{"id":"e3cf0594-86ce-4ebc-a9ac-aa04a36a8305","arxiv_id":"2412.06905","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Dazzle reconstructs an over-sampled reference from undersampled Roman dithered images and uses difference imaging with matched filtering and PSF fitting to detect and measure microlensing events.","lead":"Dazzle builds over-sampled reference images from dithered Roman Space Telescope frames and subtracts them to find faint changing stars. It provides open-source tools for detecting microlensing events in crowded bulge fields and measuring their brightness changes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pure-translation assumption at the heart of the over-sampled reference construction (Eq. 2) is violated by Roman's geometric distortion, so the simulated recovery rates may not transfer to real data without a distortion correction that also alters the PSF.","rationale":"The reader's weakest assumption is that the frames are pure translations with no differential rotation and a known over-sampled PSF. I agree this is the most load-bearing premise. The paper's entire reference construction rests on Eq. 2, which models each pixel as a direct sample of a single global over-sampled image at a fixed offset. For Roman, geometric distortion is a well-known, large effect that couples neighboring pixels and makes the per-pixel independent fit invalid. The paper acknowledges only that there is 'no differential rotation,' but does not address the position-dependent distortion that is characteristic of WFI. Without a distortion-correction step, the reference will contain field-dependent artifacts, and difference images will inherit correlated noise that suppresses faint-transient detection. The offset-correction sign inconsistency and the underestimated error bars are real issues, but they are secondary to the geometric assumption because the headline recovery rates and photometric light curves come from simulations that already satisfy the pure-translation idealization. I therefore keep the reader's CONDITIONAL verdict unchanged: the method is plausible and the simulations are internally consistent, but the transfer to real Roman data is not demonstrated and the geometric assumption is the primary risk.","tokens_in":9513,"tokens_out":5766,"duration_ms":63704,"concrete_test":"Generate a set of dithered synthetic images with RomanISIM including the true WFI distortion model, feed them to Dazzle without any distortion-correction preprocessing, and measure (a) the RMS of fixed-star residuals in the difference images and (b) the microlensing recovery rate for W146 < 23. Compare with the same pipeline run on pure-translation images. If either metric degrades by more than about 10%, the pure-translation assumption is load-bearing for the headline claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on Eq. 2: each pixel of every dithered image is a direct sample of one global over-sampled image at a fixed translation offset. This is only true if the only difference between frames is a linear shift. Roman's WFI has significant geometric distortion, so real dithered images are not pure translations of a fixed scene. The mapping between sky coordinates and detector coordinates varies nonlinearly across the 4k×4k field, meaning the per-pixel independent Legendre fit in Eqs. 3–8 will absorb a field-dependent distortion pattern into the reference. Difference images will then contain correlated residuals that scale with local intensity gradients and the relative distortion between frames, exactly the regime where faint microlensing signals are lost. The paper does not describe any distortion-correction or resampling step. If images were resampled to a distortion-free grid, the effective PSF would change, breaking the 'known over-sampled effective PSF' assumption used in the photometry. Thus, the 80–90% recovery rates in §3.5 are demonstrated only under an idealized geometry that Roman does not provide.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents Dazzle, an open-source Python package for constructing oversampled reference images from dithered, undersampled Roman WFI frames and for performing difference-image detection and photometry. The method represents each reference pixel as a low-order Legendre polynomial in the subpixel dither coordinate, solves for the coefficients using a design matrix assembled from the dithered images, iteratively masks outlier pixels, forms difference images, detects microlensing events with a three-dimensional Gaussian matched filter, and fits PSF photometry to the difference images. The validation is entirely simulation-based: SYNTHPOP and RomanISIM images of the Galactic bulge with injected PSPL microlensing events, recovery rates stated as above 90% for W146 < 23 and above 80% for W146 < 25, illustrative light curves, and a photometric scatter diagram. The paper is clearly written and the public code is a strength.","tokens_in":9748,"tokens_out":11401,"duration_ms":124509,"significance":"If the central claims hold, Dazzle would be a practical component of a Roman GBTDS transient pipeline, especially for faint and blended sources where direct PSF fitting is expected to struggle. The paper's strengths are its analytic, interpolation-free reference model, its injection-recovery validation against an external benchmark (the injected PSPL light curves), and the fact that the implementation and reproduction scripts are publicly available. However, two equation-level errors and the idealized translation-only geometry