{"id":"cda696de-faa1-4e9a-b288-fa32b28c1d8d","arxiv_id":"2411.18967","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Plug-and-Play HIO interleaves HIO projections with a deep denoiser and achieves state-of-the-art Fourier phase retrieval quality on natural images.","lead":"This paper derives and tests a phase retrieval method that combines hybrid input-output iteration with a deep learning denoiser. It reports better reconstruction quality and faster runtime than existing deep phase retrieval baselines on a large image benchmark.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (7) does not satisfy the claimed optimality condition: it is one fixed-point iterate, not a solution of the HQS x-subproblem, so the paper's mathematical derivation is unsupported even though the empirical SOTA claim may survive.","rationale":"The paper's useful contribution is an empirical algorithm: PnP HIO with a pretrained DnCNN prior and inner HIO iterations improves PSNR/SSIM on 236 images across three noise levels. The table is internally consistent and the improvements are modest but systematic, with error bars that do not overlap at M=2 and M=3. I do not see a fatal flaw in the experiments themselves: the same HIO-refined initialization is used for all methods, the comparison uses author implementations, and the runtime numbers are plausible. The load-bearing weakness is the claimed mathematical derivation. The x-subproblem of HQS is not solved by Eq. (7). The text asserts that one fixed-point iteration initialized at z_N satisfies the necessary condition; this is false in general, and a scalar counterexample demonstrates it. Therefore the paper's stated contribution of 'analytical update steps based on HQS' is unsupported, and the algorithm's behavior remains an empirical PnP-HIO heuristic. This warrants keeping the conditional verdict: the empirical claims can stand only if the derivation is corrected or explicitly downgraded to a heuristic, and ideally if code and data are released. The Data availability statement says data are not public; no code is mentioned, further limiting verification. My primary concern matches the reader's weakest_assumption, so agreement is agree.","tokens_in":14428,"tokens_out":15073,"duration_ms":134245,"concrete_test":"Analytically substitute Eq. (7) into the left-hand side of Eq. (6) for the scalar instance A=2, y=2, z_N=0.1, Q_N=1. The update gives x=0.55, the exact minimizer of Eq. (5a) is 0.82, and the residual in Eq. (6) is nonzero. This refutes the claim that Eq. (7) satisfies the optimality condition; if the authors disagree, they should provide a corrected derivation in which Eq. (7) is derived, not asserted, from Eq. (6).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest point is the claimed HQS derivation in Section 3, specifically the jump from the necessary condition (6) to the update (7). Equation (5a) asks for a minimizer of ||y-|Ax|||^2 + Q_N||x-z_N||^2. Equation (6) is the first-order condition for that subproblem. The paper states that initializing the fixed-point iteration with z_N 'provides the following solution that satisfies' the condition and then writes x = U_N Re{A^{-1}(y ⊙ Az_N/|Az_N|)} + (1-U_N)z_N. A single iterate of a fixed-point map is not a fixed point: substituting (7) into (6) leaves a nonzero residual for generic z_N and A. For example, with scalar A=2, y=2, z_N=0.1, Q_N=1, Eq. (7) gives x=0.55, while the exact minimizer of (5a) is x=0.82; the gradient condition fails. The update is instead the standard error-reduction/HIO phase-update step evaluated at the previous z_N together with a relaxation toward z_N; it does not minimize (5a) unless x and z_N happen to be aligned. Thus the 'analytical update steps based on half-quadratic splitting' are not actually consequences of HQS, and the mathematical-development contribution is unsupported. This does not by itself refute the empirical SOTA claim, but it makes the method a heuristic PnP-HIO whose behavior needs to be justified empirically or by a corrected derivation. The Data availability statement provides no code or data, so the exact schedules and DNN switching used for Table 1 cannot be independently checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two phase-retrieval algorithms, Plug-and-Play PR and Plug-and-Play HIO, that combine half-quadratic splitting (HQS) with a pre-trained CNN denoiser. The authors claim to derive analytical measurement and image update steps from HQS and to combine them with learning-based denoising. The experimental section compares the proposed methods against HIO, PrDeep, and