{"id":"361fee66-7bef-4f30-949a-1fb8a1379c8e","arxiv_id":"2607.29525","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PnP-FBS with MMSE denoisers recovers the true signal with explicit pointwise and Wasserstein error bounds, provided the denoiser's noise covariance is matched to the preconditioned observation noise.","lead":"This math paper proves recovery guarantees for Plug-and-Play reconstruction when the denoiser is the optimal MMSE estimator, even for correlated or degenerate noise. It shows the denoiser's noise model must be matched to the forward model and preconditioner, not chosen independently.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The contraction condition (20)/(21) is the load-bearing gap: (20) is exactly a nonexpansive denoiser and (21) requires near-isotropic BA, so the advertised guarantees do not cover ill-conditioned problems with non-contractive denoisers.","rationale":"The reader's weakest assumption identifies the contraction condition q < 1 as the main gap, and my analysis agrees and sharpens it: condition (20) is not merely 'very large training noise' but is exactly equivalent to the denoiser being nonexpansive by the paper's own Lemma 4.3, undermining the claim that contractiveness is not required. Condition (21) is seen to require near-isotropy of BA even for mildly expansive denoisers, so the theorem's practical applicability is far narrower than the abstract suggests. This is a load-bearing concern because the paper's central 'physics-aware' conclusion is derived from error bounds that only contract under these restrictive conditions. I do not see an internal inconsistency in the proofs; the theorems likely hold as stated. The issue is the gap between the stated assumptions and the advertised 'mild assumptions' and 'no contractiveness' claims. Since the reader already recommended CONDITIONAL acceptance, my verdict remains UNCHANGED. No further concern about data fabrication, circularity, or mathematical soundness was found.","tokens_in":38644,"tokens_out":8551,"duration_ms":88063,"concrete_test":"Compute, for a concrete ill-conditioned problem (e.g., 1D Gaussian deblurring with κ(A)=20, M=1, B=A^T, and an MMSE denoiser trained with isotropic noise σ=0.05 so L=M²/σ²=400), whether either (20) or (21) holds: κ(BA)=400, while the RHS of (21) is approximately (1−(σ²/M²)^2)^{−1/2} ≈ 1.00000016, so both fail. Then run the PnP-FBS iteration with γ=(κ(BA)λ_max(BA))^{-1} and measure the empirical contraction ratio ρ = lim ||x_{k+1}−x||/||x_k−x||. If ρ ≥ 1, the theorem gives no guarantee in this standard regime, confirming the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the paper's central claim to be substantiated, the recovery bounds of Theorems 3.2 and 3.6 must hold under the advertised 'mild assumptions.' Both theorems hinge on the contraction factor q = (M²/λ_min(ΣΥ))·√(1−κ(BA)^{-2}) < 1, guaranteed only by condition (20) or (21). By Lemma 4.3, L := M²/λ_min(ΣΥ) is a Lipschitz bound for the MMSE denoiser, so condition (20) is exactly L ≤ 1 — i.e., the denoiser is nonexpansive, the very assumption the paper claims to relax. Condition (21) allows L > 1 only if κ(BA) < (1−L^{-2})^{−1/2}; for L=2 this is κ < 1.15, and for L=1.1 it is κ < 2.4. Thus non-contractive denoisers are covered only when BA is nearly the identity on the relevant cone. For standard ill-posed operators (non-injective A, or κ(BA) ≳ 10 with realistic training noise so λ_min ≪ M²), neither condition holds, and the bounds (23) and (30) are vacuous because (1−q)^{-1} is undefined when q ≥ 1. Consequently, the headline conclusion that the denoiser 'cannot be chosen physics-agnostically' is only established in a restricted near-contractive/near-isotropic regime, not for the ill-conditioned problems the introduction targets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the forward-backward splitting (FBS) variant of Plug-and-Play (PnP) in which the proximal operator is replaced by an MMSE denoiser designed for zero-mean Gaussian noise with possibly degenerate covariance Σ_Υ, and the gradient of the data fidelity is preconditioned by a matrix B. The main theoretical results are: (i) regularity properties of the MMSE denoiser, including a Jacobian formula, a Lipschitz bound L ≤ M²/λ_min(Σ_Υ), monotonicity, cocoercivity (for isotropic noise), and stability with respect to the prior measure; (ii) pointwise and Wasserstein recovery bounds for the PnP-FBS iteration, stating that if either L≤1 or a near-isotropic condition on BA holds, then the error is controlled by a contraction factor q plus a term involving the