1 does not imply nonexpansiveness, and the condition μ>L_H does not make H contractive. The remark in Section 2.3 claiming convergence under this weaker assumption is therefore not justified by the cited fixed-point theory.","section":"Section 2.3, Theorem 2"},{"comment":"Theorem 1 concludes that the KAN denoiser is Lipschitz continuous only under assumptions of bounded spline coefficients and bounded derivatives, but the training procedure described in Section 3 does not enforce or verify these bounds, and no Lipschitz constant L_H is computed for the trained network. The paper itself concedes in Section 2.2 that the implicit regularizer R defined through (16) need not exist because Hσ may not be firmly nonexpansive. Consequently, the condition μ>L_H required by Theorem 2 is not checked for the algorithm actually run, and the boundedness of the regularisation functional invoked in the abstract is not one of the theorem's hypotheses.","section":"Section 2.2 and Theorem 1"},{"comment":"The experiments use a logarithmically decreasing penalty parameter and dynamically adjusted noise levels, whereas Theorem 2 assumes a constant penalty parameter μ and a fixed denoiser Hσ. The theorem therefore does not apply to the algorithm evaluated in Tables 1, 2, and 4. In particular, the empirical claim of convergence in five iterations has no theoretical backing from the only convergence result offered in the paper.","section":"Section 3, training scheme"}],"minor_comments":[{"comment":"Several entries in Table 1 are implausibly low (for example, FFDNet-PnP on Giraffe at 2× is 4.29 dB and on Koala at 4× is 1.09 dB); these should be explained or corrected because they affect the reported comparison.","section":"Table 1"},{"comment":"Many KAN-PnP results in Table 1 are identical or nearly identical to INR-PnP (e.g., 15.40, 16.21, 16.93, 23.05, 14.79, 17.46/17.47), which weakens the claim that KAN-PnP consistently outperforms INR; the text should acknowledge the closeness of these numbers.","section":"Table 1"},{"comment":"The paper does not provide code, seeds, or a complete specification of the single-shot KAN training procedure, which makes the empirical results difficult to reproduce.","section":"Reproducibility"},{"comment":"Equation (5) uses λR(Hσ(z)) while the algorithm only uses Hσ; the paper states this is conceptual, but the notation should be flagged more clearly as an interpretation rather than a formal objective, because otherwise it conflicts with Section 2.2's admission that R need not exist.","section":"Notation"}],"recommendation":"reject","confidential_remarks":"The central convergence theorem is false as stated, and the proof sketch does not use the hypotheses of the cited PnP literature correctly. Since the advertised contribution is precisely the theoretical convergence guarantee, this is a load-bearing error rather than a local gap. The empirical comparisons may be of interest, but the current manuscript would need substantially different theoretical claims and a re-run of the experiments under the corrected assumptions before it could be considered further."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about arXiv:2412.06204. First, the central convergence claim, Theorem 2, is not just unproven; it is false. Second, despite that, the paper has a legitimate kernel of an idea — using a KAN as a single-instance denoiser inside PnP — and the experiments, while thin, are honest enough to warrant a careful look if the theory is fixed.\n\nWhat is new: the combination of KANs with PnP for single-shot inverse problems. That is a reasonable extension of the authors' earlier INR-based work [27], and the paper gives a straightforward Lipschitz bound for the B-spline KAN architecture. The empirical study covers super-resolution and joint demosaicing/deconvolution, and the ablation over KAN basis functions is a nice touch.\n\nThe soft spot is load-bearing. Theorem 2 claims convergence of PnP-ADMM whenever H is Lipschitz with constant L_H and the penalty parameter μ > L_H. The proof sketch asserts that composing a Lipschitz operator with a firmly nonexpansive proximal operator gives an averaged contraction. That is false. A concrete counterexample: H(x)=2x (L_H=2), f≡0, μ=3. All assumptions hold, yet the iterates double each step and diverge for any nonzero initial point. Lipschitz continuity alone with L_H≥1 does not imply nonexpansiveness, and μ>L_H does not make H contractive. The paper itself concedes in Section 2.2 that the implicit regulariser R need not exist, so the boundedness condition cannot repair the proof. The assumptions are also not verified for the trained network — they never compute L_H or enforce boundedness.\n\nThe experiments are a mixed bag. KAN-PnP usually beats or ties INR by a hair (e.g., 25.82 vs 25.80, 27.07 vs 27.06), and some entries are identical. There are no error bars, no code, and parameter counts differ by design. The comparison in Table 4 uses 5 iterations for KAN-PnP against 100–200 for BM3D/DIP-PnP, which is not apples-to-apples.\n\nWho this is for: someone working on single-instance priors for inverse problems might find the empirical idea worth exploring, but the theoretical guarantee should be ignored until it is corrected. The paper needs major revision with a real fixed-point argument (or a more modest claim, e.g., convergence under a nonexpansive denoiser with L_H≤1, which the authors mention as future work).