{"id":"8d3c31a8-3015-4fa7-a082-286799697463","arxiv_id":"1908.07461","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A sliding-window inference method exploiting Fisher-information locality reconstructs compound objects from higher-order photon correlation measurements with linear complexity and experimentally demonstrated super-resolution.","lead":"Researchers built an image-reconstruction algorithm that solves hard estimation problems piece by piece, using the structure of the measurement to decide which parts of the object can be treated independently. They show it beats the classical resolution limit in low-light quantum imaging with correlated photon sources, and they predict and test an optimal amount of photon correlation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The linear-complexity claim presumes a fixed window-border size, but in the super-resolution regime the FIM bandwidth n0=Δl/d grows as the pixel grid is refined, so per-window cost and truncation error are not controlled.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the FIM bandwidth n0 = Δl/d grows as the pixel size d is reduced below the Rayleigh limit, so the window border cannot remain a small constant in the super-resolution regime. This concern is supported by the manuscript itself: Supplementary Note 3 defines n0, and the Methods pseudo-code explicitly truncates probabilities to the current window with no stated error bound. The experimental sections demonstrate the method at moderate super-resolution with a fixed detector array, which is genuine and useful evidence, but it does not test the claimed asymptotic scaling. The optimal-correlation-width prediction is separately experimentally supported and is not affected by this concern. Since the central algorithmic complexity claim remains conditional on a locality assumption that is not established for arbitrarily fine pixel grids, the conditional verdict is appropriate and no change is needed.","tokens_in":18335,"tokens_out":7818,"duration_ms":87673,"concrete_test":"Take the 1D pseudo-thermal model of Supplementary Note 3 with a fixed object length and Rayleigh width Δl, and refine the pixel grid from d = Δl to d = Δl/2, Δl/4, and Δl/8 while scaling the detector resolution accordingly. Measure the total SWM runtime and the reconstruction error as functions of M. If runtime grows superlinearly in M, or if error rises because n0 = Δl/d grows with M, the linear-complexity claim fails in the super-resolution regime. An independent analytic count of the required border size as a function of d, using the D-coefficient locality conditions of Eq. (20), would settle the same point without running the experiment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's efficiency claim, stated as 'the complexity of the SWM is linear on the number of shifts required to cover the whole set of parameters' and in the abstract as 'linear on the total number of parameters', presumes that the sliding window has a border whose size is independent of the total parameter count and that pixels outside the window contribute negligibly. In the near-field application, Supplementary Note 3 defines n0 = Δl/d, the width of the PSF measured in object pixel units. For super-resolution, d is chosen below Δl, and for a fixed physical object the pixel count M grows as d shrinks, so n0 = Δl/d grows proportionally to M. The pseudo-code in Methods then requires the window border to cover this growing number of FIM bands: each local optimization involves at least O(n0) unknown or known pixels, and the number of shifts falls only as M/n0. Consequently the total cost is not linear in M unless n0 is artificially held fixed while the grid is refined. The same growth undermines the truncation step, which computes theoretical probabilities 'assuming all pixels but those inside the window to be zero': the omitted-pixel error is governed by the same PSF tails that set n0, and no error bound is given. The experiments use a fixed 32×32 detector array and a moderate super-resolution factor, so they do not exercise the scaling regime where the problem is claimed to become hard.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an iterative sliding-window method (SWM) for nonlinear parameter estimation in problems with parametrically local measurements, i.e., measurements where each observed probability depends on a limited subset of the unknown parameters. The authors connect this locality to the banded structure of the Fisher information matrix (FIM), use the FIM to choose the window and border sizes, and apply the method to near-field quantum imaging with higher-order correlation functions. They derive the relevant detection coefficients for pseudo-thermal and SPDC sources, demonstrate reconstructions on simulated and experimental data, predict an optimal source correlation width from the trace of the inverse FIM, and experimentally confirm the prediction. They additionally discuss how estimation bias arising from parameter constraints can improve resolution.","tokens_in":18707,"tokens_out":6212,"duration_ms":65989,"significance":"If the linear-complexity claim is sustained, the SWM is a potentially valuable tool for quantum imaging and tomography with large parameter spaces. The paper has several concrete strengths: the FIM-based locality analysis is clearly presented, the pseudo-code is implementable, and the experimental demonstrations with a 32x32 SPAD array for both pseudo-thermal and SPDC sources are credible. The prediction of an optimal correlation width from the Fisher-information model and its confirmation with experimental reconstruction infidelity (Fig. 4a,b) is a nontrivial, falsifiable result. The biased-estimation discussion correctly draws on the constrained-estimation literature. However, the central scaling claim is not rigorously established for