Bures metric between γBE and Υ; (iii) an extension to neural-network denoisers with approximation guarantees. The paper concludes that the denoiser cannot be chosen independently of the forward model.","tokens_in":39068,"tokens_out":7325,"duration_ms":67549,"significance":"If correct, the paper provides a rigorous framework for designing physics-aware denoisers in PnP and identifies the Bures-metric mismatch as the key quantity controlling distributional recovery. The proofs are detailed and largely self-contained; the cocoercivity result for MMSE estimators (Proposition 4.10) and the derivative identity (40) are useful contributions. The Wasserstein analysis with the explicit Bures-metric term (Remark 3.7) is a genuine step forward. However, the advertised 'mild assumptions' and the claim that nonexpansiveness is not required are substantially overstated: the contraction conditions (20)/(21) essentially reduce to nonexpansiveness or to BA being near the identity, which excludes typical ill-conditioned inverse problems. This does not invalidate the mathematics, but it materially narrows the scope of the headline claim.","major_comments":[{"comment":"The factor q = (M²/λ_min(Σ_Υ))√(1−κ(BA)^{-2}) must be <1 for the bounds (23) and (30) to be meaningful. Since Lemma 4.3 gives L := M²/λ_min(Σ_Υ) as a Lipschitz bound for the MMSE denoiser, condition (20) is exactly L≤1, i.e. nonexpansiveness. Condition (21), equivalently κ(BA) < (1−L^{-2})^{-1/2}, permits L>1 only when κ(BA) is very close to 1 (for L=2, κ<1.15; for L=1.1, κ<2.4). Thus the 'non-contractive' regime is essentially the near-isotropic regime. For the ill-conditioned, non-injective forward operators targeted in the introduction, neither condition will hold in general, and the bounds are vacuous because (1−q)^{-1} is undefined when q≥1. The abstract and Remark 3.3(i) present these as 'mild assumptions'; this is misleading. Please revise the claims and state explicitly the restricted regime in which the guarantees apply.","section":"§3.2, Theorems 3.2 and 3.6, conditions (20)-(21)"},{"comment":"The pointwise recovery bound contains the term 2M(N−1)exp(−ε/(2λ_max(Σ_Υ))). Here ε is the separation gap between the squared distances of the nearest and second-nearest samples; for a typical continuous distribution this gap tends to 0 as N increases (often at rate 1/N), so the term need not vanish and may even grow with N. The phrase 'there exists ε>0' is not quantitative. The modification in Remark 5.2 replaces it with an extra nonvanishing √(λ_max(Σ_Υ))ε term. Hence Theorem 3.2 does not demonstrate consistency of the pointwise error as N grows. Please provide a quantitative treatment of ε or state the bound with an explicit, controlled extra term and discuss the resulting asymptotics.","section":"Lemma 5.1 and Eq. (23)-(24)"},{"comment":"The neural-network recovery bound contains a √N ε term, where ε is the uniform approximation error of the identity on the support manifold S. This term grows with the sample size N, so the bound degrades as more training data are used, unless ε is chosen to decay with N. But then, by Lemma 3.10, the required network width scales like ε^{-d_S}, which is polynomial in N (or worse). The current statement does not make this trade-off explicit and gives the misleading impression that the NN bound is a convergence-in-N result. Please optimize ε as a function of N and state the resulting rate, or clarify the non-asymptotic nature of the bound.","section":"Theorem 3.11, Eq. (36)-(37)"}],"minor_comments":[{"comment":"There are missing spaces in the typeset text ('Wefurtherdeviate', 'thatthedenoiser'), likely a formatting issue; please fix.","section":"Abstract"},{"comment":"The sentence 'This condition imposes that BA cannot be too ill-conditioned' refers to condition (20), which controls M²/λ_min(Σ_Υ) rather than the conditioning of BA. Reword to avoid confusion.","section":"Remark 3.3(i)"},{"comment":"The proof states that T maps co(supp(μ_X)) 'onto' itself; by Brouwer it suffices that it maps into a compact convex set. Please correct 'onto' to 'into'.","section":"Lemma 3.1"},{"comment":"The notation for seκ(BA), seq, and seγ is typographically heavy and easy to confuse with the un-tilded versions. Consider renaming these quantities for readability.","section":"Section 3.2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's own prior work [60] is used as the motivating choice B=A^T Σ_E^{-1/2}, and the analysis builds on it. This is not improper, but the novelty relative to [60] should be made explicit to the editor. The paper is purely theoretical; a simple numerical illustration showing where conditions (20)/(21) hold or fail would help calibrate the restrictiveness of the main theorems, but this is not required for a theory journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this is the first rigorous treatment I have seen of PnP-FBS with MMSE denoisers designed for degenerate and non-diagonal Gaussian noise, and the Bures-metric mismatch term is a real contribution. But the main theorems only bite when either the denoiser's Lipschitz bound is no more than 1 or BA is nearly isotropic; the abstract's \"mild assumptions\" overstates the coverage.