{"id":"97158214-78de-4269-a787-71a66892853f","arxiv_id":"2411.17566","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A new background estimation algorithm uses local pixel minima and empirical corrections to recover sky background within about 10-14% in simulated low-photon, crowded UV images.","lead":"Astronomers must subtract sky background to see faint objects, but standard tools fail in low-light and crowded images. This paper proposes using the darkest pixels, local minima, plus statistical corrections to estimate background, and tests it on simulated UV images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 10–14% recovery claim is calibrated to one smoothing/box configuration; without a mapping for F(G1,G2,b) or an estimate of the source-contamination term, it does not generalize to other PSFs or pixel scales.","rationale":"The paper's analytic minima statistics for IID Gaussian fields are derived cleanly, and the source-free Poisson calibration is internally consistent. The stress-test concern is not that the method fails for the tested configuration, but that the headline quantitative claim is calibrated to one instrument-specific configuration and one smoothing schedule. The reader's weakest_assumption already identifies the empirical F and 0.1-offset constants; my check sharpens this by asking whether the claim survives a change of pixel scale and PSF, which is exactly the scope implied by 'astronomical images' in the abstract. The Discussion and Appendix A honestly acknowledge the F-dependence and the unestimated source-contamination term Sc, so the reader's CONDITIONAL verdict is warranted. If the proposed recalibration shows the constants are stable and the error distributions remain within the claimed bounds, the claim would be strengthened; if not, the paper should be explicitly scoped to the exact calibration and UVIT-like pixel scale used in Table 1.","tokens_in":20419,"tokens_out":10026,"duration_ms":98696,"concrete_test":"Recompute the §4.2 calibration on source-free Poisson fields for a different pixel scale/PSF (e.g., 0.6\"/px, PSF FWHM≈2.5 px), choosing G1/G2 and b by the paper's own '5–10× PSF' rule, then run the §4.4 UVIT-like validation (500 and 4000 sources; λ=0.348 and 5.8) for 100 realizations using the published F=0.213 and 0.1-offset. If the 95th percentile of relative deviation exceeds 10%/14%, or if the best-fit F differs from 0.213 by more than ~20%, the abstract's accuracy claims are calibration-specific and need to be scoped accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's quantitative claims—10% recovery in low-photon, uncrowded fields and 14% in crowded fields—rest on Eqs. (22)–(24) correctly removing the minima bias from source-contaminated images. These relations are calibrated only for a fixed configuration: G1 FWHM=15.3 px, G2 FWHM=3.5 px, box b=7 px, on UVIT-like 0.417\"/px simulated fields (§4.2, Figs. 6 and 12). Appendix A explicitly states that F(G1,G2,b) varies with these parameters and provides only qualitative trends (Figs. 11–12), not a closed-form mapping. A user who follows the paper's '5–10× PSF' prescription for another instrument will therefore get a different F, and the claimed accuracy has no guarantee. Separately, the source-contamination term Sc in Eq. (25) is listed but never estimated; the Discussion concedes it 'cannot be estimated independently of the background.' The Table 1 headline entries are single realizations with no repeated-trial statistics or real-data ground truth. Thus the 10–14% figures are a property of one calibration, not of the geometric minima method as presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a geometric method for estimating astronomical background levels by identifying local minima via steepest descent and exploiting the statistics of these minima to recover the parent distribution's mean. For IID Gaussian fields the authors derive an analytic recovery relation μ = μ_min + 2.496 σ_min (Eq. 16) and verify it numerically. For Poisson noise they add an empirical power-law correction (Eq. 18). For crowded and low-count fields they propose a multi-resolution recipe: locate minima on a heavily smoothed image, track them through progressively narrower kernels, and sample boxes around the final minima