{"id":"6c10e88b-c547-459d-a01e-adeb3814d781","arxiv_id":"2501.08732","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Stacking up-the-ramp frames with weights optimized for SNR=1 faint pixels improves SNR over equal-weight and ramp-fitting and yields about 0.5 mag deeper CSST near-infrared limits.","lead":"This paper proposes a weighted stacking scheme for up-the-ramp infrared detector readouts, choosing weights that are optimal for faint pixels with signal-to-noise ratio of one. For CSST near-infrared images, stacking 30 readout frames would improve limiting magnitude by about 0.5 mag and cut effective readout noise by about 62%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SNR and CSST gain claims rest entirely on a white-noise covariance model; correlated readout noise could erode them, and the paper tests no real detector ramps.","rationale":"I read the paper as making two claims: (1) a mathematical claim that C^{-1}s weights maximize SNR for a single pixel under the stated covariance model, and (2) an empirical claim that choosing the SNR_target=1 case and applying one weight set to all pixels beats equal-weight and ramp fitting in simulations, and would improve CSST limiting magnitudes by about 0.5 mag. The mathematical core is standard, and I verified Eqs. (2)-(6); the target-mismatch robustness argument in Fig. 2 is plausible and consistent with the reported flat-field/point-source ordering. I did not find an internal inconsistency in the white-noise simulations. The load-bearing weakness is the transfer of those results to real detectors: the covariance model in Eq. (2) is the only place detector physics enters, and it assumes temporally white, spatially uncorrelated readout noise with no persistence or cosmic-ray jumps. Real NIR ramps violate this, and the paper's own citation [6] shows correlated readout noise affects flux measurements in Euclid NISP. Since the CSST 0.5 mag and 62% numbers are predicted improvements for a real instrument, an unmodeled off-diagonal readout covariance could shrink or reverse them. This is a correctness risk for the practical conclusion, not an ad hominem or a consensus dispute. It also supports the reader's CONDITIONAL verdict: the method is well-founded under its assumptions but not yet validated on real ramps. I would not change the reader's verdict; the condition should be to re-test with correlated noise or real detector ramps before adopting the method in a CSST pipeline. The absence of code or data artifacts independently limits verification, but that is secondary to the noise-model concern.","tokens_in":9619,"tokens_out":15103,"duration_ms":161122,"concrete_test":"Re-run the Section 3 flat-field and point-source pipelines and the Section 4 CSST calculation with the readout-noise generator replaced by a correlated process, e.g., a covariance C_r(i,j) = sigma^2 rho^{|i-j|} with rho = 0.5 and 0.9, plus a 1/f frame-frequency component, while keeping rms readout noise fixed at 50 e-. Recompute the SNR curves in Fig. 3 and the N=30 limiting-magnitude row of Table 4. If quasi-optimal stacking no longer dominates equal-weight/ramp fitting in any regime, or if the 0.5 mag gain shrinks by more than about 0.1 mag, the claims need to be restated as conditional on white readout noise; if the gains persist, the concern is retired.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's quantitative claims—quasi-optimal stacking always beating equal-weight and ramp fitting, and the 0.5 mag / 62% CSST improvement—are derived from and tested against the covariance model C_ij = s_min(i,j) + b_min(i,j) + r^2 delta_ij (Section 2). This model treats readout noise as independent, identically distributed Gaussian draws. Real NIR arrays, including HgCdTe devices like CSST's, exhibit correlated readout noise: 1/f and common-mode drifts, row/column correlations, reset anomalies, and persistence. With off-diagonal readout covariance, C^{-1}s is no longer the true optimal weight vector, and the performance gap against equal-weight and ramp-fitting methods has no reason to be preserved. The paper itself cites the Euclid correlated-readout-noise study (Ref. [6]) but none enters the simulations or the CSST estimate. Because Table 4 and the 'always' claim are the headline results, this untested structural assumption is the single most load-bearing weakness. The internal mathematics under the stated model appears correct; the risk is that the stated model is not the detector's noise model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a weighted stacking method for up-the-ramp (non-destructive readout) images. The authors derive the optimal weights ω ∝ C^{-1}s