{"id":"9fed453f-9d2f-4e30-a948-fd42bb0f235a","arxiv_id":"2608.00354","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Readout glow, not only 1/f drift, explains why averaging many non-destructive reads gives less noise reduction than 1/sqrt(N), and it sets a floor sigma_min ~ 1.5 sigma_RN^{1/2} G^{1/4}.","lead":"This paper shows that “readout glow”—light emitted by an infrared detector's own electronics during readout—adds a small extra electron count on every read, so averaging more reads to reduce noise stops helping after a point. It derives a noise-floor formula and tests it on two detector types, giving a practical limit for future space telescopes and faint-object spectrographs.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Noise-floor formula and glow-not-1/f attribution rest on per-read glow being independent Poisson; LmAPD validation fits that assumption rather than testing it, leaving the central claim not fully established.","rationale":"The reader's weakest assumption matches my own: the independence and Poisson variance of per-read glow is the load-bearing condition. The paper's derivation (Appendix B) is internally consistent under that assumption, and the simulation validation confirms the algebra, but the laboratory evidence does not independently establish the assumption. The LmAPD fits use the same model, so they are consistency checks, not tests of independence; the direct glow measurement only fixes the mean, not the variance or correlation structure. The H2RG data explicitly show glow is not limiting in that instrument, so the central validation rests entirely on LmAPD, where a 1/f or correlated-noise alternative could plausibly produce a similar noise floor. The proposed covariance test would settle this by directly examining whether the residual read-to-read covariance follows the assumed min(i,j) structure. If it does, the concern is resolved and the paper's conditional acceptance stands; if it does not, the fundamental-limit claim is not supported. I agree with the reader's conditional verdict and see no reason to move it, hence UNCHANGED.","tokens_in":15681,"tokens_out":5217,"duration_ms":55751,"concrete_test":"Using the 100 LmAPD ramps at 10 V, subtract the best-fit linear ramp (signal + dark + glow) from each read, then compute the empirical residual covariance matrix C_ij across reads (after 5-sigma outlier rejection). Compare C_ij to the model prediction C_model(i,j) = G·min(i,j) + σ_RN^2 δ_ij, with G and σ_RN estimated from the direct measurements. If the off-diagonal residuals are consistent with zero within uncertainties, the independent-Poisson assumption is supported. If significant excess covariance grows with |i−j|, or the per-read residual power spectrum exhibits a 1/f-like signature, the noise-floor formula and the 'glow, not 1/f' conclusion require revision. This directly tests the assumption that the entire derivation and validation rest on.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that readout glow sets a fundamental noise floor with σ_min ∝ σ_RN^{1/2} G^{1/4} (Eqs. 16, 22) and that glow, not 1/f drift, explains observed averaging limits—depends on the assumption that per-read glow g_k is an independent Poisson random variable. Appendix B makes this explicit: Eq. 39 uses 'independence of the g_k' to evaluate the glow variance, and Eqs. 44–46 use independence to compute the Fowler group covariance. The same assumption underpins the SUTR covariance substitution in Eq. 20. If glow values are correlated across reads (e.g., a persistent 1/f component in the source follower) or have variance exceeding the Poisson mean (overdispersion), the covariance structure changes, the optimal-N scaling is no longer σ_RN^{1/2} G^{1/4}, and the attribution to glow over 1/f fails. The LmAPD validation does not test this: the solid curves in Fig. 6 are two-parameter fits to the very model that assumes independence, and the forward-model wedge sweeps cF_g over 0.3–1.0, so it is permissive rather than decisive. The H2RG data are background-limited and measured noise exceeds the theory, so they provide no independent confirmation. Thus the strongest claim is load-bearing on an assumption that is analytically convenient but empirically unchecked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes that readout glow — photoelectrons generated by the detector readout electronics during each nondestructive read — sets a fundamental limit on how far read noise can be reduced by averaging. The authors model glow as an independent Poisson random variable added on each read, derive noise-optimal averaging formulas for Fowler sampling (Eq. 16) and sample-up-the-ramp (Eq. 22), giving a floor σ_min ∝ σ_RN^{1/2} G^{1/4}, and extend the model to linear-mode avalanche photodiodes by introducing partial-gain factors c and F_g (Section 4.2, Eqs. 27–28). The analytic results are checked against numerical simulations (Appendix A). Laboratory data are presented for a Magellan/FIRE H2RG detector and an Ike Pono LmAPD; the authors conclude that glow, rather than 1/f drift, explains the observed inability to average down noise in the LmAPD, with agreement to 0.1 e- over two decades of averaging.","tokens_in":16043,"tokens_out":3746,"duration_ms":34430,"significance":"If the central claim holds, this is a practically important result for infrared detector development, with direct implications for ELT instruments and the Habitable Worlds Observatory. The analytic derivations are