on the original image; the sampled mean is corrected by an empirical relation (Eq. 24). The method is tested on simulated Poisson fields and UVIT-like images with injected Sérsic sources, reporting recovery within 10% (uncrowded, λ=0.348) and 14% (crowded) of the input background, while SExtractor's background shows much larger deviations. A background-map extension and several limitations are discussed.","tokens_in":112,"tokens_out":14326,"duration_ms":228971,"significance":"The central idea of using 2D spatially selected minima to avoid source contamination is a genuine methodological contribution, and the analytic Gaussian result (K≈2.496) provides a clean, numerically checked asymptotic baseline. The contamination-fraction comparison (Fig. 8) convincingly demonstrates that the minima sampler collects 1–2 orders of magnitude less source flux than SExtractor's 3σ-clipping over the tested range, which is the paper's most promising result. The paper honestly lists important limitations (λ≲0.1 failure, crowded-field degradation, diffuse-object caveats). However, the headline 10% and 14% accuracies are currently calibrated for a single smoothing/box configuration and a specific simulated source model, and a sign error in Eq. (15) affects the printed Gaussian derivation. With these issues addressed, the method could be a useful tool for low-count UV surveys.","major_comments":[{"comment":"The printed formula for σ_min contains a sign error. For the minimum of N=9 IID Gaussians, the variance is σ_min^2 = 18 σ^2 ∫_0^1 [Erf^{-1}(2Φ−1)]^2 (1−Φ)^8 dΦ − (μ−μ_min)^2, whereas Eq. (15) as written puts a minus sign before the 18, making the right-hand side negative for a non-negative integrand and producing an imaginary σ_min. This is load-bearing for the central recovery constant K=C/D: the reported numerical values C_gaussian=1.48501 and D_gaussian=0.59779 cannot follow from the printed formula. Please correct the sign and confirm the derivation.","section":"Section 2.2, Eq. (15)"},{"comment":"The quantitative recovery relations used for the source-free and crowded-field tests are μ−μ_min = (0.213±0.002)σ (Eq. 22) and λ_recovery = μ_min + 0.213(σ_min + 0.1) (Eq. 24), with the constant 0.213 fitted for one configuration: G1 FWHM=15.3 px, G2 FWHM=3.5 px, b=7 px. Appendix A shows only qualitative dependencies of F on G1, G2, and b (Figs. 11–12) and gives no closed-form mapping or tabulated values. Consequently the abstract's 'within 10%' and 'within 14%' claims are properties of this single calibration, not of the geometric minima method as presented. A user applying the stated '5–10× PSF' prescription at a different pixel scale or PSF cannot compute F and has no accuracy guarantee. Please provide at least a fitted formula or lookup table for F(G1,G2,b) over the tested ranges, or explicitly restrict the claimed accuracy to the tested configuration.","section":"Appendix A and Eq. (22)"},{"comment":"The background recovery equation includes a source contamination term Sc, but the paper states that Sc 'cannot be estimated independently of the background.' The implementation therefore does not correct for Sc; it only attempts to minimize it by sampling near minima. The claimed 10–14% accuracies in Table 1 are for a specific simulated source model (Gaussian PSF, SNR 0.7–7, one source-count value for each crowding level) with no independent estimate of Sc. Since the Discussion acknowledges that the source distribution and PSF 'could have a significant impact' on background estimation, the paper should either provide a method to estimate or bound Sc, or revise the claims to state they are validation results for the simulation setup, not universal accuracy guarantees.","section":"Section 4.3, Eq. (25)"},{"comment":"The text states the Minima routine recovers B_MS = 1.03×10^-4 cps/pixel for the 500-source, texp=3000 s image (figure 10a), while Table 1 lists 1.05×10^-4; the corresponding deviations are ~11% and 9.32% from the input 1.16×10^-4, so the abstract's 'within 10%' claim is sensitive to this discrepancy. Additionally, each entry in Table 1 appears to be a single realization (or a single averaged value) with no repeated-trial statistics, uncertainties, or scatter; the reader cannot assess whether the 10–14% figures are