for a pixel whose covariance C follows from Poisson signal/background fluctuations plus frame-independent Gaussian readout noise (Eqs. (2)-(6)). Because a single weight set must serve all pixels of different brightness, they select the weights that maximize SNR for a target case of SNR=1 in the last frame, calling the result quasi-optimal. They compare this method with equal-weight stacking, ramp fitting, and simply taking the last frame, using simulated flat-field images and point-source images (Section 3). The paper reports that the quasi-optimal method yields the highest SNR in the tested regimes (with the last frame eventually winning in the high-SNR photon-noise-dominated regime), and it estimates that for CSST NIR observations, stacking 30 up-the-ramp frames improves the limiting magnitude by about 0.5 mag and reduces effective readout noise by about 62% (Section 4, Table 4).","tokens_in":9832,"tokens_out":6446,"duration_ms":65171,"significance":"The analytic derivation in Section 2 is clean and correct under the stated covariance model, and the method has practical appeal: a single set of frame-wise weights preserves easy flux calibration, and the SNR_target=1 choice is well motivated for faint extended or point sources. The simulations are clearly described and support the relative ranking of the methods within the tested parameter space. The CSST forecast, if validated against realistic detector noise, is actionable for the instrument's observing strategy. However, the quantitative claims—both the 'always better' simulation result and the 0.5 mag/62% CSST improvement—rest entirely on an idealized noise model with uncorrelated readout noise and pure Poisson fluctuations, and the paper itself cites but does not incorporate the known correlated-readout-noise behavior of near-IR arrays. The practical significance is therefore real but not yet fully established.","major_comments":[{"comment":"The covariance model C_ij = s_min(i,j) + b_min(i,j) + r^2 δ_ij assumes pure Poisson signal/background and readout noise that is independent across frames with identical variance. This is the single load-bearing assumption behind the claim that the quasi-optimal method outperforms equal-weight and ramp fitting, and behind the CSST limiting-magnitude and effective-readout-noise numbers. Real HgCdTe NIR arrays, including those planned for CSST, exhibit correlated readout noise (1/f, common-mode, row/column correlations), and the paper cites the Euclid NISP study of this effect (Ref [6]) without letting it enter the simulations or the CSST estimate. As a result, the reader cannot tell whether the 0.5 mag and 62% improvements survive realistic detector behavior. I request either additional simulations with a correlated-readout-noise component (for example, a simple 1/f or common-mode model) or a quantitative literature-based argument bounding the degradation; alternatively, the paper should explicitly frame the conclusions as conditional on the white-readout-noise model.","section":"Section 2, Eq. (2); Section 4, Table 4"},{"comment":"The sentence defining the flat-field SNR reads: 'The SNR is given by the ratio of the standard deviation of all pixels in the reduced image to the mean of these pixels.' This is the inverse of the SNR definition in Eq. (2) and of the values actually plotted in Figure 3 (which must be mean/std to reach SNR_last≈1 for the Table 2 parameters). Please correct this typo and write the formula explicitly so that the simulation metric is unambiguous.","section":"Section 3.1"}],"minor_comments":[{"comment":"The abstract says the quasi-optimal method 'always enhances the SNR more than the equal-weight stacking method and the ramp fitting method.' Although this is not formally contradicted by the conclusion (which concedes that the last-frame method wins at high SNR in the photon-noise-dominated regime), the word 'always' is stronger than what the finite, idealized simulations support. Please qualify it as 'in the simulated regimes under the stated noise model' to match the scope of the evidence.","section":"Abstract and Section 5"},{"comment":"The symbol b in Eq. (13) is used without definition; from context it is the background count per pixel accumulated by the last frame, but this should be stated explicitly to make the effective-readout-noise formula reproducible.","section":"Section 4, Eq. (13)"},{"comment":"The limiting magnitudes and effective readout noise values in Table 4 are quoted to two decimals with no uncertainties or sensitivity analysis. Given the assumed system throughput, sky background, and readout noise are all stated with inequalities or handbook values, a simple