\n\nMy recommendation: send it to peer review, because the idea is novel enough to deserve expert scrutiny, but the current form would not survive — the theorem must be fixed or removed. I would not cite it as it stands.","headline":"The paper's central convergence theorem is false as stated, so the current version cannot be accepted; the KAN-as-single-shot-prior idea is worth revisiting after a serious fix.","tokens_in":16396,"tokens_out":3756,"would_cite":false,"duration_ms":35851,"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":"The paper claims KAN denoisers enable Plug-and-Play inverse problems from a single noisy instance, with provable Lipschitz continuity and fixed-point convergence.","keywords":["Inverse Problems","Plug-and-Play Methods","Deep Denoiser Priors","Kolmogorov-Arnold Networks","Single-Instance Prior","Super-Resolution","Convergence Analysis","ADMM"],"falsifier":"Measure the actual Lipschitz constant of the trained KAN denoiser used in the experiments (for instance, by computing the maximum singular value of the Jacobian over the test images) and compare it with the ADMM penalty parameter $\\mu$; if $L_H \\ge \\mu$ in any reported configuration, Theorem 2's premise fails and the claimed fixed-point guarantee does not cover that run.","tokens_in":1833,"feed_emoji":"🖼️","tokens_out":2801,"duration_ms":118808,"temperature":0.7,"pith_summary":"The paper introduces KAN-PnP, a Plug-and-Play optimization scheme that uses a Kolmogorov-Arnold Network (KAN) as the denoiser, trained on only the single noisy observation that is being restored. The central claim is that this single-instance prior removes the need for large training datasets while still giving stable, converging reconstructions. To back this, the authors prove that a KAN built from B-spline layers is Lipschitz continuous, and they argue that when the data fidelity term is convex and the ADMM penalty parameter exceeds the denoiser's Lipschitz constant, the iterates converge to a fixed point. Their experiments on super-resolution and joint demosaicing/deconvolution report that KAN-PnP matches or outperforms single-shot baselines such as INR, Noise2Self-FFDNet, and Noise2Self-UNet, often in as few as five iterations.","feed_headline":"KAN denoisers solve inverse problems from one noisy image","feed_subtitle":"Trains on the single noisy image itself, converges in five ADMM iterations, and beats dataset-trained baselines.","key_machinery":"The central object is the KAN denoiser $H_\\sigma$, a composition of layers in which each layer computes $x_{l+1,j} = \\sum_i \\phi_{l,j,i}(x_{l,i})$ for univariate B-spline functions $\\phi_{l,j,i}$ with bounded derivatives. The key estimate is Theorem 1: each spline has Lipschitz constant bounded by $C G^{k-1}$ (with $G$ grid points and smoothness order $k$), so the full network is Lipschitz with constant equal to the product of layer-wise maxima of these sums. This Lipschitz bound is what Theorem 2 feeds on: viewing the PnP-ADMM updates as a fixed-point iteration of a composed operator, the paper argues that a convex data fidelity term plus a penalty parameter $\\mu > L_H$ makes the composition an averaged contraction, so the iterates converge by the Banach fixed-point theorem. The architecture's smooth B-spline interpolation also supplies the inductive bias that is claimed to make single-shot training avoid overfitting to noise.","core_discovery":"The paper's central discovery is that Kolmogorov-Arnold Networks, whose layers sum univariate B-spline functions of their inputs, are naturally well-conditioned as Plug-and-Play denoisers in the single-shot regime. Because each spline has bounded derivatives, the whole network has a Lipschitz constant bounded by the product of per-layer sums of spline bounds, so the denoiser is stable under input perturbations. The authors then show that PnP-ADMM with this denoiser converges to a fixed point $(x^*, z^*, u^*)$ satisfying $x^* = H_\\sigma(z^* - u^*)$, $z^* = \\operatorname{Prox}_{f/\\mu}(x^* + u^*)$, and $x^* = z^*$, provided $f$ is proper, closed, and convex and the penalty parameter satisfies $\\mu > L_H$. Empirically, the method achieves the best or tied PSNR/SSIM on super-resolution (2x, 4x, 8x) and on the joint demosaicing-plus-deconvolution task across datasets including Raccoon, Fractals, Wolf, and Butterfly, while converging in five ADMM iterations.","pith_inferences":["A practical follow-up would certify the convergence condition by measuring or upper-bounding the trained KAN denoiser's Lipschitz constant and checking it against the ADMM penalty before deployment.","The paper's five-iteration convergence is reported with a scheduled noise-level and penalty descent; varying that schedule while fixing the denoiser family would isolate how much of the speed comes from the KAN prior itself.","Because the single-instance prior trains on the noisy observation itself, it necessarily absorbs some noise statistics; testing KAN-PnP under structured or non-Gaussian noise would clarify whether the B-spline smoothness