Iterative DNN-HIO on 236 test images at three Poisson noise levels, using 5 Monte Carlo runs, and reports average PSNR, SSIM, and runtime. The results show that Plug-and-Play HIO achieves the best average PSNR/SSIM for natural images at all tested noise levels, with runtime roughly half that of PrDeep, and exhibits improved robustness to initialization and noise.","tokens_in":14855,"tokens_out":12055,"duration_ms":102107,"significance":"The empirical evaluation is a clear strength: 236 test images, 5 Monte Carlo runs per setting, comparisons against the authors' implementations of competing learning-based methods, and separate sensitivity analyses for initialization and noise. If the empirical results are correct, Plug-and-Play HIO is practically useful because it combines the computational efficiency of classical projection methods with the prior power of deep denoisers and appears less initialization-sensitive than the compared alternatives. However, the paper's central mathematical claim—that the update steps are derived from half-quadratic splitting—is not supported: Eq. (7) is a single fixed-point iterate rather than a solution of the stated optimality condition, and the HIO inner iterations in Algorithm 2 are introduced heuristically without an associated objective. The theoretical contribution therefore fails as written, and the paper should be reframed as an empirical study of a well-motivated heuristic or accompanied by a corrected derivation.","major_comments":[{"comment":"The derivation of the analytical update step is incorrect. The necessary condition for the x-subproblem in Eq. (5a) is a nonlinear equation in x; for general invertible A it contains an additional term involving A^H A x, and even for unitary A the phase term Ax/|Ax| depends on x. The paper evaluates the phase at the previous iterate z_N and then solves a linear equation for x, obtaining Eq. (7). This is one iteration of a fixed-point map, not a point satisfying the stated necessary condition. A concrete scalar example illustrates the failure: with A=2, y=2, z=0.1, Q=1, Eq. (7) gives x=0.55, while the exact minimizer of Eq. (5a) is x=0.82. Thus the claim that Eq. (7) 'provides the following solution that satisfies the above necessary condition for optimality' is false, and the paper's stated contribution of deriving the update steps from half-quadratic splitting is not established.","section":"Sec. 3, Eqs. (6)-(7)"},{"comment":"The paper presents Plug-and-Play HIO by replacing the error-reduction projection steps of Algorithm 1 with a few HIO inner iterations, but this substitution is not derived from HQS or from any stated objective function. The inner HIO loop is introduced as a heuristic improvement (justified by the known behavior of HIO for nonconvex constraints), and no argument is given that alternating the measurement update, HIO iterations, and the denoiser corresponds to minimizing a common cost. Consequently, the 'mathematical development' announced in the abstract and introduction covers only the simplified PnP-PR update, not the algorithm that actually achieves the reported state-of-the-art results. The authors should either provide a corrected derivation for the HIO variant or explicitly characterize it as an empirically validated heuristic.","section":"Algorithm 2, steps 3-12"}],"minor_comments":[{"comment":"The noise model uses the symbol L for the normal distribution; this should be N, and the relation between the scaling parameter M and the SNR should be defined more explicitly.","section":"Sec. 2, Eq. (1)"},{"comment":"The factor 1/2 appears inconsistently between the objective in Eq. (4) and the subproblem in Eq. (5a); this is cosmetic but should be made consistent.","section":"Sec. 3, Eqs. (4)-(5)"},{"comment":"The plus-or-minus values in Table 1 are not identified as standard deviations or standard errors; please specify the measure of variability.","section":"Sec. 4, Table 1"},{"comment":"The pseudocode uses the inner loop variable Z but refers to iterates as v_{P+1} and v_P; this inconsistent indexing should be corrected to avoid confusion.","section":"Algorithm 2"},{"comment":"The parameters S_max=40, S_min=5, X=5, and W=200 are described as optimized for the M=3 case only; the paper should state explicitly whether the same schedule is used for M=2 and M=4 or whether it was retuned for those noise levels.","section":"Sec. 4, parameter selection"},{"comment":"Because the quantitative conclusions depend on the exact denoiser-switching schedule and parameter schedules, providing the reconstruction code or a fully specified pseudo-code would substantially improve reproducibility; the current data availability statement does not mention code.