{"id":"7b4d111f-bcb5-447d-b9da-ff4bdea4d3c5","arxiv_id":"2608.09923","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"An unrolled graph denoiser combining a truncated Taylor expansion with fixed-step conjugate gradient is proven to realize only polynomial graph filters of degree at most K(m-1), a measure-zero subset of its nominal degree budget.","lead":"An unrolled graph-based image denoiser that learns coefficients in a truncated Taylor expansion and in a fixed-step conjugate-gradient solver is proven to realize only polynomial graph filters of bounded degree. The learned coefficients therefore shrink the hypothesis space to a thin subset of an already known filter class, and a direct Bernstein parameterization of that class is cheaper and better conditioned.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The collapse theorem is sound under its stated fixed-step CG scope; the load-bearing assumption is that α_k and β_k are shared scalars, which the numerical parameter count supports but the original implementation must confirm.","rationale":"The central claim of the paper, Theorem 4 plus Proposition 5, is mathematically sound. The induction proof correctly shows that for any fixed scalars {α_k} and {β_k}, the CG iterates remain in the Krylov subspace K_m(A_K, y), and since A_K = t_K(Ψ) is a polynomial of degree at most K, the output is a polynomial in Ψ of degree at most K(m−1). The reachable set is the image of a lower-dimensional parameter space and is therefore measure zero whenever the dimension inequality holds, which it does for the practical orders K=10, m=3. The reader correctly identifies the fixed-step CG assumption as the weakest link: if the coefficients were input-dependent, the map would not be a fixed polynomial and the thinness result would not follow. However, the paper explicitly discloses this scope in Remark 6, and the numerical setup with 17 scalar parameters is consistent with it. The other issues raised by the reader—the abstract overstating the closed-form initialization result, Proposition 12's unstated explicit threshold, and the absence of released code—are presentation and reproducibility concerns that do not threaten the central theorem. They warrant conditional acceptance rather than outright rejection. Therefore the reader's CONDITIONAL verdict remains appropriate, and no verdict adjustment is needed.","tokens_in":7983,"tokens_out":20445,"duration_ms":175646,"concrete_test":"Inspect the released or original implementation of the TSE+CG unrolled network from [3] and confirm that the CG step sizes α_k and β_k are implemented as scalar parameters shared across all input patches, not computed from the residual. If they are input-dependent, check whether the output remains a polynomial in Ψ; if they are shared scalars, the collapse theorem and measure-zero conclusion apply as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 4 and Proposition 5 depend critically on the CG step sizes α_k and β_k being fixed scalars shared across all inputs, i.e. the unrolled regime. If those coefficients were functions of the residual, as in textbook CG, the output x_m would still lie in a Krylov subspace for each input y, but the map y -> x_m would not be a single fixed polynomial P(Ψ); the collapse to a fixed polynomial, and the measure-zero reachable set, would fail. The paper explicitly scopes to the unrolled regime in Remark 6, and the numerical section's parameter count (17 total: 11 for TSE, 1 for s, 3 for α, 2 for β) indicates scalar coefficients. However, the theorem's applicability to the original architecture in [3] rests on that implementation choice. A secondary caveat is that Ψ itself is constructed from the input patch in the experiments, so the end-to-end map is not a fixed polynomial graph filter; the theorem applies per-operator, a point the paper acknowledges but which limits the universality of the title. The proof itself is correct: the induction shows Krylov subspace containment for arbitrary scalar step sizes, and the degree bound K(m−1) follows from polynomial composition.