{"id":"eb57a9c5-4ee4-4b62-9fd7-15f5b72316b3","arxiv_id":"2412.05343","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ERED generalizes equivariant and stochastic plug-and-play denoisers into one framework, proves convergence and critical-point behavior, and finds only modest practical gains from equivariance.","lead":"This paper introduces ERED, a unified way to make plug-and-play image restoration respect symmetry properties such as flips, rotations, and translations, and it proves convergence of the resulting stochastic algorithm. A generalist reader might care because equivariance is a cheap structural constraint that could improve reconstruction, and this paper tests exactly how much it helps.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sign mismatch between Algorithm 1 and the analyzed update in Eq. (15): the convergence proofs target the opposite regularization sign, so as written they do not cover the proposed ERED update.","rationale":"The paper proposes an appealing unified equivariant PnP framework, and the experiments are reported with appropriate caveats. However, the theoretical section as written contains a systematic sign error: the update rule analyzed in Eq. (15) and throughout the proofs has the opposite sign to Algorithm 1 when the exact MMSE denoiser is used. Because the sign determines whether the method performs gradient descent on the stated equivariant regularizer or on its negative, the convergence, bias, and critical-point theorems cannot be said to apply to the proposed algorithm without correction. This is not a matter of external consensus or boundedness assumptions; it is an internal inconsistency between the algorithm definition and the analysis. The reader's boundedness concern is valid but secondary, so I partially agree. A corrected version that fixes the sign and re-derives the corresponding estimates could well salvage the main claims, so conditional acceptance with a mandatory correction is the appropriate posture.","tokens_in":14590,"tokens_out":11628,"duration_ms":108721,"concrete_test":"Run a one-dimensional sanity check with prior p = N(0,1), G = {id}, f = 0, λ = 1, and the exact MMSE denoiser. From x0 = 1, Algorithm 1 (with its minus-before-(G-D*) sign) gives x_{k+1} = x_k + δλ∇log p_sigma(x_k), which converges to 0, the minimizer of -log p_sigma. The update in Eq. (15) gives x_{k+1} = x_k - δλ∇log p_sigma(x_k), which moves away from 0 and diverges. This immediately shows the two updates define different dynamics and settle which one Proposition 2 actually covers. In parallel, independently re-derive Eq. (15) from Algorithm 1 and the chain rule using definition (6); the signs must match. If they do not, the theorem statement and proof require correction.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central convergence claims (Propositions 2 and 3, and the critical-point analysis in Propositions 4-5) are derived for a stochastic gradient update whose sign is opposite to the one in Algorithm 1. With the exact MMSE denoiser, Tweedie's formula (3) gives (G(x)-D*_sigma(G(x)))/sigma^2 = -∇log p_sigma(G(x)). Substituting into Algorithm 1 line 6 yields x_{k+1} = x_k - δ∇f(x_k) + δλ J_G^T(x_k)∇log p_sigma(G(x_k)). However, Eq. (15) in Section 4.1 states x_{k+1} = x_k - δ_k∇f(x_k) - λδ_k J_G^T(x_k)∇log p_sigma(G(x_k)), i.e., the opposite sign. The proof of Proposition 2 then defines ξ_k = ∇f(x_k) + λ J_G^T∇log p_sigma(G(x_k)) - ∇F^pi_sigma(x_k), and the identity x_{k+1}=x_k-δ_k(∇F^pi_sigma(x_k)+ξ_k) holds only for the Eq. (15) sign, not for Algorithm 1. Moreover, with the regularizer defined in (6) as r^pi_sigma = -E[log(p_sigma∘G)], its gradient is ∇r^pi_sigma = -E[J_G^T∇log p_sigma(G)]; gradient descent on F^pi_sigma = f + λr^pi_sigma must use the Algorithm 1 sign (plus). Eq. (15) is instead gradient descent on f + λE[log(p_sigma∘G)], which maximizes the prior log-likelihood. The same sign flip appears in the proof of Proposition 4, which writes s^pi_sigma(x) - ∇log p(x) = E[J_G^T(∇log p_sigma - ∇log p)(G(x))]; this is only consistent with s^pi_sigma = +E[J_G^T∇log p_sigma(G)], contradicting definition (7). Consequently, the convergence results and bias bounds apply to a different algorithm than the one proposed. The boundedness assumption on Λ_K flagged by the reader is a real limitation, but the sign inconsistency is more fundamental: if Algorithm 1's sign is the intended one, the proofs need to be re-derived and the bias expression in Eq. (23) changes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Equivariant Regularization by Denoising (ERED), a plug-and-play image restoration framework