parameter variation (e.g., ±10% in throughput and sky background) would help the reader assess the robustness of the 0.5 mag and 62% headlines.","section":"Section 4, Table 4"},{"comment":"Section 3.2 reports point-source SNR boosts of '57–70%' for four specific input flux levels, but the figure labels also list the resulting aperture SNRs. Adding the noise realizations or the number of trials used for those aperture photometry measurements would improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central derivation is sound and the paper is likely to be a useful contribution to ramp-image reduction practice. My main concern is the gap between the idealized white-readout-noise covariance model and the unqualified quantitative claims made for real detectors, particularly the CSST numbers. If the authors add a robustness analysis with correlated readout noise (or clearly bound its impact), I would be inclined to accept. The 'always' wording in the abstract should also be toned down to match the simulation scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the paper's actual new idea is the choice to derive stacking weights for a per-pixel SNR_target=1 and to show that this choice is robust over a wide range of target SNRs. That is a practical and useful extension of the textbook result w ∝ C^{-1}s, and the CSST projection is a legitimate forward calculation. The math in Section 2 is clean; I verified the covariance structure and the weight derivation. The simulations cover the three relevant noise regimes, and the point-source test with SEP photometry supports the ranking. The paper is also honest in its conclusion about the high-SNR exception, which earns credit.\n\nWhat I would want fixed before quoting the headline numbers:\n\n- The abstract says the method \"always\" beats equal-weight stacking and ramp fitting. The paper's own conclusion says the last-frame method wins for SNR_last ≳ 130 in the photon-noise-dominated regime. That is not a contradiction of the literal claim, but it shows the abstract's \"always\" is doing more work than the paper's actual conclusion. A reader should not have to read to Section 5 to find the exception.\n\n- The whole quantitative structure rests on the white-noise covariance model C_ij = s_min(i,j) + b_min(i,j) + r^2 δ_ij. The paper cites the Euclid correlated-readout-noise study but does not test correlated noise, 1/f, persistence, or cosmic-ray jumps. Real NIR ramps have these, and each one changes C and can shrink the claimed gains. The 0.5 mag / 62% CSST numbers should be labeled as model-dependent estimates until real detector ramps are analyzed. This is the load-bearing soft spot.\n\n- No error bars on the simulation curves, and no code or data artifacts, so reproducibility is limited. The point-source boost of 57–70% is from one parameter set.\n\nWho this is for: infrared detector data-reduction folks, especially anyone setting up CSST or Roman NIR pipelines. The qualitative recommendation—take at least 30 up-the-ramp frames and stack with C^{-1}s weights tuned to a low-SNR target—is likely to survive even if the exact magnitude gain shifts. I would send this to a serious referee; the derivation is sound, the simulations are helpful, and the missing correlated-noise testing is a fixable weakness rather than a fatal one.","headline":"A practical, well-derived stacking recipe for faint-source detection in up-the-ramp readouts, with a solid robustness study and a CSST estimate that needs a caveat about correlated readout noise before its headline numbers are used.","tokens_in":10367,"tokens_out":3296,"would_cite":true,"duration_ms":35160,"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":"For non-destructive readout images, weighting frames by the inverse noise covariance with a target of SNR=1 per pixel outperforms equal-weight stacking and ramp fitting, and for CSST NIR data it would add about 0.5 magnitude of depth.","keywords":["up-the-ramp readout","non-destructive readout","image stacking","signal-to-noise ratio","ramp fitting","infrared detectors","China Space Station Telescope","astronomical data reduction"],"falsifier":"Take a real near-infrared detector ramp sequence with a known flat background and measure the covariance matrix $C_{ij}$ from many pixels. If the off-diagonal entries deviate from $\\min(s_i+b_i,s_j+b_j)$ beyond the readout noise, or if the diagonal variances are not equal across frames, the derived weights are no longer optimal; a direct test would compare the SNR of real ramp stacks reduced with the paper's weights against the last-frame and equal-weight methods and check the predicted gain of roughly 60% in