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes a two-module unrolled graph denoiser: a truncated Taylor expansion (TSE) module that forms a system matrix from a graph-Laplacian denoising operator Ψ, followed by m conjugate-gradient steps with learned scalar coefficients. Theorem 4 proves that, for every setting of the learned coefficients, the end-to-end map is q_{m-1}(t_K(Ψ)), a polynomial in Ψ of degree at most K(m-1). Proposition 5 shows that this reachable set has dimension at most K+2m+1 and is therefore a measure-zero subset of P^{K(m-1)} whenever K(m-2)>2m. Proposition 7 derives a closed-form spectral response g_K in the exact-solve limit and shows it is bounded below by 1/(K+1); Proposition 9 gives a Bernstein basis spanning all polynomial filters with a structural nonexpansiveness guarantee; Proposition 12 proves a condition-number lower bound for shift-invariant graph weights. The paper concludes that the unrolled block does not enlarge the polynomial graph-filter class and compares the block against a direct Bernstein parameterization in small image-denoising experiments.","tokens_in":8235,"tokens_out":22815,"duration_ms":205300,"significance":"If its scope qualifications are made precise, this is a significant negative result for the algorithm-unrolling literature: it identifies a broad family of learned TSE+CG modules whose hypothesis space is provably confined to polynomial graph filters, and it quantifies that confinement as measure-zero within the nominal degree budget. The proof of Theorem 4 is a clean Krylov-subspace induction and is correct under the stated fixed-step assumption; Proposition 5's dimension count is valid; and Proposition 9 provides a useful convex baseline. The paper is also honest about its main idealizations, noting in the Fig. 2 caption that the finite-m initialization is far from an exact solve and that the closed-form response is not the m-step response. The numerical experiments are limited but adequate as an illustration. With the overstatements corrected, the result should be of interest to researchers designing and analyzing unrolled graph filters.","major_comments":[{"comment":"The theorem is stated for 'arbitrary step sizes {α_k},{β_k}, learned or otherwise,' and Remark 6 explicitly scopes it to the unrolled regime in which these are constants shared across inputs. The manuscript does not, however, state whether the original architecture in [3] is in this regime. If [3] used textbook CG with residual-dependent steps, the end-to-end map is not a single fixed polynomial and Theorem 4 would not characterize that construction. Please add an explicit statement confirming that [3]'s unrolled CG uses fixed scalar step sizes, or alternatively narrow the introduction and abstract to the architecture defined in Section 2.","section":"§3 (Theorem 4, Remark 6)"},{"comment":"Proposition 7 derives g_K under the condition 'If the CG module solves exactly,' so Eq. (8) is an exact-solve idealization. The abstract nevertheless says 'the realized spectral response is obtained in closed form,' and Section 6 states 'the realized gain is g_10(λ_min)=...' while the Fig. 2 caption acknowledges that at initialization the m=3 CG module is far from an exact solve and 'the response there is not g_K.' This is an internal inconsistency in how the closed form is presented. Please either report the actual degree-K(m−1) response at the initialization used in the experiments, or explicitly label Fig. 1 and the spectral-response paragraph as the exact-solve limit and remove 'realized' from the abstract and the contribution list.","section":"§4 (Prop. 7), §6, Abstract"},{"comment":"The experimental network constructs Ψ from each input patch, so the end-to-end map is not a single fixed polynomial filter; Theorem 4 applies per operator. The conclusion's last sentence acknowledges this, but the title and the abstract's claim that 'the resulting map is a polynomial in the denoising operator' are likely to be read as a statement about a fixed hypothesis class. Please state the per-instance qualification in the abstract and in the numerical section, and clarify that the measure-zero statement of Proposition 5 concerns the reachable set for a fixed Ψ.","section":"§6, Title"}],"minor_comments":[{"comment":"The initialization of {α_k} and {β_k} is not specified, so the ablation is not fully reproducible; please give the default values or state that they are initialized randomly and describe the distribution.","section":"§6"},{"comment":"The caption calls g_K(λ) of Eq. (8) the 'realized response'; it should be qualified as the exact-solve response, not the m-step response, to match Proposition 7 and the Fig. 2 caption.","section":"Fig. 1 caption"},{"comment":"The all-ones vector is written as 1 without definition; please define it near Eq. (3) for clarity.","section":"§2 (Lemma 1)"},{"comment":"The symbols p and q are used both as adjacency indices and as the test-vector index in x_p=(-1)^p; consider a less overloaded notation.","section":"§6 (Prop. 12)"},{"comment":"The sentence 'this paper adds a proof of the correspondence' is slightly ambiguous: the graph-filter/denoiser correspondence already appears in [2], and what is new here is the collapse theorem for the unrolled block. Please rephrase to distinguish the new contribution from the quoted result.