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The main technical issue is the incorrect derivation in Section 3: Eq. (7) does not solve the stated optimality condition, and the HIO variant in Algorithm 2 is heuristic. The empirical study is extensive and the performance claims are credible, but the paper's framing as a mathematical development overstates what is proven. If the authors correct the derivation or reframe the contribution as an empirical heuristic, a major revision could be sufficient. I would also note that the method relies heavily on the authors' prior conference paper and thesis, and on a pre-trained denoiser from another group; the incremental novelty is mainly in the empirical evaluation and the PnP-HIO variant."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's empirical contribution is real, but the mathematical derivation it claims is flawed. The update in Eq. (7) does not solve the fixed-point condition (6) for the HQS subproblem; it is one iterate of a fixed-point map. The scalar counterexample in the stress-test is correct. So the paper does not actually provide an HQS derivation of the method. What it does provide is a careful, large-scale evaluation of a plug-and-play HIO heuristic, and that evaluation is the paper's strength.\n\nThe benchmark on 236 images across three noise levels is solid: PnP HIO beats HIO, PrDeep, and Iterative DNN-HIO on average PSNR/SSIM for natural images, with runtime comparable to the faster baselines. The robustness to initialization experiment is also informative. The method is simple and could be useful to practitioners. The authors are honest that the algorithm was first presented at COSI 2020 and in a thesis; the novel parts here are the attempted derivation and the evaluation.\n\nThe soft spots are the derivation, the lack of code/data, and the small effect size. The gains over the closest baseline are around 1 dB PSNR, which is not huge but consistent across noise levels. The parameters were tuned on one noise level and then applied to others; that is mild overfitting, but they admit it. The data availability statement says data 'may be obtained from authors upon reasonable request,' which is not a real release.\n\nFor peer review: I would send it out. The empirical claims are substantial enough to justify referee time, and a referee can ask the authors to correct the derivation claim or reframe the method as heuristic. As it stands, the paper should not be accepted without that change. I would not cite the derivation, but the benchmark table might be useful as a reference for comparison. It is a good candidate for a reading group discussion about the gap between theoretical claims and practical algorithms.","headline":"Useful empirical study, but the paper's central HQS derivation does not hold; the method is a heuristic that works well in experiments.","tokens_in":15329,"tokens_out":3269,"would_cite":false,"duration_ms":28962,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Plug-and-Play HIO plugs a deep denoiser into HIO-type projections and reports state-of-the-art Fourier phase retrieval on 236 test images.","keywords":["Fourier phase retrieval","plug-and-play regularization","half-quadratic splitting","hybrid input-output","deep denoiser prior","image reconstruction","alternating projections","noisy intensity measurements"],"falsifier":"Compute the left-hand side of the necessary condition in Eq. (6) at the output $x$ produced by Eq. (7) for a set of test images; if the residual is far from zero and stays nonzero when the update is applied repeatedly, then the claim that Eq. (7) satisfies the optimality condition is false. A direct algorithmic version: replace the single update with several inner applications per outer iteration and see whether PSNR improves substantially; if it does, the one-step solution claim is not the reason the method works.","tokens_in":14242,"feed_emoji":"🔬","tokens_out":10588,"duration_ms":88353,"temperature":0.7,"pith_summary":"The paper develops Plug-and-Play HIO, a phase-retrieval algorithm that alternates half-quadratic-splitting update steps with a pre-trained deep denoiser, and claims it achieves state-of-the-art reconstruction quality for noisy Fourier phase retrieval. On a 236-image test set, the method reports higher average PSNR and SSIM than conventional HIO, PrDeep, and Iterative DNN-HIO at three Poisson noise levels, while running in about half the time of PrDeep. The paper