{"id":"0dbf00b3-7b46-4354-97b2-a2d342e44772","arxiv_id":"2501.07308","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"pFPM reinterprets DPC as one Gauss-Newton iteration, then adds annular dark-field patterns and TV regularization to achieve high-resolution quantitative phase imaging with five to six LED-array measurements.","lead":"This paper introduces perturbative Fourier ptychographic microscopy (pFPM), which recasts differential phase contrast as the first step of an iterative Gauss-Newton reconstruction and adds annular dark-field illumination to recover high-resolution phase images from about five measurements.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on an empirically validated but unproven basin-of-attraction assumption for the outer nonconvex Gauss-Newton loop; Appendix D certifies only the inner convex substeps, not convergence from the DPC initialization to true high-frequency content.","rationale":"The reader identified the basin-of-attraction assumption as the weakest point, and my review converges on the same issue after examining the paper's derivations and appendices. The internal-equivalence result in Appendix B is sound, and the inner-solver stepsize conditions in Appendix D are technically correct, but neither supports the outer nonconvex convergence on which the headline claim depends. The simulation and the single USAF experiment provide encouraging evidence, not a guarantee. Because the paper already received a CONDITIONAL verdict based on this concern, my analysis does not move the verdict; it sharpens the reason: the missing piece is not merely 'convergence unproven' but a specific unverified assumption that the DPC estimate lies in the basin of attraction of the proximal Gauss-Newton scheme for realistic samples. The proposed ablation and stress-amplitude test would directly settle whether this assumption holds. No additional objection to novelty or internal consistency was found.","tokens_in":14016,"tokens_out":5172,"duration_ms":57306,"concrete_test":"Using the provided code and the Section 4.1 simulation setup, run DF-pFPM on phase-only objects with phase amplitudes 1.0, 1.5, and 2.0 rad and on objects with added absorption (mu = 0.1 and 0.2), keeping all other settings unchanged. In parallel, ablate the initialization by starting the outer Gauss-Newton loop from a random o0 and from a low-pass-filtered version of the ground truth. If the SNR for strong-phase or absorbing objects drops by more than about 5 dB relative to the reported 17.35 dB, or if different initializations converge to substantially different reconstructions, the basin-of-attraction assumption is the limiting factor and the method's applicability range must be restricted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's key advantage over DPC is the iterative use of dark-field measurements (Section 3.1), where 'subsequent iterations of the proximal Gauss-Newton algorithm do not rely on the weak-object approximation.' That statement is true only if the current estimate ok lies close enough to the true object that the linearization G(ok)+G'(ok)(o-ok) is a reliable surrogate for the nonconvex objective. No argument is provided for this. Appendix D establishes stepsize conditions (equations 42-44) that guarantee convergence of the inner Condat-Vu algorithm for a fixed linearization, and Proposition 6 in [15] only equates the proximal update to the linearized subproblem (36). Neither addresses convergence of the outer proximal Gauss-Newton iterations to a global or even local solution of (33). The practical success is demonstrated on one simulation with phase bounded in [-0.5, 0.5] rad and one USAF phase phantom; the DPC initialization itself has SNR of only 3.24 dB in that simulation. If the object has stronger phase, appreciable absorption, or structure outside the assumed annular illumination geometry, the dark-field iterations could converge to a local minimum or produce streak artifacts rather than recovering high-frequency phase. Because the method's claimed advantage over DPC is precisely the nonlinear refinement enabled by dark-field measurements, this unproven convergence is the most load-bearing assumption in the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes perturbative Fourier ptychographic microscopy (pFPM), an iterative extension of differential phase contrast (DPC) that combines a proximal Gauss-Newton reconstruction with tailored annular dark-field illumination patterns. The authors show that the first proximal Gauss-Newton iteration with quadratic regularization is equivalent to DPC, and then generalize this to multiple iterations and TV regularization. They report that adding two dark-field annuli to the two DPC bright-field patterns enables recovery of higher spatial frequencies, and they demonstrate the method on a simulated cameraman phase object and on an experimental USAF-1951 phantom, with acquisition times around 550 ms versus 126 s for conventional FPM. The appendices contain derivations of the transfer functions, the DPC-equivalence proof, discretization details, the inner solver algorithm, and comparisons with alternative illumination patterns.","tokens_in":14302,"tokens_out":5417,"duration_ms":49710,"significance":"If the reported results hold, pFPM provides a practical speed-resolution tradeoff for LED-array microscopes: it retains the speed and simplicity of DPC while extending resolution into the dark-field region with only five measurements and no learned components. The paper's derivation of the transfer functions and the identification of DPC as the first Gauss-Newton step are clean