{"id":"47f9f974-54f4-4177-a397-8b7b148fb8a8","arxiv_id":"2412.02548","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A half-quadratic splitting plug-and-play method for ptychography achieves better amplitude and phase reconstruction than classical PIE algorithms at reduced overlap and high noise.","lead":"This paper derives a plug-and-play half-quadratic splitting algorithm that reconstructs complex-valued images from ptychography measurements, with an explicit closed-form data consistency step. It demonstrates on simulated data that the method keeps reconstruction quality high at low probe overlap and high noise, which could cut scanning time in coherent diffraction imaging.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central practical claim rests on simulated data plus an unvalidated real/imag denoiser heuristic; no real ptychographic measurements support the low-overlap conclusion.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing point: the complex-valued denoiser construction in Section 2.3 is a heuristic proxy for the true complex proximal operator, and it is never validated on real ptychographic data. My reading of the paper agrees that the mathematical derivation of HQS is clean and the simulated comparisons are internally consistent, so this is not a reason to reject the work outright. However, the central practical claim about lower overlap and noise robustness depends on the transfer of a natural-image denoiser to complex-valued reconstructions through an ad hoc real/imag decomposition. The paper's own conclusion acknowledges that probe reconstruction and realistic ptychography systems remain future work, and the abstract's 'real test objects' wording is not supported by the experiments. Because the reader already conditioned acceptance on addressing this gap, my stress-test does not move the verdict; the requested validation on real data or at least a head-to-head comparison with a genuinely complex-valued denoiser would settle whether the concern lands.","tokens_in":9785,"tokens_out":6470,"duration_ms":68563,"concrete_test":"Run the same PnP-HQS pipeline on a real ptychographic dataset (for example, synchrotron measurements of a test object with an independently reconstructed ground truth) at 38% overlap and a comparable noise level, and report amplitude and phase PSNR against SimPIE. If the DRUNet gain over SimPIE vanishes or reverses, the complex denoiser heuristic does not transfer. As a complementary check, rerun the BSD500 experiments with a complex-valued denoiser (e.g., applied to magnitude and phase, or a CNN trained on complex images) and compare with the real/imag DRUNet results in Table 1; a large gap would isolate the heuristic as the cause.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The claim that PnP-HQS 'was able to produce good reconstructions while using lower probe overlap than classical methods' is an empirical statement about practical ptychography, but every quantitative result is generated from simulated measurements on synthetic complex-valued images built from natural RGB images (Section 3.1) or from a brain phantom with simulated measurements (Section 3.2). The abstract's phrase 'real test objects' overstates this evidence. The complex-valued proximal step is implemented by splitting into real and imaginary parts, adding a constant equal to the maximum amplitude, and applying a pre-trained real-valued DRUNet independently to each part (Section 2.3). This is only a heuristic proxy for the complex proximal operator: DRUNet was trained on natural grayscale images in a standard intensity range, and the shift-and-scale preprocessing plus independent-component treatment has no supporting experiment. If the heuristic fails on realistic samples, the central performance advantage over PIE would not transfer, and the paper provides no real diffraction data to rule this out.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a plug-and-play half-quadratic splitting (PnP-HQS) algorithm for ptychographic phase retrieval. The authors derive an explicit solution for the data-fidelity subproblem in Fourier phase retrieval, extend it to ptychography via multiple auxiliary variables, and use the resulting update with classical (TV, WCRR) and neural (DRUNet) denoisers. For complex-valued images, they construct the denoiser step by applying real-valued denoisers separately to the real and imaginary parts after a constant shift. Experiments on synthetic complex images from BSD500 and on a brain phantom show consistent PSNR improvements over unregularized sequential and simultaneous PIE, particularly at low overlap and high noise.","tokens_in":9991,"tokens_out":10518,"duration_ms":103270,"significance":"If the results hold, the paper makes a useful contribution by showing that PnP-HQS, with its closed-form data-consistency updates, is a practical and flexible framework for ptychographic reconstruction with reduced overlap. The derivation in Section 2 is algebraically sound, and the evaluation covers overlap and noise sweeps with multiple denoisers; the use of equivariant evaluation for DRUNet is a positive methodological detail. The significance is tempered by the fact that the quantitative evidence is entirely simulation-based and by two heuristics (the real/imaginary-part