SNR.","tokens_in":1689,"feed_emoji":"🔭","tokens_out":4111,"duration_ms":77791,"temperature":0.7,"pith_summary":"The paper proposes a way to combine up-the-ramp detector frames—successive non-destructive reads of the same pixels during one exposure—into a single image that is nearly optimal for finding faint objects. Because noise in these frames is correlated, the best linear combination uses unequal weights, and the paper derives the weights that maximize signal-to-noise ratio for a chosen target brightness. The authors then argue that targeting the faintest useful case, one signal-to-noise ratio unit per pixel in the last frame, is a safe choice: bright objects lose little, faint objects gain the most. They claim this quasi-optimal stacking always beats equal-weight stacking and ramp fitting in their simulated flat-field and point-source tests. For a 150-second CSST near-infrared exposure, stacking 30 frames would deepen the limiting magnitude by about 0.5 mag and cut effective readout noise by about 62%.","feed_headline":"Stacking 30 ramp frames adds 0.5 mag of depth for CSST","feed_subtitle":"A single weight set tuned to SNR=1 maximizes faint-source signal while barely costing bright objects, simulations show.","key_machinery":"The load-bearing object is the covariance matrix $C_{ij}=s_{\\min(i,j)}+b_{\\min(i,j)}+r^2\\delta_{ij}$, which encodes the fact that Poisson fluctuations persist once accumulated, while readout noise appears only on the diagonal. Maximizing $\\mathrm{SNR}=\\mathbf{w}^T\\mathbf{s}/\\sqrt{\\mathbf{w}^T C\\mathbf{w}}$ gives the optimal weight vector $\\mathbf{w}_{\\mathrm{opt}}\\propto C^{-1}\\mathbf{s}$; the paper's method fixes the target SNR per pixel in the last frame to unity, computes these weights once, and applies the same weights to every pixel in a frame to preserve calibratability. The key insight is an asymmetry: using brighter-than-truth target weights hurts faint objects, while using fainter-than-truth target weights barely hurts bright objects, so targeting the faintest recoverable case is quasi-optimal.","core_discovery":"The central discovery is that a single global set of stacking weights, derived from the noise covariance under the target assumption $\\mathrm{SNR_{last}}=1$, achieves signal-to-noise ratio per pixel at or near the theoretical optimum across a wide range of source brightness, whereas weights tuned to brighter targets degrade faint sources severely. The paper expresses the optimal weights as $\\mathbf{w}_{\\mathrm{opt}}\\propto C^{-1}\\mathbf{s}$, where $C$ is the covariance of pixel values across frames and $\\mathbf{s}$ is the signal vector, and shows numerically and with simulations that the SNR loss is under 1% for $\\mathrm{SNR_{last}}$ up to 10 when targeting $\\mathrm{SNR_{target}}=1$. In flat-field and point-source simulations spanning readout-, background-, and photon-noise dominated regimes, the method outperforms equal-weight stacking, ramp fitting, and the last-frame benchmark, except that at very high SNR in the photon-noise dominated regime the last frame eventually wins. For the CSST NIR imager, the method converts 30 up-the-ramp frames into roughly 0.5 mag of additional limiting depth.","pith_inferences":["The derivation is independent of the telescope, so the same weights should transfer to other non-destructive-readout cameras, provided the noise is dominated by uncorrelated Poisson and readout terms.","If correlated readout noise or 1/f noise is measured, the covariance matrix $C$ could be augmented with off-diagonal terms beyond the diagonal readout term, and the same $C^{-1}\\mathbf{s}$ recipe would produce weights for those detectors.","A practical extension for cosmic-ray rejection would be to flag and clean jumps before weighting, since the weights assume no jumps; the paper does not model this.","The CSST 0.5 mag estimate assumes a perfectly Poisson background and Gaussian readout, so it would need end-to-end verification on real detector calibration data before being used in an observing plan."],"forward_implications":["For any up-the-ramp data set with the assumed noise covariance, choosing $\\mathrm{SNR_{target}}=1$ yields stacked SNR within a few percent of the per-pixel optimum up to high $\\mathrm{SNR_{last}}$, so one weight set serves an entire survey.","In readout-noise-dominated observations, the method recovers most of the SNR that destructive-readout observers lose, improving faint-source detection without extra exposure time.","The method avoids the instability of ramp fitting when only a few up-the-ramp exposures are available.","For CSST NIR with 30 frames, limiting magnitude improves by about 0.5 