and self-contained, and the authors ship code and data for reproducibility. The main caveat is that the outer nonconvex Gauss-Newton loop lacks a convergence or basin-of-attraction analysis, so the central claim that the method 'does not rely on the weak-object approximation' is currently only empirically supported for the tested objects. Overall, this is a solid, clearly presented contribution to fast quantitative phase imaging, with a limitation that should be addressed in revision.","major_comments":[{"comment":"The statement that 'subsequent iterations of the proximal Gauss-Newton algorithm do not rely on the weak-object approximation' in Section 3.1 is not supported by the provided analysis. The proximal Gauss-Newton update (13) remains a local method: its accuracy depends on how well G(ok)+G'(ok)(o-ok) approximates G(o) near the current iterate. Appendix D proves only that the inner Conduit-Vu stepsizes (42)-(44) satisfy the convex-convergence condition, and Proposition 6 of [15] merely identifies the proximal update with the linearized subproblem (36). No result is given on whether the DPC initialization lies in the basin of attraction of the outer nonconvex iteration or whether the iterates avoid local minima. Because the resolution advantage over DPC is explicitly attributed to this nonlinear refinement, this missing convergence argument is load-bearing. I recommend either adding a local convergence analysis (e.g., regularity conditions on G and a neighborhood estimate around the DPC solution) or reformulating the claim to state that the method does not rely on the weak-object approximation only where the linearized model remains accurate, with convergence validated empirically.","section":"Section 3.1 / Appendix D"}],"minor_comments":[{"comment":"The caption in Figure 3 states 'SNR and RMSE in Fourier space', but the definitions in (15) and (16) are image-domain metrics applied to the phase of the reconstruction. Please correct the caption or the equations to avoid confusion.","section":"Section 4.1 / Figure 3 caption"},{"comment":"The identification of DPC as the first proximal Gauss-Newton iteration requires the phase-only and zero-mean assumptions stated at the end of Appendix B. This assumption should be stated explicitly in Section 3.1 where the claim is made, so that readers do not infer the equivalence for general complex objects with absorption.","section":"Section 3.1 / Appendix B"},{"comment":"The comparison with conventional FPM is informal: the FPM reconstruction in Figure 6 is obtained by 50 iterations of a simple gradient descent and shows visible artifacts. Please report the number of FPM measurements, the reconstruction settings, and ideally a quantitative comparison metric (e.g., line-profile error or correlation with the DF-pFPM result) to support the 126 s versus 550 ms comparison.","section":"Section 4.2 / Figure 6"},{"comment":"The inner solver is run for a fixed K=100 iterations with warm start, which makes the method inexact. Please state explicitly whether the reported reconstructions use this inexact setting and whether the outer iteration count (8, or 4 for BF-pFPM*) was chosen by performance on validation data.","section":"Appendix D / Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a good fit for the journal and the core algorithmic contribution is interesting. The main revision point is the overstatement in Section 3.1 regarding the weak-object approximation; a tempered claim and an explicit discussion of the empirical nature of the convergence would make the paper acceptable. The comparison to FPM acquisition time is a bit promotional, and should be made fairer by reporting quantitative results for the FPM run. No concerns about novelty or citation practices."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid methods paper that earns its central claim despite one unproven theoretical step. The new idea is to read DPC as the first proximal Gauss-Newton step and then keep iterating, using dark-field annuli to push the recovered spectrum outward. That interpretation is derived cleanly in Appendix B, not asserted, and the annular patterns are motivated and tested against sectors and random patterns. The simulation and the USAF phantom experiment both show real gains: DF-pFPM resolves Group 9 element 5 (with hints of element 6) at five measurements, with a 550 ms acquisition versus 126 s for conventional FPM in the reported setup. Code and data are public. That is real evidence.\n\nThe stress-test concern about convergence is on target but not damning. Appendix D certifies only the inner convex subproblems; there is no proof that the outer nonconvex Gauss-Newton loop converges from the DPC initialization to the true high-frequency content. The claim that subsequent iterations \"do not rely on the weak-object approximation\" should be read as conditional on the linearization around the current estimate being a good surrogate. Empirically it works for phase bounded at ±0.5 rad and one phantom, but there is no proof, no sensitivity sweep over phase amplitude, absorption, or object structure, and no error bars on the experimental numbers. These are real limitations, but they are the usual kind for a computational imaging paper, not evidence of a broken method.\n\nThe paper is honest about its own scope. It explicitly discourages use of the asterisked variants, reports the limitations of unregularized reconstruction, and credits prior work on multiplexed FPM, learned illumination, and proximal Gauss-Newton appropriately. The acquisition-time comparison is meaningful, and the TV-versus-L2 comparison in Appendix E is a nice practical touch.