robust. Please reconcile the printed values and report mean ± scatter over multiple realizations.","section":"Section 4.4 and Table 1"}],"minor_comments":[{"comment":"The integral contains a stray '(Φ)' after Erf^{-1}(2Φ−1); it should read ∫_0^1 Erf^{-1}(2Φ−1) (1−Φ)^8 dΦ.","section":"Section 2.2, Eq. (14)"},{"comment":"The text 'images containing 500, 2000 and 4000 thousand sources' should read '4000 sources'; similarly, the phrase in Section 3 '300 to 3600 (69 to 828 sources/arcminutes²)' should clarify the units of source density.","section":"Section 4.3, after Figure 7"},{"comment":"The offset 0.1 is not dimensionless; the paper notes it becomes 0.1/texp for cps units, but the units should be stated explicitly in the equation itself.","section":"Section 4.2, Eq. (23)"},{"comment":"The claims of 'within 10%' for low-photon-count images should be accompanied by the λ≳0.1 applicability restriction stated in Section 5, since the paper itself says the method is not confident below λ~0.1.","section":"Abstract and Section 5"},{"comment":"The bibliography lists Saha et al. 2024a and 2024b with the same arXiv identifier (arXiv:2408.03629); please verify the citations.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is well within the scope of an astronomical methods journal. The principal concerns are: (i) the sign error in Eq. (15), which should be easy to fix but is load-bearing; (ii) the calibration-specificity of the F coefficient; and (iii) the unqualified quantitative claims that go beyond the single calibration. If the authors can provide a closed-form mapping or a table for F(G1,G2,b) and report repeated-trial statistics, the paper could be acceptable after revision. I would recommend not accepting in the present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Pandey and Saha have a genuinely useful idea. Instead of trying to clip sources away, sample the image at local minima, where source contamination is least. The Gaussian order-statistics derivation is clean, the constant K=2.496 is analytic, and the Poisson corrections are empirically fitted but candidly presented as such. On the simulated UVIT-like images, the method beats SExtractor by large margins in low-λ and crowded regimes. That is a real contribution to a practical problem.\n\nThe stress-test note is right, and it matters. The 10–14% recovery figures are properties of one calibration — G1 FWHM=15.3 px, G2 FWHM=3.5 px, box b=7 px — not of the method as presented. Appendix A shows F(G1,G2,b) varies with these parameters, but gives no closed-form mapping. A user following the '5–10× PSF' prescription for another instrument will get an unknown F, and the claimed accuracy has no guarantee. The source-contamination term Sc is listed in Eq. 25 but never estimated; the paper concedes it 'cannot be estimated independently of the background.' That leaves a real gap for images with specific source distributions. Table 1 also shows single realizations with no repeated-trial statistics, and there is no real-data validation beyond a mention of applying it to AUDF GOODS South.\n\nNone of this is fatal to the core idea. The math is sound, the improvements over SExtractor are plausible and demonstrated in simulation, and the failure modes are spelled out honestly. What is missing is a generalization argument: a mapping or sensitivity study for F over a range of PSFs, pixel scales, and box sizes, plus some bracketing of Sc. I would send this to peer review with those requests, because the method deserves to be tested properly and the claims need to be reined in to the calibrated regime.\n\nFor your reading group: worth a look, but go in knowing the headline numbers are narrower than the abstract suggests. I would cite it as a promising technique, not as a settled calibration.","headline":"Solid minima-sampling background estimator, but the headline 10-14% accuracy is calibrated to one smoothing configuration and needs a mapping for F before it generalizes.","tokens_in":21197,"tokens_out":2695,"would_cite":true,"duration_ms":23494,"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":"A geometric minima estimator recovers astronomical background within 10-14% where 3-sigma clipping fails.","keywords":["background estimation","minima statistics","steepest