transparent, the simulations in Appendix A strongly support the variance calculations, and the paper usefully identifies the weak G^{1/4} dependence of the noise floor. However, the experimental validation is the load-bearing part of the claim that glow—not 1/f drift—limits averaging, and that validation is currently incomplete: the close-looking agreement in Fig. 6 is largely a two-parameter fit, and the forward prediction is a broad band. The manuscript also never empirically tests the independence/Poisson assumption on which the scaling and the glow-not-1/f attribution rest.","major_comments":[{"comment":"The abstract's claim that the LmAPD 'follows the predicted noise value to within 0.1 e-' is not supported by the figure as presented. The solid curves in Fig. 6 are two-parameter fits of the amplified noise model (σ_RN/M and F_g c^2 G free) to the very same noise-vs-N curves, so they cannot be called a prediction. The only forward-prediction element is the shaded wedge, built from the directly measured mean glow cG and sweeping cF_g over 0.3–1.0. That wedge is permissive; many curves, including the 1/f-drift hypothesis, could pass through such a band. Please separate 'fit' from 'prediction' in the text and quantify the agreement using the forward wedge, not the fitted curves.","section":"§6.2.2, Fig. 6"},{"comment":"The entire derivation of the noise floor and the attribution of the averaging limit to glow rather than 1/f drift depends on the assumption that per-read glow values g_k are independent Poisson random variables with variance equal to their mean G. Eq. (39) explicitly invokes independence of the g_k, and Eq. (44) does the same for the covariance. The LmAPD validation fits a model that assumes this independence; it does not test it. The H2RG data (Sec. 6.1) are background-limited and the measured noise exceeds the theory, so they do not discriminate either. If glow has temporal correlations (e.g., a persistent component from the source follower) or is overdispersed, the covariance structure changes, the σ_RN^{1/2} G^{1/4} scaling no longer follows, and the conclusion that glow, not 1/f drift, limits averaging would not be established. Please add a direct empirical check: estimate per-read","section":"Appendix B, Eqs. (39), (44); §4.2"},{"comment":"The internal consistency of the LmAPD glow measurements is not as good as the text suggests. At 6 V, the fitted glow variance F_g c^2 G is 20.6×10^-3 e-/pix/read while the directly measured mean glow cG is 9.82×10^-3 e-/pix/read, giving a ratio of about 2.1, which lies well outside the paper's stated range cF_g = 0.3–1.0. At 8 V the ratio is about 1.2, also above unity. The text says the two measurements are 'consistent within the uncertainties but persistently offset'; for 6 V that seems inconsistent. This discrepancy is directly relevant to the quantitative validation claim and should be addressed, ideally with an explicit error budget for the ratio and a discussion of what would make the 6 V point compatible with the model.","section":"§6.2.2, Table 2"},{"comment":"The H2RG test does not validate the glow-floor prediction: the measured noise lies above the theory and the paper attributes the difference to structured noise and dark current, with dark current exceeding glow. As the text itself says, 'glow is not limiting in this particular instrument.' Including this data set in the abstract's claim of 'laboratory data using both sensor architectures' is misleading. The H2RG data can be retained as a glow detection and a qualitative check, but the manuscript should not present it as confirmation of the noise-floor model.","section":"§6.1.2, Fig. 5"}],"minor_comments":[{"comment":"The 'Independent per read' designation for glow in Table 1 is an assumption, not an established property. Please mark it as an assumption and point forward to the empirical test suggested above.","section":"§2, Table 1"},{"comment":"The caption reads 'See text in this section for an explanation,' but no explanation of Figure 1 appears in Section 5. Either add a proper description or refer to the specific paragraph that interprets the plot.","section":"Figure 1"},{"comment":"The approximations leading to N_opt and σ_min are not stated. For a reader trying to reproduce the coefficients 1.7 and 1.5, it would help to show the intermediate step (e.g., minimizing 2σ_RN^2/N + 2NG/3).","section":"§3.2, Eq. (16)"},{"comment":"The paper states 'We focus on 10 V in what follows' but then Table 2 includes 6 and 8 V with no further interpretation. Either discuss those rows or move them to an appendix.","section":"§6.2.2"},{"comment":"There are several minor typos: inconsistent capitalization of 'Fowler' (e.g., 'fowler' in Appendix B), and the use of 'f texp' versus 'f·t_exp' in equations. A careful proofread would improve readability.