\n\nWho gets value: anyone working on LED-array phase imaging, particularly for dynamic or live-cell samples where acquisition speed matters. The paper deserves a serious referee. I would send it out and ask for additional experiments, a sensitivity analysis over object phase strength and iteration counts, and a more careful statement of what the convergence guarantee would require. But desk rejection would be wrong.","headline":"A clean, well-documented extension of DPC to iterative dark-field Gauss-Newton with a real speed-resolution win; the main risk is the unproven basin-of-attraction assumption, but the empirical case is strong enough to referee.","tokens_in":14831,"tokens_out":2104,"would_cite":true,"duration_ms":21826,"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":"Five illuminations match FPM resolution at DPC speed.","keywords":["Fourier ptychographic microscopy","differential phase contrast","quantitative phase imaging","proximal Gauss-Newton","dark-field illumination","LED array microscope","phase retrieval","computational imaging"],"falsifier":"Image a strong phase object with phase excursions well beyond 0.5 rad, or with significant absorption, under the same five illumination patterns; if DF-pFPM fails to resolve the high-frequency elements it resolves for weak objects, or produces streak artifacts that do not diminish with more iterations, the assumed basin of attraction is too optimistic.","tokens_in":13790,"feed_emoji":"🔬","tokens_out":4201,"duration_ms":40810,"temperature":0.7,"pith_summary":"Perturbative Fourier ptychographic microscopy (pFPM) claims that differential phase contrast (DPC) is only the first step of a more general iterative scheme: interpret DPC as the initial iteration of a proximal Gauss-Newton algorithm, add dark-field illumination, and the same fast few-measurement setup recovers the high-frequency phase that DPC misses. The paper argues that with five illumination patterns — the classic DPC antisymmetric bright-field pair, one full bright-field measurement, and two dark-field annuli — the method reaches resolutions comparable to conventional Fourier ptychographic microscopy while cutting acquisition time from 126 s to 550 ms. If correct, this turns a two-measurement low-resolution technique into a five-measurement high-resolution one without any learned components or hardware changes beyond the already-programmable LED array.","feed_headline":"Five illuminations match FPM resolution at DPC speed","feed_subtitle":"Two dark-field annuli push the DPC reconstruction beyond its bright-field limit, cutting acquisition from 126 s to 550 ms.","key_machinery":"The proximal Gauss-Newton algorithm with quadratic or total-variation regularization, in which each iteration solves a regularized linear least-squares problem around the current estimate rather than around a fixed weak-object assumption; the paper proves that the first such iteration starting from $o=1$ reproduces DPC exactly. The tailored illumination strategy is the second half: two annular dark-field patterns spanning $[1,1.5]\\nu_{\\text{obj}}$ and $[1.5,2]\\nu_{\\text{obj}}$ are chosen to 'push' recovered Fourier frequencies outward, and the paper recommends two annuli because more give diminishing returns.","core_discovery":"The paper establishes that DPC is exactly the first iteration of a proximal Gauss-Newton algorithm starting from the uniform-transmission guess $o=1$ with quadratic regularization, and then removes the weak-object restriction by running more iterations. Because later iterations do not rely on the weak-object approximation, the same linearized-update structure can be applied to dark-field measurements, whose scattered light carries the high spatial frequencies that DPC's bright-field-only model discards. Adding two annular dark-field illumination patterns, spanning $[1,1.5]\\nu_{\\text{obj}}$ and $[1.5,2]\\nu_{\\text{obj}}$, allows the recovered Fourier support to be 'pushed' outward, yielding simulated SNR 17.35 dB and experimentally resolving USAF Group 9 Element 5 with hints of Element 6, in roughly 5 total measurements and 550 ms acquisition time.","pith_inferences":["The same interpretative move — linearize, iterate, then add dark-field patterns — could be applied to other linearized phase-retrieval techniques, turning any single-shot linear inversion into a multi-iteration nonlinear refinement with few extra measurements.","Because each annulus corresponds to a controlled Fourier-support extension, the design suggests a direct tradeoff curve between measurement count and resolution that could be probed systematically for X-ray or electron imaging, where LED-array-style illumination geometries differ but the Fourier-support logic persists.","The paper's own comparison shows diminishing returns beyond two annuli, implying that the practical limit of this scheme may be set by model mismatch and noise rather than by information content, so better forward models (including pupil aberrations) could push resolution further at the same measurement count.","A natural testable extension is to replace the hand-set regularization with a data-driven prior inside the same proximal Gauss-Newton loop; if that prior keeps the high-frequency content while reducing artifacts, the method could tolerate stronger or more absorbing objects than the currently demonstrated weak-phase regime."],"forward_implications":["DPC users can upgrade to near-FPM resolution with only three additional exposures and no change of hardware, provided their microscope has a programmable LED array.","The identification