assumed in Eq. (2) currently limit the strength of the Roman-readiness claim; these issues are fixable but require new or corrected analysis.","major_comments":[{"comment":"There is an internal sign inconsistency in the offset-correction derivation. Equation (9) defines the difference image as D = T - R, but Eq. (10) writes the corrected difference image as D' = R(x+beta, y+gamma) - T, which is -D before the shift terms. Expanding Eq. (10) gives D' = -D + beta dR/dx + gamma dR/dy, yet Eq. (12) states D' = D + beta dR/dx + gamma dR/dy. If D is kept as T - R, the corrected residual is D - beta dR/dx - gamma dR/dy, so the minimisation in Eq. (13) has the wrong sign for the shift terms. As written, the normal equations (14)-(16) will return offset corrections of the wrong sign. This section should be rewritten with a single sign convention and tested on simulated frames with deliberately misregistered dither offsets.","section":"Sec. 3.4, Eqs. (9)-(16)"},{"comment":"Equation (18) is not the variance of the flux estimate in Eq. (17). For constant pixel noise sigma, Eq. (17) reduces to F = sum(P D)/sum(P^2), whose variance is sigma^2 / sum(P^2); Eq. (18) instead evaluates to sigma^2, independent of the PSF normalization. The correct variance of the weighted least-squares estimate is (sum P^2/sigma^2)^(-1). Because the paper uses these error bars to judge photometric scatter and then states that the Poisson error bars underestimate the scatter, this formula needs to be corrected and the comparison in Fig. 7 reassessed. If the PSF grid is normalized in a special way, that normalization must be stated explicitly.","section":"Sec. 3.6, Eqs. (17)-(18), and Fig. 7"},{"comment":"The reference construction assumes a single global translation (integer plus subpixel) between each frame and the reference grid, with no differential rotation (Sec. 3.1). Roman's WFI has strong geometric distortion over its 4k x 4k field, so real dithered frames are not pure translations of a fixed scene. A field-dependent mapping violates Eq. (2), and the per-pixel Legendre fit can absorb slowly varying distortion into the reference, leaving correlated residuals that scale with local intensity gradients. The simulations do not appear to include or correct for this effect, and no test with differential rotation is reported. To support the claim that Dazzle is ready for Roman GBTDS data, the recovery rates in Sec. 3.5 should be recomputed on simulated images with a realistic distortion model, or with images resampled to a distortion-free frame and the ePSF appropriately transformed. At a minimum, the quoted recovery rates should be explicitly stated as applying to the idealized translation-only case.","section":"Secs. 3.1-3.5, Eq. (2)"}],"minor_comments":[{"comment":"The text contains duplicated words: 'the of the input stars' near the Fig. 2 discussion and 'by by' in Sec. 3.6; these should be corrected.","section":"Secs. 2 and 3.6"},{"comment":"The recovery statistic is defined only as 'within 3 pixels' and the false-positive rate is described as 'a few' without a number. Reporting completeness and false-positive curves as a function of magnitude, u0, spatial kernel, temporal kernel, and detection threshold would make the detection claim quantitative and reproducible.","section":"Sec. 3.5"},{"comment":"The choices N = 5 and f = 1.2 are justified only as working well 'from experimentation'; a short sensitivity test showing reference residuals or recovery rates versus N and f would make this tuning reproducible.","section":"Sec. 3.2"},{"comment":"In Eqs. (15)-(16), the vector b should be written as b_k to make clear that the normal equations are solved separately for each image k, while the matrix A is common to all frames.","section":"Sec. 3.4"},{"comment":"The statement that the Poisson error bars underestimate the scatter should be quantified, for example by plotting the ratio of observed scatter to formal error versus magnitude, rather than left as a qualitative remark.","section":"Sec. 3.6"},{"comment":"The introduction suggests difference imaging will be superior to direct PSF fitting for faint, blended sources, but no side-by-side comparison on the same simulated frames is shown; such a comparison would substantiate the motivation.","section":"Sec. 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for an instrumentation/data-analysis journal. I agree with the reader that there is no circularity: the injected PSPL events are an external benchmark, and the code is genuinely public. The two equation-level issues in Sec. 3.4 and Sec. 3.6 are real but local, and the distortion question is a load-bearing robustness gap rather than an internal inconsistency. These are fixable within the paper's scope, so I recommend major revision rather than rejection. One editorial point: the abstract's 'high recovery rates' should either be accompanied by the idealized-geometry caveat or by a distortion-inclusive simulation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a serious, useful methods paper, not a breakthrough. It takes known ingredients—per-pixel Legendre-basis oversampled references, iterative outlier masking, 3D matched filtering, PSF fitting on difference images—and assembles them into a coherent, working pipeline for Roman's undersampled WFI data. The code is open-source, the validation uses injected synthetic microlensing events with known truth, and the paper explicitly flags that its Poisson error bars understate scatter. That last point is refreshing; too many pipeline papers hide that kind of thing.