mag and effective readout noise falls by about 62%, with modest additional gain to about 0.6 mag and 71% at 70 frames.","At very high SNR in the photon-noise dominated regime the last-frame method can surpass it, so the gain is concentrated exactly at the detection threshold where it matters."],"supporting_citations":[{"why":"Provides the analytic ramp-fitting noise formulas used in the SNR comparisons.","marker":"[23]"},{"why":"Defines the equal-weight stacking benchmark and supplies the sky-background spectrum used for the CSST estimate.","marker":"[20]"},{"why":"Describes the multiband photometer reduction that motivates ramp fitting, the method the paper compares against.","marker":"[21]"},{"why":"Introduces the slope-fitting approach for infrared ramps that is the ramp-fitting method under test.","marker":"[22]"},{"why":"Provides the aperture-photometry implementation used to measure point-source SNR in the simulated star images.","marker":"[25]"},{"why":"Underlies the source-extraction library used for the photometry measurements.","marker":"[26]"}],"fun_headline_variants":["One weight set gives near-optimal up-the-ramp stacking, 0.5 mag deeper for CSST","SNR=1 tuning yields quasi-optimal stacking, adds 0.5 mag for CSST","Faint up-the-ramp sources get near-optimal SNR from one weight set","Stacking up-the-ramp frames: near-optimal SNR with single weights","Up-the-ramp stacking method: 0.5 mag deeper for CSST"],"cache_read_input_tokens":12544,"weakest_assumption_plain":"The whole derivation rests on the assumption that each pixel's noise is exactly the sum of two independent Poisson processes (source and background) plus Gaussian readout noise with the same variance in every frame and no correlation between frames; real detectors often have correlated readout noise, 1/f noise, persistence, or cosmic-ray jumps that would make the derived weights suboptimal and could shrink the claimed gains.","fun_headline_variants_meta":{"raw":{"variants":["One weight set gives near-optimal up-the-ramp stacking, 0.5 mag deeper for CSST","SNR=1 tuning yields quasi-optimal stacking, adds 0.5 mag for CSST","Faint up-the-ramp sources get near-optimal SNR from one weight set","Stacking up-the-ramp frames: near-optimal SNR with single weights","Up-the-ramp stacking method: 0.5 mag deeper for CSST"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000753,"raw_usage":{"total_tokens":3416,"prompt_tokens":1080,"completion_tokens":2336,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":696,"completion_tokens_details":{"reasoning_tokens":2217}},"tokens_in":696,"tokens_out":2336,"duration_ms":16502,"temperature":1.0,"reasoning_tokens":2217,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:19:09.375545+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real near-infrared detector ramp sequence with a known flat background and measure the covariance matrix $C_{ij}$ from many pixels. If the off-diagonal entries deviate from $\\min(s_i+b_i,s_j+b_j)$ beyond the readout noise, or if the diagonal variances are not equal across frames, the derived weights are no longer optimal; a direct test would compare the SNR of real ramp stacks reduced with the paper's weights against the last-frame and equal-weight methods and check the predicted gain of roughly 60% in SNR.","supporting_citations":[{"cited_title":"2007, Analysis of the sampling schemes for WFC3-IR, WFC3 ISR, 12","cited_arxiv_id":null,"evidence_quote":"Provides the analytic ramp-fitting noise formulas used in the SNR comparisons."},{"cited_title":"2023, in WFC3 Instrument Handbook for Cycle 31 v","cited_arxiv_id":null,"evidence_quote":"Defines the equal-weight stacking benchmark and supplies the sky-background spectrum used for the CSST estimate."},{"cited_title":"D., Rieke, G","cited_arxiv_id":null,"evidence_quote":"Describes the multiband photometer reduction that motivates ramp fitting, the method the paper compares against."},{"cited_title":"D., & Forrest, W","cited_arxiv_id":null,"evidence_quote":"Introduces the slope-fitting approach for infrared ramps that is the ramp-fitting method under test."},{"cited_title":"2016, SEP: Source Extractor as a library, Journal of Open Source Software, 1 (6), 58","cited_arxiv_id":null,"evidence_quote":"Provides the aperture-photometry implementation used to measure point-source SNR in the simulated star images."},{"cited_title":"1996, SExtractor: Software for source extraction, Astronomy and astrophysics 10 ASTRONOMICAL TECHNIQUES & INSTRUMENTS supplement series, 117 (2), 393–404","cited_arxiv_id":null,"evidence_quote":"Underlies the source-extraction library used for the photometry measurements."}],"review_version":1}