of DPC as the first Gauss-Newton iteration gives a principled way to design illumination patterns: each dark-field annulus extends the recovered Fourier support by a controlled amount, replacing ad-hoc or learned pattern choices.","Total-variation regularization inside the proximal Gauss-Newton framework suppresses the high-frequency noise and ringing that appear in L2-regularized DPC reconstructions, as demonstrated on the USAF phantom.","A five-pattern acquisition is fast enough for time-lapse phase imaging of moving or live samples, where the 126 s needed for conventional FPM at 1 s exposure per LED is impractical."],"supporting_citations":[{"why":"Defines conventional Fourier ptychographic microscopy and its single-LED acquisition scheme, which sets the high-resolution baseline and the hundreds-of-measurements cost that pFPM aims to reduce.","marker":"[24]"},{"why":"Introduces quantitative DPC in an LED-array microscope, supplying the linearized measurement model, the antisymmetric bright-field illumination patterns, and the weak-object assumption that pFPM generalizes.","marker":"[16]"},{"why":"Establishes multiplexed coded illumination for Fourier ptychography, providing the multiplexed measurement model and the random-pattern baseline that pFPM compares against.","marker":"[18]"},{"why":"Provides the convergence analysis of the proximal Gauss-Newton method that justifies the iterative update scheme used by pFPM.","marker":"[15]"},{"why":"Shows the benefit of separating bright-field and dark-field patterns in computational illumination, a direct precedent for pFPM's hybrid pattern design.","marker":"[17]"},{"why":"Proposes a five-pattern acquisition (two DPC bright-field plus three dark-field sectors) with a data-driven reconstruction; pFPM uses the same measurement budget with interpretable annuli and outperforms it.","marker":"[20]"},{"why":"Its learned dark-field illumination patterns are observed to be symmetric and clustered, motivating the partition of the dark-field region into annuli rather than sectors.","marker":"[7]"}],"fun_headline_variants":["Five illuminations, 550 ms: FPM resolution at DPC speed","Two dark-field annuli unlock FPM-class phase at DPC speed","From DPC to pFPM: 5 measurements, 550 ms, FPM resolution","Quantitative phase imaging in 550 ms with FPM resolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The initial DPC estimate, obtained by linearizing around a uniform transparent object, must lie in the region where the nonlinear Gauss-Newton iterations converge to the true high-frequency content rather than to a local minimum, with only five multiplexed measurements and no proof for the outer nonconvex iterations.","fun_headline_variants_meta":{"raw":{"variants":["Five illuminations, 550 ms: FPM resolution at DPC speed","Two dark-field annuli unlock FPM-class phase at DPC speed","From DPC to pFPM: 5 measurements, 550 ms, FPM resolution","Quantitative phase imaging in 550 ms with FPM resolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000552,"raw_usage":{"total_tokens":2618,"prompt_tokens":915,"completion_tokens":1703,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":1620}},"tokens_in":531,"tokens_out":1703,"duration_ms":13487,"temperature":1.0,"reasoning_tokens":1620,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:44:08.482464+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Image a strong phase object with phase excursions well beyond 0.5 rad, or with significant absorption, under the same five illumination patterns; if DF-pFPM fails to resolve the high-frequency elements it resolves for weak objects, or produces streak artifacts that do not diminish with more iterations, the assumed basin of attraction is too optimistic.","supporting_citations":[{"cited_title":"Wide-field, high-resolution Fourier ptychographic microscopy","cited_arxiv_id":null,"evidence_quote":"Defines conventional Fourier ptychographic microscopy and its single-LED acquisition scheme, which sets the high-resolution baseline and the hundreds-of-measurements cost that pFPM aims to reduce."},{"cited_title":"Quantitative differential phase contrast imaging in an LED array microscope","cited_arxiv_id":null,"evidence_quote":"Introduces quantitative DPC in an LED-array microscope, supplying the linearized measurement model, the antisymmetric bright-field illumination patterns, and the weak-object assumption that pFPM generalizes."},{"cited_title":"Multiplexed coded illumination for Fourier Ptychography with an LED array microscope","cited_arxiv_id":null,"evidence_quote":"Establishes multiplexed coded illumination for Fourier ptychography, providing the multiplexed measurement model and the random-pattern baseline that pFPM compares against."},{"cited_title":"Convergence analysis of a proximal Gauss-Newton method","cited_arxiv_id":null,"evidence_quote":"Provides the convergence analysis of the proximal Gauss-Newton method that justifies the iterative update scheme used by pFPM."},{"cited_title":"Reliable deep-learning-based phase imaging with uncertainty quantification","cited_arxiv_id":null,"evidence_quote":"Proposes a five-pattern acquisition (two DPC bright-field plus three dark-field sectors) with a data-driven reconstruction; pFPM uses the same measurement budget with interpretable annuli and outperforms it."},{"cited_title":"Data-driven design for Fourier ptychographic microscopy","cited_arxiv_id":null,"evidence_quote":"Its learned dark-field illumination patterns are observed to be symmetric and clustered, motivating the partition of the dark-field region into annuli rather than sectors."}],"review_version":1}