complex denoiser and the D²≈γI approximation for DRUNet) that are not validated or quantified. The paper would be strengthened by tests against a true complex-valued proximal/denoiser and by a clearer statement of the simulation-only scope of the low-overlap claim.","major_comments":[{"comment":"The complex-valued denoiser is implemented by adding a constant equal to the maximum amplitude to the real and imaginary parts, applying a real-valued pre-trained denoiser to each part independently, and (implicitly) subtracting the constant afterward. This is a heuristic proxy for the complex proximal operator, and it is load-bearing: every PnP result in Tables 1–3 uses this construction. The paper provides no evidence that this proxy behaves like a true complex-valued proximal mapping, and the text does not specify the subtraction step explicitly. Please either validate the heuristic (for instance, by comparing with an explicitly complex proximal operation for a known regularizer, or with a magnitude-phase decomposition) or soften the claims that depend on it.","section":"Section 2.3"},{"comment":"For the DRUNet experiments, the spatially varying operator D is replaced by γI through the approximation D²≈γI, but the approximation error is never quantified. For the binary circular probe used in Section 3.1, D is not a multiple of the identity: it takes different values in the interior, the partially illuminated boundary, and the corners of the reconstructed field. Since all DRUNet results in Tables 1 and 2 rely on this approximation, the paper should quantify its effect (e.g., by comparing with the exact spatially varying update used for TV and WCRR on a subset of images) or justify that the chosen τ_k schedule makes the reconstruction insensitive to it.","section":"Section 2.2, Remark 1"},{"comment":"The abstract states that the method is evaluated on 'real test objects,' and the conclusion states that PnP-HQS 'was able to produce good reconstructions while using lower probe overlap than classical methods.' However, all quantitative experiments in Section 3 use simulated measurements: the BSD500 images are synthetic complex images with a simulated circular probe, and the brain phantom in Section 3.2 also uses simulated measurements ('We simulate measurements using a circular probe'). No real ptychographic diffraction data are presented. Please either add real-data experiments or revise the abstract/conclusion to state explicitly that the low-overlap claim is demonstrated only in simulation.","section":"Abstract, Section 3.2, Section 4"}],"minor_comments":[{"comment":"The formula 'c = n/(n+µσ²)' contains an undefined σ²; it should be c = n/(n+µ) to match Algorithm 1 and the derivation.","section":"Section 2.1, after Eq. (5)"},{"comment":"The text says 'The results are in Table 1,' but the noise-sweep results are in Table 2; please correct the cross-reference.","section":"Section 3.1, 'Influence of Noise'"},{"comment":"The paper does not state how many BSD500 images are used for testing or how the validation set is split; please specify the number of test images and random seeds so that the standard deviations in Tables 1 and 2 are interpretable.","section":"Section 3.1"},{"comment":"The claim that the algorithm recovers simultaneous PIE when R is the indicator of the nonnegative orthant is imprecise: SimPIE's update is a step-size-weighted averaging update, whereas the HQS update with R=ι_{[0,∞)ⁿ} is a projection of the auxiliary least-squares solution; the two coincide only under additional assumptions, which should be stated.","section":"Remark 1"},{"comment":"The denoising procedure for complex images is not fully specified: after adding the maximum-amplitude constant, applying the real-valued denoiser to each component, and recombining, the constant must be subtracted; the paper should state this explicitly to ensure reproducibility.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable algorithmic contribution with a sound derivation, but the central practical claim (low-overlap advantage) rests on simulated measurements and on two unvalidated heuristics. In revision, the authors should either add validation for the complex denoiser and the D²≈γI approximation, or substantially temper the abstract and conclusion. The self-citations are appropriate and do not carry the central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nThe paper is a clean, useful extension of PnP-HQS to ptychography. The genuinely new piece is the analytic data-consistency update per probe position (Eq. 5) inside a multi-auxiliary-variable HQS scheme, plus the way the overlap weighting is folded into a spatially varying denoising problem (Eq. 9). The derivations are correct and the experiments are well organized: overlap sweep, noise sweep, several denoisers, and a brain phantom. On the simulated tests, the claim that PnP-HQS beats unregularized PIE at low overlap and high noise holds up.