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":"This manuscript is a self-contained negative result about the author's own previously published architecture. The self-citation is disclosed and the analysis does not appear circular. The main risks are the overstatement of the closed-form spectral response and the lack of explicit confirmation that [3] used fixed scalar CG steps. If the author can clarify these two points, the paper would be suitable for publication. The Bernstein-baseline comparison is appropriate and strengthens the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central result is correct and worth knowing: an unrolled TSE+CG block, with learned coefficients held fixed across inputs, realizes only a polynomial in the denoising operator Ψ, of degree at most K(m−1). The proof in Theorem 4 is a clean Krylov subspace induction; the dimension count in Proposition 5 shows the reachable set is a measure-zero subset of that polynomial class when K(m−2)>2m. That is a genuine negative result for algorithm unrolling, and it is new relative to the prior work the paper cites.\n\nThe paper does several things well. It is explicit about scope: Remark 6 makes clear the result concerns the unrolled regime with shared scalar step sizes, not textbook CG. The Bernstein parameterization in Proposition 9 is a useful counterpoint, showing a direct convex parameterization spans the same polynomial class and even contains the target operator exactly. The numerical ablation is consistent with the theory: the unrolled and Bernstein arms land within about a decibel of each other at equal parameter counts, and the Jacobian rank observation (12 of 17) fits the dimension bound. The condition number bound in Proposition 12 is an interesting addition, and the paper honestly notes the per-operator nature of the result.\n\nSoft spots are real but not fatal. The abstract states the initialization response is obtained in closed form, but that holds only for an exact CG solve; the paper footnotes the finite-m deviation in the Fig. 2 caption, yet the abstract lacks that caveat. Proposition 12 promises an explicit threshold but the proof only says 'above an explicit O(1) threshold' without giving the formula—a minor promise-reveal mismatch. No code is shipped, so the numerical claims (0.34% condition-number match, rank 12 of 17) cannot be independently checked. Also, in the experiments Ψ is built from the input patch, so the end-to-end map is not a fixed polynomial; the theorem is per-operator. The paper acknowledges this, but the title overstates it.\n\nThe self-citation to [3] is fine; the proof does not assume the conclusion. I share the reader's view that the central theorem is sound and the issues are addressable. This deserves a serious referee. I would send it to peer review, with the expectation that the authors tighten the abstract, supply the threshold formula, and ideally release code.","headline":"A sound negative result for unrolled graph denoisers—worth refereeing despite minor overreach in the abstract and missing code.","tokens_in":8793,"tokens_out":2895,"would_cite":true,"duration_ms":25301,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An unrolled graph-Laplacian denoiser with learned coefficients still realizes only polynomial graph filters.","keywords":["algorithm unrolling","graph Laplacian denoiser","conjugate gradient","polynomial graph filters","Krylov subspace","graph signal processing","image denoising","Bernstein polynomials"],"falsifier":"Run the trained unrolled module on a basis of inputs $y_1,\\dots,y_N$; if any output $x_m$ has a component outside $\\operatorname{span}\\{\\Psi^j y_i:0\\le j\\le K(m-1)\\}$, then Theorem 4 is false.","tokens_in":7753,"feed_emoji":"🧮","tokens_out":7731,"duration_ms":63662,"temperature":0.7,"pith_summary":"This paper proves that a recently proposed unrolled network for graph-based image restoration—a truncated Taylor expansion of the graph-Laplacian denoiser followed by a fixed number of learnable conjugate-gradient steps—cannot learn anything beyond a polynomial in that denoising operator. For every setting of the learned coefficients, the end-to-end map is a polynomial of degree at most $K(m-1)$, and at the orders used in practice the reachable set is a measure-zero subset of that polynomial class, so unrolling confines the hypothesis space rather than enlarging it. The paper also shows that the standard initialization is a leaky denoiser whose spectral response exceeds the intended response throughout the spectrum and approaches a positive floor $1/(K+1)$, and