descent","Poisson noise","low-photon count","crowded fields","faint source detection","sky background"],"falsifier":"Generate a Poisson background image with known $\\lambda$ from 0.03 to 100, with sources injected as in the paper, but run the recipe with a different final kernel and box size (for example $G_2$ FWHM = 6 pixels and $b = 15$ pixels); if the recovered background leaves the paper's 10-14% band, the claimed accuracy is not a general property of the method.","tokens_in":20234,"feed_emoji":"🔭","tokens_out":9930,"duration_ms":77652,"temperature":0.7,"pith_summary":"The paper proposes a geometric background estimator for astronomical images: instead of collapsing pixels into a histogram and iteratively clipping outliers, it finds local minima by steepest descent and samples pixels around those minima. The paper argues these minima-dominated samples are far less contaminated by faint sources, so the recovered mean background stays close to the true sky level. In low-photon-count images (mean count $\\lambda \\approx 0.35$) the method recovers the background within about 10% in sparse fields and 14% in crowded fields, where conventional 3-$\\sigma$ clipping underestimates the background by orders of magnitude in sparse fields and overestimates it by roughly 200% in crowded fields. An accurate background matters because an overestimated background suppresses exactly the faint sources that deep surveys target.","feed_headline":"Minima sampling recovers sky background within 10 percent","feed_subtitle":"Searching for local minima avoids the faint-source contamination that biases 3-sigma sky estimates.","key_machinery":"The load-bearing object is the minima distribution of a random field and the identity $P_{\\rm min}(x) = N P(x)(1-\\Phi(x))^n$, which gives the isomorphic CDF relation $\\Phi = 1 - (1 - \\Phi_{\\rm min})^{1/(n+1)}$ and the linear moment-shift law $\\mu = \\mu_{\\rm min} + K\\sigma_{\\rm min}$. The operative mechanism in real images is the iterative smoothed-minima search: wide-to-narrow Gaussian smoothing with steepest descent, followed by fixed box sampling at the final minima, which targets source-free regions without collapsing the 2D pixel distribution into a 1D histogram. The quantitative claim is carried by empirically fitted coefficients: $K \\approx 2.496$, the Poisson correction $A \\approx 0.231$ and $\\gamma \\approx -0.2085$, and, for the smoothed-field recipe, the mean-shift coefficient $F = 0.213$ with the $\\sigma$ offset $\\sqrt{\\lambda} = \\sigma_{\\rm min} + 0.1$.","core_discovery":"The central discovery is a statistical relation between a random field and its local minima: for an $M$-dimensional iid field, the minima PDF is $P_{\\rm min}(x) = N P(x)(1-\\Phi(x))^n$, and the minima CDF is isomorphic to the parent CDF, so the parent mean can be recovered as $\\mu = \\mu_{\\rm min} + K\\sigma_{\\rm min}$, with $K \\approx 2.496$ for Gaussian fields and a Poisson correction $\\lambda \\approx \\mu_{\\rm min} + 2.496\\sigma_{\\rm min} + A\\mu_{\\rm min}^{\\gamma}$ ($A \\approx 0.231$, $\\gamma \\approx -0.2085$). For images with sources, the paper replaces direct minima sampling with an iterative recipe: smooth with a wide Gaussian kernel (full width at half maximum 15.3 pixels), find minima by steepest descent, then re-find minima through progressively narrower kernels down to 3.5 pixels, and sample $7 \\times 7$ pixel boxes around the final minima. On simulated UVIT-like images this recovers the input background within about 10% at $\\lambda = 0.348$ in an uncrowded field, within 14% at $\\lambda = 0.348$ in a crowded field, within about 10% at $\\lambda = 5.8$ in a crowded field, and within 0.65% at $\\lambda = 5.8$ uncrowded.","pith_inferences":["Because the minima CDF is isomorphic to the parent CDF for any continuous parent distribution, the same moment-correction logic should apply to non-Poisson noise mixtures such as read noise and dark current, provided the coefficients are recalibrated; the paper calibrates only Gaussian and Poisson cases.","The fitted coefficients $F$, $A$, and $\\gamma$ are measurements rather than derivations, so a closed-form mapping from kernel widths and box size to $F$ would let the method be used without per-instrument recalibration; the paper's