in which the prior is encoded through an equivariant denoiser and a stochastic gradient update over a transformation group. The authors introduce a π-equivariant regularization term r^π_σ, relate it to previous equivariant PnP methods and to their own SNORE method, and analyze convergence of the resulting stochastic algorithm. They prove convergence to critical points of F^π_σ = f + λ r^π_σ when the exact MMSE denoiser is used (Proposition 2), give bias bounds for an inexact denoiser (Proposition 3), and study the behavior of critical points as the denoiser noise level σ tends to zero for π-equivariant priors (Propositions 4 and 5). Numerical experiments on deblurring, super-resolution, and despeckling show modest but consistent gains over RED.","tokens_in":15114,"tokens_out":14353,"duration_ms":128107,"significance":"The proposed framework is conceptually appealing: it unifies equivariant PnP, stochastic PnP, and SNORE under a single formalism, and it targets a real gap, namely the lack of convergence guarantees for equivariant stochastic PnP methods. The use of the Doucet-Tadic theorem is appropriate for biased stochastic gradient descent, and the paper explicitly states its assumptions and provides complete proofs in the appendix. The experimental section is reasonably broad, covering multiple restoration tasks and denoisers. However, the theoretical core as written contains sign inconsistencies that make Propositions 2 and 4 incorrect as stated relative to the definitions and algorithm used. These issues are local and repairable, but they must be fixed before the convergence claims can be accepted.","major_comments":[{"comment":"The definition of s^π_σ is inconsistent with the proof of Proposition 4. Equation (7) defines s^π_σ(x) = -E[J_G^T(x)∇log p_σ(G(x))], which is the gradient of r^π_σ, not the score s = ∇log p. For a π-equivariant prior, s = +E[J_G^T (s∘G)], so as σ→0 one gets s^π_σ → -s. Proposition 4 claims ∥s - s^π_σ∥_{∞,K} → 0, which is generally false under definition (7). The proof's Eq. (25) instead uses s^π_σ = +E[J_G^T ∇log p_σ(G)], i.e. the opposite sign. This affects Proposition 5 as well, since its proof relies on Proposition 4 to conclude ∥F - F^π_σ∥_{∞,K} → 0. The authors should adopt one consistent sign convention throughout and adjust Eqs. (7), (8), (25), and Proposition 4 accordingly.","section":"Section 3.2, Eqs. (7)-(8); Section 4.3, Eq. (25) and Proposition 4"},{"comment":"There is a sign mismatch between Algorithm 1 and the update analyzed in Proposition 2. With the exact MMSE denoiser, Tweedie's formula gives (G(x)-D*_σ(G(x)))/σ^2 = -∇log p_σ(G(x)); substituting into Algorithm 1 line 6 yields x_{k+1} = x_k - δ∇f(x_k) + δλ J_G^T(x_k)∇log p_σ(G(x_k)). Equation (15), however, states x_{k+1} = x_k - δ_k∇f(x_k) - λδ_k J_G^T(x_k)∇log p_σ(G(x_k)), with the opposite sign. The subsequent definition of ξ_k in the proof of Proposition 2 is consistent with Eq. (15) but not with Algorithm 1. As written, Proposition 2 proves convergence for an update that maximizes the prior log-likelihood rather than for the proposed ERED algorithm. This is fixable by flipping the sign in Eq. (15) and in the definition of ξ_k, but the current proof does not establish the stated result.","section":"Section 4.1, Eq. (15) and proof of Proposition 2"},{"comment":"The convergence results are stated 'almost surely on Λ_K', where Λ_K is the set of random seeds for which all iterates stay in a fixed compact K. The paper calls this the boundedness assumption but gives no conditions on f, G, or the denoiser that ensure Λ_K has positive probability, nor any argument that trajectories remain bounded under Assumptions 1-4. Thus the central guarantees are conditional on an unverified property of the algorithm's own dynamics. Please either provide sufficient conditions for almost-sure boundedness, or explicitly state and discuss this as a limitation that must be assumed separately.","section":"Section 4.1, definition of Λ_K and Propositions 2-3"},{"comment":"Proposition 1 states that r^π_σ is π-equivariant, but 'π-equivariant' is defined only for densities (Definition 2), and r^π_σ is not a density. The proof also takes logarithms of r^π_σ, which is not guaranteed to be positive. Please reformulate the property that is actually being proved for the regularizer and justify the proof under the stated assumptions on G and π.","section":"Section 3.2, Proposition 1"}],"minor_comments":[{"comment":"The sentence 'As shown in [16], natural images densities tend to be invariant to some set of transformations such as rotations or flips have been studied in.' is