\n\nThe genuinely new material is substantial. The derivative identity for the MMSE Jacobian (Prop 4.1), the cocoercivity result (Prop 4.10), the pointwise and Wasserstein recursions, and the neural-network extension are all carefully derived and do not reduce to earlier work. The Wasserstein analysis correctly identifies the Bures metric between gamma BE and Upsilon and shows the choice Sigma_Upsilon = gamma^2 B Sigma_E B^T kills the mismatch. That is a quantitative version of the folklore \"physics-aware denoiser\" rule, and it is derived rather than assumed. The proof structure is coherent, and the main statements follow from the stated assumptions.\n\nThe load-bearing soft spot is the contraction factor q = (M^2 / lambda_min(Sigma_Upsilon)) sqrt(1 - kappa(BA)^-2). Since M^2 / lambda_min is the paper's own Lipschitz bound for the MMSE denoiser (Lemma 4.3), condition (20) is exactly the nonexpansive-denoiser regime the paper claims to go beyond. Condition (21) allows L > 1 only for very well-conditioned BA: for L = 2, kappa < 1.15, and for L = 1.1, kappa < 2.4. For standard ill-posed operators with non-injective A or kappa(BA) in the tens, neither condition holds and the bounds are vacuous because (1 - q)^-1 is undefined. The statement \"we do not require non-expansiveness\" is technically true, but the covered regime is narrow. The paper would be stronger if the theorems used the actual Lipschitz constant L_bD rather than the bound M^2/lambda_min, or if the restricted regime were stated prominently in the abstract.\n\nMinor point: the NN section relies on coarse Lipschitz estimates and approximation assumptions; it is a reasonable extension but not the main event. No circularity or misattribution: dependence on their own prior work [60] is legitimate here.\n\nThis paper is for researchers working on PnP theory or learned denoisers. It deserves a serious referee. I would send it out with a request to revise the claims around (20)-(21), add an honest discussion of the restricted regime, and then it is a solid contribution.","headline":"Solid, useful theory paper on PnP with MMSE denoisers, but the advertised recovery guarantees require a contraction condition that is much more restrictive than the \"mild assumptions\" language suggests.","tokens_in":39568,"tokens_out":3020,"would_cite":true,"duration_ms":27393,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65J22","62F15","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Plug-and-play inverse problems require physics-aware denoisers for guaranteed recovery.","keywords":["plug-and-play","MMSE denoiser","forward-backward splitting","recovery guarantees","Wasserstein distance","inverse problems","neural networks","noise covariance matching"],"falsifier":"Compute q for a simple ill-conditioned inverse problem (e.g., a diagonal A with condition number 100, B = A^T, M fixed) and a denoiser trained with isotropic noise of variance σ². For σ² small enough that M²/σ² > 1 and κ(BA) large, q exceeds 1, so the pointwise bound (23) does not contract; run the iteration and show the error does not converge to the bound, contradicting the theorem's hypotheses.","tokens_in":38560,"feed_emoji":"🧩","tokens_out":3404,"duration_ms":28057,"temperature":0.7,"pith_summary":"This paper proves that in the forward-backward splitting version of Plug-and-Play (PnP), the denoiser must be designed in tandem with the forward model, not chosen independently. The authors analyze an iteration where the standard gradient step is replaced by a flexible operator B, and the denoiser is an MMSE estimator specialized to Gaussian noise with possibly degenerate, non-diagonal covariance. They derive pointwise and Wasserstein recovery bounds that depend on how closely the denoiser's training noise matches the noise transformed by B. The findings imply that conventional choices like B=A^T tacitly assume the denoiser is agnostic to the physics, which harms recovery; instead, the covariance of the training