bias is what prevents overfitting to noise.","The same Lipschitz-based framework should transfer to other forward operators (e.g., nonlinear phase retrieval), since only the proximal data-fidelity step changes, providing a ready test of generality."],"forward_implications":["A denoiser trained on the single image being restored can replace dataset-trained denoisers in PnP, so inverse problems in data-scarce domains such as medical or scientific imaging become tractable without large training collections.","Because the method converges in about five ADMM iterations, per-image computational cost is low, making interactive or on-device restoration feasible.","The explicit Lipschitz bound gives a design rule for KAN denoisers: by choosing grid points, spline smoothness, and widths, practitioners can target $L_H \\le 1$ (non-expansiveness) for even stronger convergence guarantees.","The same single-instance prior handles both pure super-resolution and joint tasks like demosaicing plus deconvolution, indicating the mechanism is not tied to one forward operator.","Theorem 2's fixed-point characterization provides a checkpoint condition $x^* = z^*$ that can be monitored in practice to detect convergence without knowing the implicit regularizer."],"supporting_citations":[{"why":"supplies the KAN architecture and the Kolmogorov-Arnold representation-theorem basis for the denoiser.","marker":"[28]"},{"why":"supplies the fixed-point convergence framework that Theorem 2 adapts to the single-instance setting.","marker":"[10]"},{"why":"supplies the proximal-denoiser convergence analysis used alongside [10] for the composed-operator argument.","marker":"[12]"},{"why":"supplies the prior single-shot PnP work (INR-based) that the paper differentiates from and compares against.","marker":"[27]"},{"why":"introduces the Plug-and-Play framework that KAN-PnP operates within.","marker":"[8]"},{"why":"supplies Deep Image Prior, the single-instance baseline motivating the single-shot comparison.","marker":"[25]"},{"why":"supplies the Noise2Self self-supervised denoiser baseline used in the experiments.","marker":"[36]"},{"why":"supplies the FFDNet denoiser baseline used in the experiments.","marker":"[37]"},{"why":"supplies the UNet denoiser baseline used in the experiments.","marker":"[23]"}],"fun_headline_variants":["KAN-PnP: one noisy image trains the denoiser itself","Single-shot KAN denoiser converges in 5 iterations, beats baselines","Kolmogorov-Arnold Networks enable PnP without training datasets","From a single image, KAN-PnP solves super-resolution and deblurring","KAN denoisers: stable, fast, and dataset-free for inverse problems"],"cache_read_input_tokens":18560,"weakest_assumption_plain":"The convergence proof assumes the trained network is smooth enough that its sensitivity to input changes stays below the algorithm's penalty strength, but the paper never measures or enforces that smoothness for the networks it actually runs.","fun_headline_variants_meta":{"raw":{"variants":["KAN-PnP: one noisy image trains the denoiser itself","Single-shot KAN denoiser converges in 5 iterations, beats baselines","Kolmogorov-Arnold Networks enable PnP without training datasets","From a single image, KAN-PnP solves super-resolution and deblurring","KAN denoisers: stable, fast, and dataset-free for inverse problems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000294,"raw_usage":{"total_tokens":1779,"prompt_tokens":1080,"completion_tokens":699,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":696,"completion_tokens_details":{"reasoning_tokens":593}},"tokens_in":696,"tokens_out":699,"duration_ms":6870,"temperature":1.0,"reasoning_tokens":593,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:54:05.874688+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual Lipschitz constant of the trained KAN denoiser used in the experiments (for instance, by computing the maximum singular value of the Jacobian over the test images) and compare it with the ADMM penalty parameter $\\mu$; if $L_H \\ge \\mu$ in any reported configuration, Theorem 2's premise fails and the claimed fixed-point guarantee does not cover that run.","supporting_citations":[{"cited_title":"In: International Conference on Machine Learning, pp","cited_arxiv_id":null,"evidence_quote":"supplies the fixed-point convergence framework that Theorem 2 adapts to the single-instance setting."},{"cited_title":"In: International Conference on Machine Learning, pp","cited_arxiv_id":null,"evidence_quote":"supplies the proximal-denoiser convergence analysis used alongside [10] for the composed-operator argument."},{"cited_title":"Transactions on Machine Learning Research (2024)","cited_arxiv_id":null,"evidence_quote":"supplies the prior single-shot PnP work (INR-based) that the paper differentiates from and compares against."},{"cited_title":"In: 2013 IEEE Global Conference on Signal and Information Processing, pp","cited_arxiv_id":null,"evidence_quote":"introduces the Plug-and-Play framework that KAN-PnP operates within."},{"cited_title":"IEEE Transactions on Image Processing 27(9), 4608–4622 (2018) 22","cited_arxiv_id":null,"evidence_quote":"supplies the FFDNet denoiser baseline used in the experiments."}],"review_version":1}