Appendix leaves that mapping open.","The residual deviation near $\\lambda \\approx 0.1$ is likely set by integer-count zeros; an exposure-time-aware version of the offset in $\\sqrt{\\lambda} = \\sigma_{\\rm min} + 0.1$ would be a direct testable extension for very short exposures.","The paper's reported accuracy concerns recovering the mean background; survey decisions depend at least as much on patch-to-patch variance and on correlated noise from artifacts, which the paper lists as an untested limitation."],"forward_implications":["Low-photon UV images with $\\lambda \\approx 0.35$ and sparse sources get a background estimate within about 10%, whereas histogram-clipping routines return values orders of magnitude too low.","Crowded fields near 500 sources per square arcminute get recovered background within about 14% at low $\\lambda$ and about 10% at higher $\\lambda$, instead of the roughly 200% and 60% overestimates of clipping-based estimators.","Source contamination in the sampled pixels stays one to two orders of magnitude below that of 3-sigma clipping, improving the signal-to-noise of faint sources near the detection limit.","The method extends to spatially varying backgrounds by patching the field and interpolating, producing background and rms maps for a 236 square arcminute field in about 30 seconds.","The paper proposes the minima routine for wide deep surveys where faint-source density is high, naming Euclid, DESI imaging, and LSST as suitable applications."],"supporting_citations":[{"why":"Supplies the standard 3-sigma-clipping background estimator that the paper uses as the baseline comparison for recovery accuracy.","marker":"Bertin & Arnouts (1996)"},{"why":"Supplies the maxima-segmentation algorithm used to mask source regions before minima sampling in the first approach.","marker":"Berry (2015)"},{"why":"Provides the Sersic modeling tools and software environment used to generate the simulated UVIT-like source fields.","marker":"Astropy Collaboration et al. (2022)"},{"why":"Supplies the UVIT GOODS South NUV catalog and the measured background level that seed the realistic simulations, and is the survey application where background maps are produced.","marker":"Saha et al. (2024b)"},{"why":"Defines the UVIT N242w pixel scale and PSF used for injecting point sources into the synthetic images.","marker":"Ghosh et al. (2022)"}],"fun_headline_variants":["Minima statistics recover sky background to 10% accuracy","Steepest descent on minima beats 3-sigma sky estimates","Geometric background estimator dodges faint-source contamination","Local minima PDF yields accurate sky background in UV images","New method estimates sky background within 10% in low-photon fields"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 10-14% accuracy claims rest on empirically fitted constants ($A \\approx 0.231$, $\\gamma \\approx -0.2085$, the 0.1 $\\sigma$ offset, and especially $F = 0.213$) that are calibrated to one fixed smoothing sequence and box size, with no closed-form mapping for other settings.","fun_headline_variants_meta":{"raw":{"variants":["Minima statistics recover sky background to 10% accuracy","Steepest descent on minima beats 3-sigma sky estimates","Geometric background estimator dodges faint-source contamination","Local minima PDF yields accurate sky background in UV images","New method estimates sky background within 10% in low-photon fields"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000755,"raw_usage":{"total_tokens":3446,"prompt_tokens":1122,"completion_tokens":2324,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":738,"completion_tokens_details":{"reasoning_tokens":2255}},"tokens_in":738,"tokens_out":2324,"duration_ms":15769,"temperature":1.0,"reasoning_tokens":2255,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:58:23.169169+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a Poisson background image with known $\\lambda$ from 0.03 to 100, with sources injected as in the paper, but run the recipe with a different final kernel and box size (for example $G_2$ FWHM = 6 pixels and $b = 15$ pixels); if the recovered background leaves the paper's 10-14% band, the claimed accuracy is not a general property of the method.","supporting_citations":[],"review_version":1}