argues that the success comes from treating the data-fidelity subproblem with closed-form measurement and image updates, and using HIO inner iterations rather than error-reduction projections when space-domain constraints are available. This matters because phase retrieval is nonlinear and ill-posed, and the method promises fast, flexible reconstruction without retraining when the measurement model changes.","feed_headline":"Deep denoiser inside HIO lifts phase-retrieval quality","feed_subtitle":"Plug-and-play HIO beats PrDeep and unrolled DNN-HIO in PSNR and SSIM at about half the runtime","key_machinery":"The load-bearing identity is the one-step fixed-point update derived from the necessary condition on the HQS data-fidelity subproblem: $$ x = U_k \\,\\mathrm{Re}\\{$A^{{-1}}$(y \\odot A z_k/|A z_k|)\\} + (1-U_k)z_k. $$ The paper rewrites it as a convex measurement update $\\tilde{y}_k = U_k y + (1-U_k)|A z_k|$ followed by an error-reduction projection using the current phase estimate, and then a deep denoiser for the $z$-subproblem. In Plug-and-Play HIO, the projection block is replaced by several hybrid input-output iterations, a relaxed alternating-projection scheme that updates only the components of the estimate that violate the space-domain constraints. The varying weight $U_k$ and denoiser noise level $\\sigma_k$ decrease with iterations, so the algorithm increasingly trusts the learned prior and the evolving data model.","core_discovery":"On its own terms, the paper's central discovery is that a plug-and-play variant of HIO, derived through half-quadratic splitting, can outperform both classical alternating projections and existing deep-learning phase-retrieval methods on Fourier intensity measurements. The derivation produces the measurement update $\\tilde{y}_k = U_k y + (1 - U_k)|A z_k|$ and the image update $x_{k+1} = \\mathrm{Re}\\{A^{-1}(\\tilde{y}_k \\odot A z_k / |A z_k|)\\}$, followed by the denoiser step $z_{k+1} = D(x_{k+1}, \\sigma_k)$; Eq. (7) presents the combined update as satisfying the necessary optimality condition for the HQS data-fidelity subproblem. When the projection step is replaced by a few hybrid input-output iterations, the resulting Plug-and-Play HIO reports the best average PSNR and SSIM on the 236-image dataset, is less sensitive to initialization and noise, and generalizes to non-natural images even though the denoiser was trained only on natural images. The method inherits the flexibility of projection-based phase retrieval, so the same algorithm can be applied to other magnitude measurements, such as ptychography or coded diffraction patterns, without retraining.","pith_inferences":["Inference: if the paper's derivation is taken literally, the natural stress test is to run several inner fixed-point updates in place of Eq. (7) and check whether PSNR improves; a material improvement would show the single-step update is only an approximation, not the claimed solution of subproblem (5a).","Inference: the same HQS-plus-denoiser recipe should transfer to ptychography or coded-diffraction phase retrieval by swapping the forward operator, since the method already accommodates non-invertible $A$ via the pseudoinverse.","Inference: for complex-valued images, applying a real-valued denoiser separately to amplitude and phase is a heuristic; a denoiser trained on complex-valued data or on the target image distribution would likely narrow the gap to natural-image performance."],"forward_implications":["Plug-and-Play HIO lifts average reconstruction quality on the full test set: at the lowest tested noise it reports 26.35 dB PSNR and 0.751 SSIM, versus 25.78 dB and 0.663 for PrDeep and 25.63 dB and 0.737 for Iterative DNN-HIO.","Reconstruction is fast: about 247 seconds per image versus 475 seconds for PrDeep, making the learned-prior quality gain available at a practical runtime.","Because the data-fidelity step is projection-based and the prior enters only as a denoiser, the algorithm can switch between measurement models (Fourier magnitude, ptychography, coded diffraction patterns) without retraining.","The HIO inner loop is what earns the improvement: Plug-and-Play PR, the error-reduction variant, is clearly worse in PSNR and SSIM, so the relaxation matters.","Parameter robustness: performance remains state-of-the-art at noise levels other than the one used for parameter tuning, and with a single random initialization the method can be made robust by increasing inner and outer iterations."],"supporting_citations":[{"why":"Introduces the hybrid input-output algorithm whose relaxed