noise should be set to match γ²BΣ_E B^T.","feed_headline":"Denoisers must match the physics for PnP recovery to work","feed_subtitle":"New error bounds show the denoiser's noise model must align with the forward operator's noise.","key_machinery":"The iteration x_{k+1} = D(x_k + γB(y - A x_k)) with MMSE denoiser D. The load-bearing quantity is the Lipschitz constant of D, bounded by M²/λmin(ΣΥ), where M bounds the signal and ΣΥ is the denoiser's noise covariance. The condition ΣΥ = γ²BΣ_EB^T makes the noise-mismatch term vanish, and the contraction factor q = (M²/λmin(ΣΥ)) sqrt(1 - κ(BA)^{-2}) governs all bounds.","core_discovery":"The central claim is that recovery guarantees for PnP hold only when the denoiser's noise model is compatible with the forward operator: the covariance of the MMSE denoiser's training noise should equal the covariance of the observation noise after transformation by B. This makes the Bures metric between the two Gaussian noises vanish, leaving a residual dominated by the operator norm of B and the noise level. The paper also establishes that MMSE denoisers are cocoercive (hence firmly nonexpansive up to scaling), which justifies a common constraint in training.","pith_inferences":["The contraction condition q<1 may be stringent for ill-conditioned problems; for a fixed conditioning, the signal-to-noise ratio of the training data must be low enough to maintain stability. This suggests that for highly ill-conditioned forward operators, the denoiser's training noise must be large, potentially limiting resolution.","The framework suggests a design principle for PnP: choose B as a preconditioner that balances the conditioning of BA and the noise matching, rather than using B=A^T.","One could test the bounds empirically by training denoisers with covariance matched to BΣ_EB^T and comparing recovery quality against physics-agnostic denoisers, especially for degenerate noise (e.g., missing cone in tomography).","The cocoercivity result may be used to regularize neural-network training to enforce convergence guarantees in practice."],"forward_implications":["If the denoiser's training noise is chosen to match the forward model, PnP-FBS has a pointwise recovery error bounded by terms involving noise amplification, denoiser bias, and sampling error.","In Wasserstein distance, the optimal choice ΣΥ = γ²BΣ_EB^T yields a residual of order (1-q)^{-1} γ ||B|| sqrt(λmax(Σ_E)) sqrt(r).","MMSE denoisers are cocoercive, providing a theoretical justification for firm nonexpansiveness in PnP.","Results extend to neural network denoisers with width and depth bounds depending on the manifold dimension of the signal support.","The conventional choice B = A^T implicitly assumes the denoiser is physics-agnostic, which is suboptimal."],"fun_headline_variants":["PnP recovery fails if denoiser ignores forward model noise","Physics-aware denoisers key to plug-and-play guarantees","Denoiser noise must align with forward operator for PnP","Recovery bound ties denoiser noise to imaging physics","Covariance match governs PnP convergence"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The error bounds only contract if the contraction factor q = (M²/λmin(ΣΥ)) sqrt(1 - κ(BA)^{-2}) is less than 1, which requires either a large denoiser noise covariance relative to the signal bound, or a nearly isotropic preconditioned operator BA.","fun_headline_variants_meta":{"raw":{"variants":["PnP recovery fails if denoiser ignores forward model noise","Physics-aware denoisers key to plug-and-play guarantees","Denoiser noise must align with forward operator for PnP","Recovery bound ties denoiser noise to imaging physics","Covariance match governs PnP convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000116,"raw_usage":{"total_tokens":875,"prompt_tokens":667,"completion_tokens":208,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":411,"completion_tokens_details":{"reasoning_tokens":139}},"tokens_in":411,"tokens_out":208,"duration_ms":2561,"temperature":1.0,"reasoning_tokens":139,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:16:42.083674+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute q for a simple ill-conditioned inverse problem (e.g., a diagonal A with condition number 100, B = A^T, M fixed) and a denoiser trained with isotropic noise of variance σ². For σ² small enough that M²/σ² > 1 and κ(BA) large, q exceeds 1, so the pointwise bound (23) does not contract; run the iteration and show the error does not converge to the bound, contradicting the theorem's hypotheses.","supporting_citations":[],"review_version":1}