the super-resolution regime, and the window-truncation approximation lacks an error bound; these issues need to be addressed before the claimed complexity advantage can be accepted.","major_comments":[{"comment":"The central complexity claim, stated in the abstract as 'this iterative scheme is linear on the total number of parameters,' is not established for the super-resolution regime in which the method is applied. In Supplementary Note 3, n0 = Delta_l / d is the width of the point-spread function in object-pixel units. For a fixed physical object, refining the pixel grid below the Rayleigh limit makes the total pixel count M grow while n0 grows proportionally to M. The Methods pseudo-code sizes the window border from the 'number and relative value of major bands of the FIM,' so each local minimization at a given shift involves O(n0) unknown and known pixels, and the number of shifts is only O(M/n0). The total cost is therefore O(M) only under the unstated assumption that n0 remains O(1) as the grid is refined. Since the experiments use a fixed 32x32 detector array, they do not exercise the scaling regime in which the problem is claimed to become hard. Please state the assumptions under which the claimed linearity holds, or qualify the claim in the abstract and introduction.","section":"Abstract; Results (Theoretical background); Methods (Sliding windows method)"},{"comment":"The truncation approximation used in both algorithms is stated as 'the theoretical probabilities are computed assuming all pixels but those inside the window to be zero' (first approximation) and 'pixels outside the full window are set to zero' (refinement), but no error bound is given for this approximation. The size of the omitted-pixel contribution is controlled by the same PSF tails that define n0; when d is reduced toward the super-resolution regime, n0 grows and the number of omitted pixels with non-negligible weight grows as well, so the bias introduced by the truncation is not controlled. A quantitative error bound, or at least a numerical convergence check as a function of the window and border sizes, is needed to justify the accuracy of the SWM reconstructions in the claimed regime.","section":"Methods (pseudo-code, first approximation and refinement)"}],"minor_comments":[{"comment":"The sentence stating that the coefficients are 'effectively zero for |j-m| << n0 or |k-n| << n0' appears to have the inequality reversed: if n0 is the PSF width in object pixels, the coefficients should be negligible for separations much larger than n0, not much smaller. Please correct this.","section":"Supplementary Note 3, paragraph after Eq. (22)"},{"comment":"The text refers to 'the Appendix C' for the spatial correlation function, but the manuscript contains no Appendix C; the relevant description appears in Supplementary Note 3. Please update the cross-reference.","section":"Supplementary Note 4"},{"comment":"The claim of resolution beyond the Rayleigh limit is supported visually by the red bars in Fig. 3, but a quantitative definition of achieved resolution (for example, estimated feature widths with uncertainties derived from the FIM or from repeated reconstructions) would make the claim more precise and easier to evaluate.","section":"Results (Experiment); Fig. 3"},{"comment":"The statement 'The code itself is available upon request' is weaker than the rest of the reproducibility effort; depositing the code in a permanent repository would allow readers to reproduce the reconstructions and the linear-complexity experiments.","section":"Data availability"},{"comment":"The definition of strict parametric l-locality is given for a one-dimensional ordering of parameters, while the imaging demonstrations are two-dimensional. Please state how the locality condition and the window construction generalize to two dimensions (for example, using a Chebyshev or Euclidean distance between pixel indices).","section":"Results (Theoretical background)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid experimental core, and the optimal-correlation-width prediction is an interesting and experimentally supported result. The main risk is that the complexity claim in the abstract and introduction is stronger than the analysis supports. If the authors can prove the linearity claim under explicit assumptions, or alternatively reframe the contribution around parametric locality and the experimental demonstrations, the paper would be suitable for publication. I do not see grounds for rejection, but the scaling and truncation-error issues are load-bearing and require a careful revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading for two things: the sliding-window method (SWM) for parametrically local nonlinear estimation, and the prediction that there is an optimal source correlation width for super-resolved imaging. The experimental confirmation of the optimal-width prediction, using both SPDC and pseudo-thermal sources, is genuine and not something I have seen before. The SWM itself is a clean idea: use the banded structure of the Fisher information matrix to design local windows, estimate only a subset of parameters per step, and iterate. The pseudo-code is understandable, the model derivations in the supplement are standard, and the reconstructions of the USAF chart and the digit are plausible evidence that the method works in practice.