\n\nThe genuinely new bit is the combination for Roman: the oversampled reference built from randomly dithered undersampled frames, with an analytic correction for small dither errors, and the matched-filter detection stack. The recovery rates (90% for W146<23, >80% for W146<25) are plausible for the simulated geometry.\n\nThe soft spots, in order of importance. First, the core assumption is that dithers are pure translations of a fixed scene with a known oversampled PSF. Roman's WFI has real geometric distortion, so real dithered images are not pure translations. The paper does not discuss how distortion would be handled, and resampling to fix it changes the effective PSF. So the recovery rates are demonstrated under an idealized geometry and may not transfer directly. This is not a fatal flaw for a methods paper, but it needs to be addressed in the discussion, and validation on more realistic simulated distortion is needed. Second, the offset-correction algebra in Section 3.4 is internally inconsistent: Eq. 9 defines D = T - R, but Eq. 10/12 effectively use R - T. Likely a typo in the write-up, but as printed the sign story doesn't add up. Third, the tuning parameters (N=5, f=1.2, detection thresholds) are chosen on the same simulated data, so the recovery rates are somewhat optimistic. That is a limitation the authors acknowledge indirectly, but it would be good to see a cross-validation or at least a sensitivity test.\n\nBottom line: this paper deserves a serious referee. It is a solid, reproducible contribution to a real need—difference imaging for Roman's bulge survey. I'd ask the authors to fix the sign inconsistency, add a distortion discussion, and ideally show how the pipeline behaves with modest distortion injected. The Poisson error-bar caveat is fine for now. I would cite this work and bring it to the attention of anyone building Roman time-domain pipelines.","headline":"Dazzle is a well-built, openly coded difference-imaging pipeline for Roman's undersampled WFI images; the simulations are honest, but the pure-translation assumption and a sign error in the offset-correction algebra need attention before relying on the recovery rates.","tokens_in":10271,"tokens_out":6004,"would_cite":true,"duration_ms":55673,"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":"The paper presents Dazzle, a pipeline for Roman that builds an oversampled reference from dithered undersampled images and claims high recovery of microlensing events in the resulting difference images.","keywords":["difference imaging","oversampled image reconstruction","Roman Space Telescope","microlensing detection","matched filtering","PSF photometry","crowded stellar fields","dithering"],"falsifier":"Apply Dazzle to a transient-free set of real Roman GBTDS frames and compare difference-image RMS with the photon-noise floor; if the lower envelope of scatter exceeds the paper's $\\log_{10}\\sigma[\\mathrm{ppt}] = -5.19 + 0.29\\, W146$ relation, or if injected events are recovered below the stated rates, the pure-translation and known-PSF assumptions are the failing point.","tokens_in":9275,"feed_emoji":"🔭","tokens_out":8419,"duration_ms":79996,"temperature":0.7,"pith_summary":"This paper aims to give the Roman Space Telescope's Galactic Bulge Time Domain Survey a difference-imaging route for transients, especially microlensing, that works when the individual images are undersampled and dithered rather than oversampled by seeing. The central move is to treat every pixel of every dithered frame as a direct sample of one global oversampled scene, solve for that scene pixel-by-pixel in a Legendre basis, and subtract it from each frame to make difference images. The author reports that on 192 simulated Roman images this recovers 90% or more of injected microlensing events with W146 < 23 and better than 80% for W146 < 25, and produces PSF-fitted light curves with a photometric scatter lower envelope of $\\log_{10}\\sigma[\\mathrm{ppt}] = -5.19 + 0.29\\, W146$. If these results hold on real data, Dazzle would be a practical complement to direct PSF photometry for faint and blended bulge sources.","feed_headline":"Dazzle pipeline recovers 90% of bright microlensing events","feed_subtitle":"A dither-built oversampled reference turns Roman's undersampled frames into difference images for faint bulge stars.","key_machinery":"The load-bearing object is the per-pixel analytic oversampled reference: each reference-grid pixel $(i,j)$ is represented as $R_{ij}(x,y)=\\sum_{l,m}\\theta_{ijlm}B_l(x-x_i)B_m(y-y_j)$, where the $B$'s are Legendre polynomials stretched to overlap neighbouring pixels. The coefficients $\\theta_{ijlm}$ are obtained from the weighted linear least-squares solution $\\theta=(X^T C^{-1} X)^{-1} X^T C^{-1} z$, with $X$ built from the sub-pixel dither offsets and $z$ from the integer-shifted pixel values. This makes the reconstruction local, parallelizable, and free of image interpolation, and it is what turns the difference-image problem into a per-pixel linear algebra problem rather than a kernel-fitting problem. Detection then rides on a three-dimensional Gaussian matched filter over the aligned difference stack, with trial temporal widths to capture different event durations.","core_discovery":"The paper's claim is that a high-precision difference image can be formed for Roman without fitting a convolution kernel at all. Because the PSF is stable in time, the only difference between frames is a dither, so each measured pixel $T_{k,ij}$ equals the unknown