projections Plug-and-Play HIO builds on.","marker":"[11]"},{"why":"Supplies the PrDeep baseline, the HIO-based initialization procedure, and the natural and unnatural test images.","marker":"[18]"},{"why":"Supplies the Iterative DNN-HIO baseline and the 230-image test set used for evaluation.","marker":"[43]"},{"why":"Provides half-quadratic splitting, the variable-splitting framework from which the analytical update steps are derived.","marker":"[23]"},{"why":"Provides the pre-trained CNN denoiser that serves as the plug-and-play prior in the experiments.","marker":"[25]"},{"why":"Establishes the plug-and-play regularization idea of substituting a denoiser for the prior subproblem.","marker":"[19]"},{"why":"Conference presentation in which Plug-and-Play HIO was first proposed with reported state-of-the-art performance.","marker":"[2]"},{"why":"Gives the uniqueness and oversampling conditions that justify the Fourier phase retrieval measurement setup.","marker":"[9]"}],"fun_headline_variants":["Plug-and-play HIO beats deep unrolled methods in phase retrieval","Deep denoiser inside HIO: half runtime, robust to noise","HIO plus learned prior generalizes beyond training data","Half-quadratic splitting powers new plug-and-play HIO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the single fixed-point update in Eq. (7) actually solves the necessary optimality condition in Eq. (6); one pass of a fixed-point map is not a fixed point in general, so if that step is only an approximation, the derivation's guarantee is not strict.","fun_headline_variants_meta":{"raw":{"variants":["Plug-and-play HIO beats deep unrolled methods in phase retrieval","Deep denoiser inside HIO: half runtime, robust to noise","HIO plus learned prior generalizes beyond training data","Half-quadratic splitting powers new plug-and-play HIO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000674,"raw_usage":{"total_tokens":3091,"prompt_tokens":990,"completion_tokens":2101,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":2029}},"tokens_in":606,"tokens_out":2101,"duration_ms":15889,"temperature":1.0,"reasoning_tokens":2029,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:40:59.120445+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the left-hand side of the necessary condition in Eq. (6) at the output $x$ produced by Eq. (7) for a set of test images; if the residual is far from zero and stays nonzero when the update is applied repeatedly, then the claim that Eq. (7) satisfies the optimality condition is false. A direct algorithmic version: replace the single update with several inner applications per outer iteration and see whether PSNR improves substantially; if it does, the one-step solution claim is not the reason the method works.","supporting_citations":[{"cited_title":"Phase retrieval algorithms: a comparison,","cited_arxiv_id":null,"evidence_quote":"Introduces the hybrid input-output algorithm whose relaxed projections Plug-and-Play HIO builds on."},{"cited_title":"prDeep: Robust phase retrieval with a flexible deep network,","cited_arxiv_id":null,"evidence_quote":"Supplies the PrDeep baseline, the HIO-based initialization procedure, and the natural and unnatural test images."},{"cited_title":"Deep iterative reconstruction for phase retrieval,","cited_arxiv_id":null,"evidence_quote":"Supplies the Iterative DNN-HIO baseline and the 230-image test set used for evaluation."},{"cited_title":"Nonlinear image recovery with half-quadratic regularization,","cited_arxiv_id":null,"evidence_quote":"Provides half-quadratic splitting, the variable-splitting framework from which the analytical update steps are derived."},{"cited_title":"Learning deep CNN denoiser prior for image restoration,","cited_arxiv_id":null,"evidence_quote":"Provides the pre-trained CNN denoiser that serves as the plug-and-play prior in the experiments."},{"cited_title":"Plug-and-play priors for model based reconstruction,","cited_arxiv_id":null,"evidence_quote":"Establishes the plug-and-play regularization idea of substituting a denoiser for the prior subproblem."},{"cited_title":"Model-based phase retrieval with deep denoiser prior,","cited_arxiv_id":null,"evidence_quote":"Conference presentation in which Plug-and-Play HIO was first proposed with reported state-of-the-art performance."},{"cited_title":"The reconstruction of a multidimensional sequence from the phase or magnitude of its Fourier transform,","cited_arxiv_id":null,"evidence_quote":"Gives the uniqueness and oversampling conditions that justify the Fourier phase retrieval measurement setup."}],"review_version":1}