\n\nThe soft spots are real but not disqualifying. First, the abstract says 'real test objects,' but every quantitative result is simulated; the brain phantom is a phantom with simulated measurements, not measured data. That overstates the evidence. Second, the complex denoiser is a heuristic: decompose into real and imaginary parts, add a constant equal to the maximum amplitude, and apply a pre-trained real-valued DRUNet to each part. No experiment checks how close this is to a true complex proximal operator, so the practical transfer is unproven. Third, Remark 1's D²≈γI approximation is used for the DRUNet runs without quantification; this may be acceptable in the tested uniform-probe regime but should be justified or tested. Fourth, there are no comparisons to earlier PnP phase-retrieval baselines like PnP-FISTA or PnP-FASTA, which would be natural competitors. Finally, no code or data is provided, which makes reproducibility harder to assess.\n\nNone of this sinks the paper. The math is solid, the method is clearly explained, and the simulated evidence supports the central algorithmic claim. What is missing is experimental validation on real ptychographic data and a sharper statement about the complex denoiser's limitations. The paper deserves a serious referee. I would send it to review, but the revision should add a real-data experiment, or substantially temper the abstract and validate the complex denoiser against a known reference.","headline":"A clean PnP-HQS extension for ptychography with a correct analytic data-fidelity step, solid simulated evidence, but an overstated abstract and an unvalidated complex-denoiser heuristic.","tokens_in":10521,"tokens_out":2569,"would_cite":true,"duration_ms":24200,"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":"Plug-and-play denoisers beat classical ptychography at low overlap.","keywords":["ptychography","phase retrieval","half-quadratic splitting","plug-and-play","image denoising","complex-valued imaging","low overlap","regularisation"],"falsifier":"Run the algorithm on a real ptychography dataset with an independently known ground truth, such as a test pattern reconstructed at very high overlap with ePIE, and compare amplitude and phase PSNR against SimPIE at 38% overlap. If the reported performance gap disappears or visible phase artifacts appear, the complex-denoiser heuristic would be the likely culprit.","tokens_in":9584,"feed_emoji":"🔬","tokens_out":4101,"duration_ms":36619,"temperature":0.7,"pith_summary":"This paper extends half-quadratic splitting, a standard variational-reconstruction scheme, to ptychography, and shows that replacing the regularisation step with a plug-and-play denoiser yields reconstructions that beat classical ptychographic algorithms. The authors derive a closed-form solution for the data-consistency subproblem, so each iteration is cheap, and they handle complex-valued images by denoising the real and imaginary parts separately with off-the-shelf greyscale denoisers. In simulated experiments on natural images and a brain phantom, the approach keeps quality high where classical methods degrade: at only 38% probe overlap and under strong shot noise. The practical payoff is that scanning time could be cut while keeping reconstruction quality.","feed_headline":"Plug-and-play denoisers beat classical ptychography at low overlap","feed_subtitle":"Half-quadratic splitting with pre-trained denoisers keeps image quality when overlap and noise are challenging.","key_machinery":"The load-bearing object is the half-quadratic splitting formulation with multiple auxiliary variables $z_\\ell = A_\\ell x$, one per probe position. It reduces each data-consistency update to the same closed-form Fourier phase retrieval step: in Fourier space, set the amplitude of $\\hat{z}$ to a convex combination of the measured amplitude and the current amplitude, and keep the current phase. The regularisation update is a Gaussian denoising problem with a spatially varying noise level controlled by the accumulated probe intensity $D^2 = \\sum_\\ell |A_\\ell|^2$; under a regularity assumption this reduces to a standard denoiser applied to the probe-weighted average of the auxiliary images. Complex values are handled by splitting into real and imaginary parts, shifting each by a constant equal to the maximum amplitude so that pretrained non-negative greyscale denoisers apply.","core_discovery":"The central claim is that plug-and-play half-quadratic splitting (PnP-HQS) provides a stable, flexible reconstruction method for ptychography that outperforms unregularised PIE variants, particularly at low overlap and high noise. In the paper's formulation, the ptychographic forward model is split into one auxiliary variable per probe position; the resulting z-subproblems are identical to Fourier phase retrieval and admit an explicit closed-form update using the measured amplitudes and the phase of the current iterate. The x-subproblem becomes a spatially varying Gaussian denoising task, which can be solved either classically (TV, weakly convex ridge regulariser) or by a pretrained deep denoiser applied independently to real and imaginary channels. Tables 1 and 2 report consistent PSNR gains over Simultaneous PIE and Sequential PIE, with the gap widening as overlap drops from 68% to 38% and as noise level alpha rises from 10 to 40. The paper concludes that this is a first step toward shorter acquisition times.","pith_inferences":["The real/imaginary decomposition with a constant shift is not validated on experimentally measured complex samples; testing a true complex-valued proximal operator instead would isolate whether this heuristic limits