\n\nThe soft spots are in the complexity claim and in the truncation approximation. The abstract and the text say the scheme is linear on the total number of parameters. That is only true if the window border size stays fixed as the parameter grid refines. In the near-field imaging model, the PSF width in pixel units is n0 = Δl/d, so when you shrink the pixel size for super-resolution, the number of relevant FIM bands grows. The window border must cover those bands, so per-window cost grows and the number of shifts does not fall as fast as the parameter count grows. The total complexity is not strictly linear. The paper does not provide an error bound for the step that sets pixels outside the window to zero, and the experimental demonstrations use a 32×32 detector, so they do not stress-test the scaling regime. These are real limitations, but they are not fatal: for moderate super-resolution factors the method works, and the optimal-width result is independent of the complexity claim.\n\nThe code is only available on request, which is a reproducibility concern, though the data are deposited. The reliance on rule-of-thumb window sizing is a minor issue; the FIM gives a way to pick the border, but the paper does not formalize when the approximation is guaranteed to be good.\n\nWho is this for? People working on quantum imaging, few-photon correlation measurements, and nonlinear multiparameter estimation. It deserves a serious referee: the information-theoretic approach and the experimental confirmation of the optimal correlation width are contributions that should be vetted properly. I would send it to review, and I would ask the authors to either prove the complexity claim under grid refinement or soften it, and to provide public code. I would cite the optimal-width prediction in my own work.","headline":"A solid, useful paper whose main algorithmic claim is real but whose linear-complexity statement is oversold in the super-resolution regime.","tokens_in":19190,"tokens_out":1740,"would_cite":true,"duration_ms":21774,"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":"Parametric locality, visible as band structure in the Fisher information matrix, enables a sliding-window estimator that reconstructs super-resolved quantum images at linear cost and predicts the optimal photon correlation width.","keywords":["quantum imaging","Fisher information matrix","sliding window method","parametric locality","super-resolution","higher-order correlation functions","pseudo-thermal light","entangled twin photons"],"falsifier":"Calculate the number of significant Fisher information bands, n0 = Δl/d, for a one-dimensional near-field image on grids with d = Δl/2, Δl/4, and Δl/8. If the required border size grows with n0 rather than remaining bounded, the per-window cost grows with the number of pixels and the linear-complexity claim is refuted; the paper reports no benchmark of this scaling.","tokens_in":18191,"feed_emoji":"🔬","tokens_out":8652,"duration_ms":89483,"temperature":0.7,"pith_summary":"Parametric locality is the property that each measurement outcome depends on only a few neighbouring parameters of the object. The paper claims that for such problems the Fisher information matrix is banded and its inverse is approximately banded, so a parameter can be estimated from a small window of data around it. On that basis it constructs an iterative sliding-window method in which each step fits only a subset of parameters, making the total reconstruction effort linear in the number of parameters instead of nonlinear. Applied to quantum near-field imaging from second- and third-order correlation functions, the method reconstructs grey transmission objects beyond the Rayleigh limit using both pseudo-thermal and entangled twin-photon sources. The same Fisher-information analysis predicts an optimal photon correlation width, close to the smallest object detail to be resolved, and the paper reports experimental confirmation of that prediction.","feed_headline":"Sliding windows make quantum super-resolution imaging linear-cost","feed_subtitle":"Banded Fisher information lets each fit touch only nearby pixels, beating the Rayleigh limit in experiments.","key_machinery":"The central object is the Fisher information matrix of a parametrically l-local measurement, where each outcome probability has nonzero derivatives only with respect to parameters within distance l, so the matrix is l-banded. The argument uses the matrix fact that inverses of banded matrices are approximately banded, which turns the error bound for one parameter into a function of nearby parameters only. The sliding window method is the estimator built from that fact: after a coarse first pass with pixels large enough for diagonal dominance, it refines the grid and, in each step, fits only a core window of unknown pixels surrounded by a border whose width exceeds the number of major Fisher information bands; values outside the window are held fixed or set to zero. The border width, not the total number of pixels, sets the cost of every local fit, and shifting the core window across the object produces the claimed linear total complexity.","core_discovery":"The paper's central discovery is that the hard nonlinear estimation problem of quantum imaging becomes tractable when the measurement is parametrically local: when each detected correlation event depends on a small cluster of nearby pixels, the Fisher information matrix is narrowly banded, and the standard lower bound on estimator variance for a given pixel depends only on data in a window around it. The authors build the sliding window method on this: first a coarse reconstruction with large pixels whose Fisher information matrix is diagonally dominant, then refinement on a finer pixel grid where each optimization step treats a core of unknown pixels, a border of known or ignored pixels, and a border size set by the number of major Fisher information bands. They demonstrate the method on simulated and experimental data, reconstructing grey transmission objects from measured second- and third-order correlation functions for pseudo-thermal light and for position-momentum entangled twin photons, and report resolution beyond the Rayleigh limit. The paper also establishes two further results from the same information analysis: an optimal