incomplete and should be rewritten.","section":"Section 1"},{"comment":"The text refers to 'Table 5' when presenting results that appear in Table 1; also, 'PNSR' is used instead of 'PSNR' in several table headers.","section":"Section 5"},{"comment":"The notation for the noising-denoising example uses σ both for the translation magnitude and for the denoiser noise level; this is potentially confusing and should be clarified.","section":"Section 3.3, Eq. (12)"},{"comment":"The constants appearing in the proofs of Lemmas 1 and 2 are introduced with inconsistent notations (C vs C2, and subscripted constants); a pass for notational consistency would improve readability.","section":"Appendix B.2-B.3"}],"recommendation":"major_revision","confidential_remarks":"The sign inconsistencies in Eqs. (7), (8), (15), and (25) are pervasive and affect the central convergence claims, but they are repairable with a careful, consistent redefinition of signs. The experimental section and the overall framework are not invalidated by these issues. I recommend major revision rather than rejection, provided the authors reconcile the sign conventions and address the boundedness assumption on Λ_K."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the pi-equivariance framework is a real contribution, but the convergence theorems as written don't cover Algorithm 1 because of a sign error. It looks fixable, but it has to be fixed before the results can be used.\n\nWhat's genuinely new: the definition of pi-equivariance and the unified view that contains equivariant PnP and SNORE as special cases; the stochastic-gradient treatment of ERED; and the sigma-to-zero critical-point analysis. The paper is clearly written and the experiments are refreshingly honest. They report gains of only 0.1-0.3 dB and explicitly say equivariance doesn't help much, even suggesting the denoiser may have already learned the invariance.\n\nThe soft spots, in order of seriousness. First, the sign mismatch. Algorithm 1 line 6 with the exact MMSE denoiser becomes x_{k+1}=x_k-delta*grad f(x_k)+delta*lambda*J^T*grad log p_sigma(G(x_k)) via Tweedie. Equation (15), which the analysis is based on, has a minus sign on the last term. The proof of Prop 2 then defines xi_k=lambda*(J^T*grad log p_sigma - E[J^T*grad log p_sigma]) and gets x_{k+1}=x_k-delta*(grad F^pi_sigma+xi_k) only for the objective f+lambda*E[log(p_sigma circ G)] -- the opposite regularization sign from the F^pi_sigma defined in (14). Prop 4's proof has the same problem against definition (7). So the theorems as stated analyze a different algorithm. I'm fairly sure the correct proofs for Algorithm 1 go through with xi_k=-lambda*(J^T*grad log p_sigma - E[...]), and the bias bound in Eq. (23) is norm-symmetric, so the fix is probably local. But it's a real correctness issue, not a typo in one line.\n\nSecond, the boundedness assumption. Convergence is stated almost surely on Lambda_K, the set of seeds whose iterates stay in a compact set, and the paper gives no conditions that make Lambda_K non-negligible. The guarantees are thus conditional on the algorithm's own dynamics. That's a genuine hole in the main theorem, as the reader's report also notes.\n\nMinor: the experiments lack error bars and the hyperparameter selection isn't clearly validation-based. This is secondary.\n\nWho this is for: people working on PnP, equivariant imaging, or stochastic regularizers. The framework is a useful organizing device and the paper is worth discussing. But it should go back to the authors for a corrected version of the analysis and a far more careful treatment of Lambda_K before the convergence claims are taken at face value. A serious editor should send it to review.","headline":"A useful unifying framework for equivariant PnP, but the convergence proofs as written contain a sign error that makes them analyze a different update than Algorithm 1; likely fixable, but it must be fixed.","tokens_in":15648,"tokens_out":10084,"would_cite":true,"duration_ms":88123,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U10","65K10","90C15","94A08"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that a single π-equivariant plug-and-play algorithm, ERED, encompasses existing equivariant and stochastic schemes, and that its iterates converge almost surely on bounded trajectories, with bias bounded by the…","keywords":["image restoration","plug-and-play","equivariance","regularization by denoising","stochastic gradient descent","convergence analysis","Tweedie formula","deep denoiser"],"falsifier":"Run Algorithm 1 with the exact MMSE denoiser on a $\\pi$-equivariant