oversampled scene $R$ sampled at a known sub-pixel phase: $T_{k,ij}=R(x_i+\\Delta x_k+\\delta x_k,y_j+\\Delta y_k+\\delta y_k)+\\epsilon$. Representing $R$ inside each reference pixel as a product of shifted Legendre polynomials reduces the reconstruction to an independent linear least-squares problem per pixel, with a design matrix shared by all pixels. Difference images are then direct subtraction $D_{ij}=T_{k,ij}-R_{ij}$, with outlier pixels masked iteratively. The paper further derives a first-order analytic correction for small dither offset errors and demonstrates a three-dimensional matched filter over the difference-image stack that detects injected microlensing events at the stated recovery rates, followed by PSF-fitting photometry on the difference images with Nelder-Mead sub-pixel coordinate refinement.","pith_inferences":["Beyond the paper's tests, the per-pixel Legendre construction should also work for other undersampled space telescopes with a stable PSF, provided the pure-translation and known-PSF assumptions hold; this is an extrapolation, not a paper claim.","The matched-filter detection, demonstrated for microlensing with a Gaussian temporal kernel, should transfer to any transient with a known time profile by swapping the kernel, making supernova or flare searches a natural next target.","The paper notes that Poisson error bars underestimate the observed scatter, which suggests that correlated noise from the reconstructed reference enters the difference images; modelling that covariance could sharpen both detection and photometric uncertainties.","Because the design matrix depends only on the dither pattern, the reference construction can be precomputed and parallelised per pixel, so the pipeline may scale to the full survey area if geometric distortion corrections stay within the linear regime."],"forward_implications":["For Roman's Galactic Bulge Time Domain Survey, faint and blended microlensing sources that are poorly served by direct PSF photometry can still be detected and measured through difference images.","Difference-image photometry reaches a bright-end floor near 0.03 mag scatter per point, indicating where direct photometry should take over.","The analytic dither-offset correction offers a way to refine small astrometric misalignments without re-solving the reference from scratch.","The three-dimensional matched filter can be tuned by trial temporal kernels to trade detection rate against false positives for different event timescales."],"supporting_citations":[{"why":"Defines the effective PSF and the direct PSF-fitting paradigm that the oversampled reference builds on.","marker":"Anderson & King (2000)"},{"why":"Establishes the ground-based convolution-kernel difference imaging that this paper replaces for undersampled space data.","marker":"Alard & Lupton (1998)"},{"why":"Provides the precedent of a per-pixel basis-function representation for HST time-series images.","marker":"Gilliland et al. (2000)"},{"why":"Shows that accurate sub-pixel coordinates are required for precise PSF photometry, motivating the Nelder-Mead refinement.","marker":"Albrow et al. (2009)"},{"why":"Supplies the survey parameters for the Roman Galactic Bulge Time Domain Survey used in the simulations.","marker":"Penny et al. (2019)"},{"why":"Provides the noise simulations to which the measured photometric scatter is compared.","marker":"Wilson et al. (2023)"},{"why":"WebbPSF is used to generate the PSF grids for both the simulated images and the photometry.","marker":"Perrin et al. (2014)"},{"why":"The optimization method used for sub-pixel coordinate refinement in the PSF fitting.","marker":"Nelder & Mead (1965)"},{"why":"The Besançon galactic stellar population model provides the input catalogue for the simulated bulge fields.","marker":"Robin et al. (2003)"}],"fun_headline_variants":["Dazzle: oversampled references enable Roman difference imaging","No-kernel difference images for Roman's crowded bulge fields","Dazzle: recover faint transients with dither-based subtraction","Sub-pixel Roman photometry via oversampled scene reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the dithered frames are pure translations of a fixed scene with a known over-sampled PSF and no differential rotation; if real Roman frames carry residual geometric distortion, rotation, or PSF errors, the reference and every difference image inherit correlated errors before any transient search begins.","fun_headline_variants_meta":{"raw":{"variants":["Dazzle: oversampled references enable Roman difference imaging","No-kernel difference images for Roman's crowded bulge fields","Dazzle: recover faint transients with dither-based subtraction","Sub-pixel Roman photometry via oversampled scene reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1413,"prompt_tokens":931,"completion_tokens":482,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":411}},"tokens_in":547,"tokens_out":482,"duration_ms":5357,"temperature":1.0,"reasoning_tokens":411,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:20:23.311881+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply Dazzle to a transient-free set of real Roman GBTDS frames and compare difference-image RMS with the photon-noise floor; if the lower envelope of scatter exceeds the paper's $\\log_{10}\\sigma[\\mathrm{ppt}] = -5.19 + 0.29\\, W146$ relation, or if injected events are recovered below the stated rates, the pure-translation and known-PSF assumptions are the failing point.","supporting_citations":[],"review_version":1}