that the operator's condition number forces the truncation order needed for accuracy far above what is used in practice. The confined class is exactly the spectral graph filters for which a direct, convex parameterization has long existed.","feed_headline":"Unrolled denoisers only learn polynomial graph filters","feed_subtitle":"A learnable fixed-step solver adds no capacity beyond direct polynomial graph filters.","key_machinery":"The load-bearing object is the polynomial graph filter, a map of the form $P(\\Psi)$ with $\\Psi$ the graph-Laplacian denoiser and $P$ a polynomial. The argument has three gears: (1) the TSE module forms a polynomial $t_K(\\Psi)$ of degree at most $K$; (2) CG iterates stay in the Krylov subspace $\\mathcal{K}_m(t_K(\\Psi),y)$, so the solver contributes a polynomial of degree at most $m-1$ regardless of the learned step sizes; (3) composing the two bounds the degree by $K(m-1)$, and counting parameters (dimension $K+2m+1$ versus $\\dim\\mathcal{P}_{K(m-1)}=K(m-1)+1$) makes the reachable set thin. A fourth piece, Proposition 12, derives $\\kappa(\\Psi)>1+\\mu$ from a sign test on a one-dimensional trigonometric sum, placing the required truncation order well above practice, and the Bernstein basis supplies a direct parameterization of the whole class with a structural nonexpansiveness guarantee.","core_discovery":"The central discovery is Theorem 4: when the $m$-step CG module is run on $t_K(\\Psi)x = y$ from $x_0 = 0$ with arbitrary fixed step sizes $\\{\\alpha_k\\}$ and $\\{\\beta_k\\}$, the output satisfies $x_m = P(\\Psi)y$ with $\\deg P \\le K(m-1)$ for every setting of the learned coefficients. The proof runs by the Krylov-subspace invariance of CG: each iteration adds one multiplication by $t_K(\\Psi)$, so the iterate lies in a Krylov subspace of dimension $m$, and composing the degree-$K$ polynomial $t_K$ with a degree-$(m-1)$ polynomial gives degree at most $K(m-1)$. Proposition 5 sharpens this: the reachable set is the image of a polynomial map from $\\mathbb{R}^{K+2m+1}$, so when $K(m-2)>2m$ it is a strict, measure-zero subset of $\\mathcal{P}_{K(m-1)}$; for $K=10,m=3$ the reachable dimension is at most 17 inside a 21-dimensional space. The paper further establishes closed-form behavior at initialization (a leaky denoiser with response $g_K(\\lambda)=\\lambda/(1-(1-\\lambda)^{K+1})$, bounded below by $1/(K+1)$), a condition-number lower bound $\\kappa(\\Psi)>1+\\mu$ for the shift-invariant case, and a Bernstein parameterization that spans the full polynomial class with a nonexpansiveness guarantee.","pith_inferences":["This paper's argument suggests a broader principle: any unrolled solver whose internal coefficients are input-independent constants and whose operator is fixed will produce maps lying in the polynomial algebra generated by that operator, so the same thinness may hold for unrolled proximal or ADMM loops, not only for conjugate gradient.","If the only difference between parameterizations is conditioning rather than capacity, then reporting test accuracy alone is misleading; reporting the effective rank of the parameter-to-response Jacobian, as done here, would be a more meaningful comparison for unrolled networks.","The condition-number bound implies that the truncation order must scale with the inverse of the graph's smallest eigenvalue, so any practical fixed-order construction either accepts large error or must let the graph or $\\mu$ be estimated from data.","A direct testable prediction: a larger unrolled block (higher $K$ or $m$) should not improve test PSNR beyond the corresponding Bernstein polynomial of the same degree on the same fixed graph, provided both receive sufficient optimization."],"forward_implications":["At any point of training, the unrolled TSE+CG module realizes exactly a polynomial graph filter of degree at most $K(m-1)$, and at $K=10,m=3$ the reachable set is a measure-zero subset of the degree-20 polynomial class, so learning the coefficients cannot cover that class.","The standard initialization acts as a leaky denoiser: its spectral response $g_K(\\lambda)$ lies above $\\lambda$ for $\\lambda<1$, approaches $1/(K+1)$ as $\\lambda\\to0^+$, and satisfies the error bound $\\lVert x_{\\mathrm{init}}-\\Psi y\\rVert_2 \\le \\frac{1}{K+1}\\lVert y\\rVert_2$, tight as $\\lambda_{\\min}\\to0$.","Because $\\kappa(\\Psi)>1+\\mu$ and $\\kappa(\\Psi)\\approx107$ at $\\mu=100$ in the measurements, the condition (11) requires $K\\ge488$ for accuracy $\\varepsilon=10^{-2}$, far above the $K=10$ used in practice.","The Bernstein parameterization spans all of $\\mathcal{P}_d$, includes the target $\\Psi$ exactly at zero approximation error, and