prior with unbounded support and a data term with no coercivity, using a step sequence satisfying Assumption 1; if the iterates escape every compact set with positive probability, the event $\\Lambda_K$ fails and the almost-sure claims do not apply to such runs, which would settle exactly how conditional the theorem is.","tokens_in":14425,"feed_emoji":"🖼️","tokens_out":10211,"duration_ms":91957,"temperature":0.7,"pith_summary":"This paper proposes ERED, a stochastic plug-and-play algorithm that replaces a denoiser by its average over a group of transformations and treats that average as the score of a symmetric prior. The authors prove that when the exact MMSE denoiser is used and the iterates happen to stay inside a fixed compact set, the algorithm converges almost surely to critical points of the surrogate objective $F^{\\pi}_{\\sigma} = f + \\lambda r^{\\pi}_{\\sigma}$. With an inexact, learned denoiser, the asymptotic gradient norm and the spread of objective values are bounded by a bias term that is controlled by the denoiser's error on the transformed images. They also show that as the denoising level goes to zero, critical points of the surrogate converge to critical points of the original variational problem whenever the prior is $\\pi$-equivariant. Why it matters: this gives one theoretical template for stochastic and equivariant plug-and-play methods, so a user can choose transformations with some guarantee about where the algorithm is heading.","feed_headline":"Equivariant averaging unifies stochastic plug-and-play restoration","feed_subtitle":"Convergence and bias bounds for symmetry-averaged denoisers, covering rotations, flips, and noising","key_machinery":"The load-bearing object is the $\\pi$-equivariant score identity built from Tweedie's formula. For the MMSE denoiser $D^*_{\\sigma}$, $-\\nabla \\log p_{\\sigma}(x) = (x-D^*_{\\sigma}(x))/\\sigma^2$; the paper defines $s^{\\pi}_{\\sigma}(x) = E_{G\\sim\\pi}[J_G^T(x)(G(x)-D_{\\sigma}(G(x)))]/\\sigma^2$ and the regularizer $r^{\\pi}_{\\sigma}$. The convergence argument rests on a theorem for biased stochastic gradient descent (recalled as Theorem 1 from [7]), which yields almost-sure convergence and bias bounds on the event $\\Lambda_K$ where all iterates remain in a compact set. The critical-point transfer at $\\sigma\\to0$ rests on the identity $s = E_{G\\sim\\pi}[J_G^T(s\\circ G)]$ for $\\pi$-equivariant priors, which lets score convergence on the orbit ball become gradient convergence of the original objective.","core_discovery":"The central discovery is a unification: equivariant PnP and stochastic denoising regularization are not separate heuristics but instances of one construction, the $\\pi$-equivariant denoiser $\\tilde D_{\\sigma}(x)=E_{G\\sim\\pi}[J_G^T(x)D_{\\sigma}(G(x))]$, with the corresponding regularizer $r^{\\pi}_{\\sigma}$. Algorithm 1 is a biased stochastic gradient descent on $F^{\\pi}_{\\sigma}$. Under the paper's assumptions, Proposition 2 shows that with the exact MMSE denoiser the iterates converge almost surely on $\\Lambda_K$ to the critical set of $F^{\\pi}_{\\sigma}$; Proposition 3 shows that with an inexact denoiser, $\\limsup \\|\\nabla F^{\\pi}_{\\sigma}(x_k)\\|$ and the objective spread are bounded by the asymptotic bias $\\eta$, which in turn is bounded by $\\lambda \\sigma^{-2} E(\\|J_G\\|)\\,\\|D_{\\sigma}-D^*_{\\sigma}\\|_{\\infty,L}$ on the orbit ball $L$. Proposition 5 closes the loop: for a $\\pi$-equivariant prior, every limit point of critical points of the surrogate as $\\sigma\\to0$ is a critical point of the original objective $F=f+\\lambda r$.","pith_inferences":["The boundedness event $\\Lambda_K$ is the main gap; proving it from coercivity of $f$ or boundedness of the denoiser would turn the almost-sure-on-$\\Lambda_K$ statements into unconditional almost-sure theorems.","One testable prediction of the theory is that the residual gradient norm of ERED should track the denoiser error on the orbit ball; measuring $\\|D_{\\sigma}-D^*_{\\sigma}\\|_{\\infty,L}$ along a run and comparing with (21) would validate or falsify the bias model.","The framework suggests a practical diagnostic for symmetry: run ERED with different groups $\\pi$ and compare restoration quality; a group that fails to improve indicates the image class is not $\\pi$-equivariant, as the subpixel-rotation results suggest."],"forward_implications":["The same convergence theory now covers equivariant PnP, translation-based noising (SNORE), and other stochastic schemes as instances of ERED, so guarantees proven for one apply to the others.","For a fixed denoiser, the bias