performance.","The same HQS derivation should transfer to other non-linear Fourier measurements, such as Fourier ptychography with coded illuminations, because the data-consistency step only relies on the magnitude-Fourier structure.","Low-overlap robustness suggests PnP-HQS could act as an initialisation or regulariser inside joint probe-object estimation, potentially reducing the risk of local minima in that harder problem.","All experiments are simulated with a known binary circular probe; testing on real beam profiles, where probe uncertainty and position errors enter, is the natural next comparison."],"forward_implications":["At 38% overlap, PnP-HQS with DRUNet or WCRR reconstructs amplitude and phase with several dB higher PSNR than SimPIE, so the required scan overlap could be relaxed.","The closed-form z-update keeps per-iteration cost low enough that 600 iterations suffice, versus 2000 iterations for the PIE baselines.","The method inherits the behaviour of any chosen denoiser, so better datasets or denoisers directly improve reconstruction without changing the algorithm.","Because the regularisation step is generic, the same framework covers TV, learned convex regularisers, and deep denoisers within one code path.","Since the probe is assumed known, the authors point to simultaneous probe and object reconstruction as the next step toward realistic ptychography."],"supporting_citations":[{"why":"Introduces the plug-and-play framework of replacing proximal operators by denoisers, which the paper builds on.","marker":"[35]"},{"why":"Supplies the half-quadratic splitting scheme that the paper adapts to ptychography.","marker":"[9]"},{"why":"Defines the Simultaneous PIE algorithm that serves as the main classical baseline.","marker":"[18]"},{"why":"Shows how PnP can be applied to Fourier phase retrieval with the same noise model and data-fidelity term used here.","marker":"[23]"},{"why":"Provides the DRUNet denoiser and the noise-level schedule that the paper adopts for its deep plug-and-play variant.","marker":"[40]"},{"why":"Defines the weakly convex ridge regulariser (WCRR) used as one of the non-deep denoisers.","marker":"[11]"},{"why":"Prior work applying PnP to Fourier ptychography, which the paper extends to the standard ptychographic forward model.","marker":"[33]"}],"fun_headline_variants":["Plug-and-play priors sharpen ptychography at low overlap","PnP half-quadratic splitting boosts ptychographic reconstruction","Pre-trained denoisers reclaim ptychography under noise","Half-quadratic splitting makes ptychography robust to low overlap","Plug-and-play denoisers outdo PIE in challenging ptychography"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that denoising the real and imaginary parts independently, after shifting them to be non-negative, is a good stand-in for a true complex-valued denoiser, and the paper gives no experimental data validating this on real samples.","fun_headline_variants_meta":{"raw":{"variants":["Plug-and-play priors sharpen ptychography at low overlap","PnP half-quadratic splitting boosts ptychographic reconstruction","Pre-trained denoisers reclaim ptychography under noise","Half-quadratic splitting makes ptychography robust to low overlap","Plug-and-play denoisers outdo PIE in challenging ptychography"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1265,"prompt_tokens":881,"completion_tokens":384,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":295}},"tokens_in":497,"tokens_out":384,"duration_ms":3488,"temperature":1.0,"reasoning_tokens":295,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:19:02.148855+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the algorithm on a real ptychography dataset with an independently known ground truth, such as a test pattern reconstructed at very high overlap with ePIE, and compare amplitude and phase PSNR against SimPIE at 38% overlap. If the reported performance gap disappears or visible phase artifacts appear, the complex-denoiser heuristic would be the likely culprit.","supporting_citations":[{"cited_title":"In: IEEE GlobalSIP","cited_arxiv_id":null,"evidence_quote":"Introduces the plug-and-play framework of replacing proximal operators by denoisers, which the paper builds on."},{"cited_title":"IEEE Transactions on Image Processing4(7), 932–946 (1995)","cited_arxiv_id":null,"evidence_quote":"Supplies the half-quadratic splitting scheme that the paper adapts to ptychography."},{"cited_title":"Ultrami- croscopy 171, 43–54 (2016)","cited_arxiv_id":null,"evidence_quote":"Defines the Simultaneous PIE algorithm that serves as the main classical baseline."},{"cited_title":"In: ICML","cited_arxiv_id":null,"evidence_quote":"Shows how PnP can be applied to Fourier phase retrieval with the same noise model and data-fidelity term used here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DRUNet denoiser and the noise-level schedule that the paper adopts for its deep plug-and-play variant."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the weakly convex ridge regulariser (WCRR) used as one of the non-deep denoisers."},{"cited_title":"In: IEEE ICASSP","cited_arxiv_id":null,"evidence_quote":"Prior work applying PnP to Fourier ptychography, which the paper extends to the standard ptychographic forward model."}],"review_version":1}