offers a convex squared-error objective with a nonexpansiveness guarantee; the unrolled block's only advantage is its optimization path, not its reachable set.","The collapse extends to stacked blocks sharing a common $\\Psi$ (degree bound becomes $LK(m-1)$) and to $\\ell^1$ priors with nonlinear shrinkage between blocks, while modules that re-estimate the graph per block leave the polynomial class."],"supporting_citations":[{"why":"Establishes the denoiser–graph-filter correspondence used to define $\\Psi$ and to read $P(\\Psi)$ as a MAP solution.","marker":"[2]"},{"why":"The unrolled TSE+CG construction analyzed in this paper, with its architecture and initialization.","marker":"[3]"},{"why":"The source quoted for Theorem 1, which the paper follows one step further through the unrolling.","marker":"[4]"},{"why":"Provides the Krylov-subspace property of CG iterates on which Theorem 4 is built.","marker":"[5]"},{"why":"Defines polynomial graph filters in the Chebyshev basis, the class that this paper proves is the destination of the unrolled block.","marker":"[7]"},{"why":"Bernstein polynomial conversion and basis properties used for Proposition 9 and the direct parameterization.","marker":"[11]"},{"why":"Compactly supported positive-definite kernels, used to show the condition-number bound is not an artifact of kernel engineering.","marker":"[13]"}],"fun_headline_variants":["Unrolled graph denoisers are just polynomial filters","Learned CG steps don't expand the reachable filter class","Graph unrolling with solver steps is capacity-neutral","Unrolling a graph denoiser cannot beat direct polynomial fits","Fixed-step unrolling only yields degree-bounded polynomials"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The collapse theorem assumes the conjugate-gradient step sizes $\\alpha_k,\\beta_k$ are fixed scalars shared across all inputs; if they were allowed to depend on the residual, as in textbook CG, the map would still lie in a Krylov subspace for each input but would not be a single fixed polynomial in $\\Psi$, and the central claim would not hold.","fun_headline_variants_meta":{"raw":{"variants":["Unrolled graph denoisers are just polynomial filters","Learned CG steps don't expand the reachable filter class","Graph unrolling with solver steps is capacity-neutral","Unrolling a graph denoiser cannot beat direct polynomial fits","Fixed-step unrolling only yields degree-bounded polynomials"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000426,"raw_usage":{"total_tokens":2252,"prompt_tokens":1082,"completion_tokens":1170,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":698,"completion_tokens_details":{"reasoning_tokens":1091}},"tokens_in":698,"tokens_out":1170,"duration_ms":9979,"temperature":1.0,"reasoning_tokens":1091,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:19:32.300966+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained unrolled module on a basis of inputs $y_1,\\dots,y_N$; if any output $x_m$ has a component outside $\\operatorname{span}\\{\\Psi^j y_i:0\\le j\\le K(m-1)\\}$, then Theorem 4 is false.","supporting_citations":[{"cited_title":"Performance Analysis of Plug-and-Play ADMM: A Graph Signal Processing Perspective","cited_arxiv_id":"1809.00020","evidence_quote":"Establishes the denoiser–graph-filter correspondence used to define $\\Psi$ and to read $P(\\Psi)$ as a MAP solution."},{"cited_title":"Constructing an interpretable deep denoiser by unrolling graph Laplacian regularizer,","cited_arxiv_id":null,"evidence_quote":"The unrolled TSE+CG construction analyzed in this paper, with its architecture and initialization."},{"cited_title":"Mixed graph signal analysis of joint image denoising / interpolation,","cited_arxiv_id":null,"evidence_quote":"The source quoted for Theorem 1, which the paper follows one step further through the unrolling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Krylov-subspace property of CG iterates on which Theorem 4 is built."},{"cited_title":"Wavelets on graphs via spectral graph theory,","cited_arxiv_id":null,"evidence_quote":"Defines polynomial graph filters in the Chebyshev basis, the class that this paper proves is the destination of the unrolled block."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Bernstein polynomial conversion and basis properties used for Proposition 9 and the direct parameterization."},{"cited_title":"Piecewise polynomial, positive definite and compactly supported radial func- tions of minimal degree,","cited_arxiv_id":null,"evidence_quote":"Compactly supported positive-definite kernels, used to show the condition-number bound is not an artifact of kernel engineering."}],"review_version":1}