correlation width of the imaging field, approximately the size of the smallest object detail, and an improvement of resolution from the estimation bias that arises when parameters sit at the boundary of their allowed range.","pith_inferences":["The same windowing principle should carry over to any nonlinear estimation problem whose Fisher information matrix is banded, not only optical imaging; the argument never uses the specific imaging model except to establish locality.","A direct stress test is to benchmark how the required border size grows as the pixel grid is refined below the Rayleigh limit; the linear-complexity claim assumes this growth stays bounded while the number of parameters increases.","The reported resolution gain from boundary bias suggests that encoding object constraints, such as known binary or grey-value bounds, is not just a prior but a measurable information resource worth quantifying separately from the correlation statistics.","One could test the optimal-correlation-width rule by measuring reconstruction infidelity versus source correlation width across several feature sizes and checking whether the minimum tracks the feature size, as the paper reports for one object."],"forward_implications":["For objects with many pixels, each reconstruction step becomes a small least-squares fit over a window, so the total computational cost scales linearly with the number of parameters instead of with a high power of it.","In the super-resolution regime, the best photon-source correlation width is not the smallest possible but roughly the smallest object feature size, giving a design rule for choosing source speckle size or twin-photon correlation width.","Higher-order correlation measurements with both pseudo-thermal and entangled twin-photon sources can reconstruct grey transmission objects beyond the Rayleigh limit.","When estimated parameters lie at the boundary of their allowed range, the resulting estimate bias can lower the total reconstruction error, so binary objects can be resolved better than the unbiased error bound suggests.","The Fisher-information analysis can be performed per imaging setup without knowing the object, so the window structure and border size can be fixed in advance from the point-spread function and source correlations."],"supporting_citations":[{"why":"Supplies the matrix result that inverses of banded matrices are approximately banded, the step that makes local estimation possible.","marker":"[21, 22]"},{"why":"Defines the near-field imaging model and its point-spread function connecting object-plane and image-plane fields.","marker":"[14]"},{"why":"Establishes the pseudo-thermal speckle source and its Gaussian correlation statistics used in the experiments and simulations.","marker":"[32, 33]"},{"why":"Provides the theoretical description of the position-momentum entangled two-photon state used for the SPDC imaging.","marker":"[38]"},{"why":"Gives the variance bounds for biased estimators, used to show that boundary-induced bias can improve resolution.","marker":"[34, 35]"},{"why":"Describes the single-photon detector array that supplied the experimental correlation data.","marker":"[36, 37]"},{"why":"Defines the Rayleigh limit used as the resolution benchmark for the super-resolution claims.","marker":"[23]"},{"why":"Provides the bounds showing that diagonal entries of the inverse matrix are controlled by nearby matrix entries, fixing the needed window border size.","marker":"[39–41]"}],"fun_headline_variants":["Quantum imaging gets linear-time reconstruction","Banded Fisher information speeds quantum super-resolution","Windowed quantum inference breaks Rayleigh limit","Higher-order photon correlations simplify quantum imaging","Twin-photon super-resolution via sliding-window fitting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the measurement is parametrically local at the finest pixel grid used: the width of the Fisher information bands stays small relative to the total number of parameters as the grid is refined below the Rayleigh limit, so a fixed-size window border can keep the local fits accurate.","fun_headline_variants_meta":{"raw":{"variants":["Quantum imaging gets linear-time reconstruction","Banded Fisher information speeds quantum super-resolution","Windowed quantum inference breaks Rayleigh limit","Higher-order photon correlations simplify quantum imaging","Twin-photon super-resolution via sliding-window fitting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000843,"raw_usage":{"total_tokens":3699,"prompt_tokens":999,"completion_tokens":2700,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":2635}},"tokens_in":615,"tokens_out":2700,"duration_ms":20510,"temperature":1.0,"reasoning_tokens":2635,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:18:17.314434+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Calculate the number of significant Fisher information bands, n0 = Δl/d, for a one-dimensional near-field image on grids with d = Δl/2, Δl/4, and Δl/8. If the required border size grows with n0 rather than remaining bounded, the per-window cost grows with the number of pixels and the linear-complexity claim is refuted; the paper reports no benchmark of this scaling.","supporting_citations":[{"cited_title":"Shih, IEEE Journal of Selected Topics in Quantum Electronics, IEEE 13, 1016 (2007)","cited_arxiv_id":null,"evidence_quote":"Defines the near-field imaging model and its point-spread function connecting object-plane and image-plane fields."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the theoretical description of the position-momentum entangled two-photon state used for the SPDC imaging."},{"cited_title":"Born and E","cited_arxiv_id":null,"evidence_quote":"Defines the Rayleigh limit used as the resolution benchmark for the super-resolution claims."}],"review_version":1}