bound says that the asymptotic error of ERED is governed by how well the denoiser approximates the MMSE denoiser on the ball containing all transformed iterates, identifying where denoiser training effort should go.","When the prior is $\\pi$-equivariant, the surrogate problem solved by ERED is not an unrelated objective: its critical points approach those of the original variational problem as $\\sigma\\to0$.","Choice of transformation set is not neutral: the experiments indicate flips and rotations improve deblurring and despeckling slightly, while subpixel rotation and combining all transformations do not, suggesting the image prior is only equivariant to certain groups."],"supporting_citations":[{"why":"Supplies Theorem 1, the biased stochastic gradient descent convergence and bias-bound result that Propositions 2 and 3 invoke.","marker":"[7]"},{"why":"Defines RED, the deterministic algorithm that ERED extends and compares against in experiments.","marker":"[24]"},{"why":"Introduces equivariant PnP for finite isometries, which ERED generalizes through the $\\pi$-equivariant construction.","marker":"[30]"},{"why":"Proposes SNORE, the stochastic denoising regularization shown to be a special case of ERED.","marker":"[23]"},{"why":"Introduces the Plug-and-Play paradigm that frames ERED.","marker":"[31]"},{"why":"Provides the GS-DRUNet denoiser and weights used in the deblurring and despeckling experiments.","marker":"[13]"},{"why":"Establishes Assumption 1(c) on sub-polynomial score growth under the manifold hypothesis.","marker":"[6]"},{"why":"Supplies smoothness of $\\log p_\\sigma$ and the stochastic PnP background used in the proof chain.","marker":"[15]"}],"fun_headline_variants":["Equivariant denoisers unify restoration toolkits","Averaging over symmetries fixes stochastic PnP","Symmetry-averaged denoisers sharpen restoration bounds","One equivariant denoiser to rule all restorations","Equivariant averaging: a unified restoration recipe"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole convergence and bias theory is conditional on the event that the algorithm's own iterates never leave a fixed compact set, and the paper gives no condition that guarantees this event has positive probability.","fun_headline_variants_meta":{"raw":{"variants":["Equivariant denoisers unify restoration toolkits","Averaging over symmetries fixes stochastic PnP","Symmetry-averaged denoisers sharpen restoration bounds","One equivariant denoiser to rule all restorations","Equivariant averaging: a unified restoration recipe"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1410,"prompt_tokens":921,"completion_tokens":489,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":408}},"tokens_in":537,"tokens_out":489,"duration_ms":4970,"temperature":1.0,"reasoning_tokens":408,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:08:22.934238+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 with the exact MMSE denoiser on a $\\pi$-equivariant prior with unbounded support and a data term with no coercivity, using a step sequence satisfying Assumption 1; if the iterates escape every compact set with positive probability, the event $\\Lambda_K$ fails and the almost-sure claims do not apply to such runs, which would settle exactly how conditional the theorem is.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Plug-and-Play paradigm that frames ERED."},{"cited_title":"Renaud, J","cited_arxiv_id":null,"evidence_quote":"Proposes SNORE, the stochastic denoising regularization shown to be a special case of ERED."},{"cited_title":"Doucet and V","cited_arxiv_id":null,"evidence_quote":"Supplies Theorem 1, the biased stochastic gradient descent convergence and bias-bound result that Propositions 2 and 3 invoke."},{"cited_title":"Romano, M","cited_arxiv_id":null,"evidence_quote":"Defines RED, the deterministic algorithm that ERED extends and compares against in experiments."},{"cited_title":"Terris, T","cited_arxiv_id":null,"evidence_quote":"Introduces equivariant PnP for finite isometries, which ERED generalizes through the $\\pi$-equivariant construction."},{"cited_title":"Hurault, A","cited_arxiv_id":null,"evidence_quote":"Provides the GS-DRUNet denoiser and weights used in the deblurring and despeckling experiments."},{"cited_title":"De Bortoli","cited_arxiv_id":null,"evidence_quote":"Establishes Assumption 1(c) on sub-polynomial score growth under the manifold hypothesis."},{"cited_title":"Laumont, V","cited_arxiv_id":null,"evidence_quote":"Supplies smoothness of $\\log p_\\sigma$ and the stochastic PnP background used in the proof chain."}],"review_version":1}