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"This is a potentially important paper with sound analytic derivations and honest simulation support. The main weakness is experimental: the LmAPD claim of prediction is actually fit-based, the forward wedge is broad, the independence assumption of glow is untested, and Table 2 shows a notable inconsistency at 6 V. These are fixable with additional analysis or reframed claims, so I recommend major revision rather than rejection. If the authors can provide a direct measurement of per-read glow variance and correlation, or clearly delimit the validity regime of the independent-Poisson model, the paper could become a strong contribution. The H2RG section should be reframed as a glow diagnostic, not a validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the closed-form noise floors for Fowler and SUTR under per-read glow are new and almost certainly correct as random-variable propagation. The simulations in Appendix A back the math. The empirical case is weaker than the abstract implies: the LmAPD 'within 0.1 e-' agreement comes from a two-parameter fit to the noise-vs-N curves, and the forward model is a wedge sweeping cF_g from 0.3 to 1.0, so it is permissive. The H2RG data are background-limited and do not confirm the floor. I would still send this to a referee, because the analytic result is useful and the paper is honest about some limitations.\n\nWhat is genuinely new: earlier work identified glow as a noise source, but no one derived the optimal-read-count and sigma_min scaling for Fowler and SUTR. The amplified treatment with partial gain and excess noise is a real extension, and the photon-transfer slope cF_g is a clever diagnostic. The FIRE glow measurement is a useful data point.\n\nSoft spots: the central scaling assumes per-read glow is independent Poisson. If glow has temporal correlations or overdispersion, the floor formula and the attribution to glow rather than 1/f do not follow. The paper never tests this. The LmAPD experiment fits the model, so it cannot validate the assumption; the wedge is broad enough that correlated or non-Poisson glow could probably fit too. The forward prediction is not as tight as the abstract suggests: the solid curves are fits, the wedge uses cF_g swept over a factor >3, and the fitted glow variance sits at the top of the wedge while the direct measurement suggests lower—offsets acknowledged but unresolved. H2RG measured noise is above theory and background-dominated, so it is a glow-rate measurement, not a validation. No code or data shipped.\n\nWho benefits: anyone building or using IR arrays for HWO or ELTs. The 1/4-power dependence is a sobering result.\n\nRecommendation: peer review, yes. Referee should ask for a direct test of glow independence/correlation, a reframing of the LmAPD agreement as a fit consistent with direct glow, and data/code release. If addressed, I'd take the main result as established.","headline":"Clean analytic floor for glow-limited averaging, but the LmAPD validation is a fit and the independence assumption is untested.","tokens_in":16569,"tokens_out":5091,"would_cite":true,"duration_ms":46999,"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":"Readout glow, not 1/f drift, sets the floor on infrared read-noise averaging","keywords":["infrared detectors","readout glow","read noise","Fowler sampling","sample-up-the-ramp","linear-mode avalanche photodiodes","noise floor","non-destructive readout"],"falsifier":"Measure the per-read glow contribution as a time series by reading out fast with a reset between every frame and differencing; then estimate the autocorrelation of consecutive glow estimates. If the autocorrelation at any nonzero lag is clearly nonzero, the independent-Poisson premise fails. Alternatively, suppress glow with a blocking layer and check whether the noise-versus-N floor moves exactly as σ_RN^{1/2} G^{1/4} with the measured change in G.","tokens_in":15576,"feed_emoji":"💡","tokens_out":6961,"duration_ms":58421,"temperature":0.7,"pith_summary":"Infrared detectors are read out many times to average down read noise, but in practice the improvement stalls after a few dozen reads. The paper argues that the culprit is readout glow: photons emitted by the pixel's own electronics during each read. Because glow adds a new Poissonian photocount on every read, its variance grows with the number of reads instead of averaging down. The consequence is a minimum achievable noise that scales as the one-fourth power of the glow rate and one-half power of the read noise—roughly 2–3 electrons for conventional H2RG-style sensors and about 0.5 electrons for linear-mode avalanche photodiodes. Laboratory data on an LmAPD track the predicted noise floor to within 0.1 electron over two decades of averaging, indicating glow rather than 1/f drift explains the plateau.","feed_headline":"Glow, not 1/f drift, sets infrared read-noise floor","feed_subtitle":"Averaging stops improving near 0.5-3 e- because every read adds glow photons; LmAPD data match to 0.1 e-.","key_machinery":"The central object is the per-read glow realization g_k, treated as an independent Poisson random variable with mean and variance G. The derivation counts how many times each g_k enters the averaged readout: in Fowler-N, the first group's average contains g_k with weight (N−k+1)/N, which generates the cumulative variance and the covariance between the two groups that survives subtraction. Balancing the linearly growing glow variance (2N/3)G against the 1/N read-noise term yields the optimum N and the σ_min ≈ 1.5 σ_RN^{1/2} G^{1/4} floor; the same machinery extends to SUTR and, with substitutions σ_RN→σ_RN/M and G→F_g c^2 G, to avalanche photodiodes.","core_discovery":"For Fowler sampling with N reads in each group, the variance of the difference is (2N/3 + 1/(3N))G + 2σ_RN^2/N, where G is the per-read glow mean and variance and σ_RN the single-read noise. The read-noise term shrinks as 1/N while the glow term grows linearly with N, so the curve turns around at N_opt ≈ 1.7 σ_RN/√G, leaving σ_min ≈ 1.5 σ_RN^{1/2} G^{1/4}. Sample-up-the-ramp gives the same structure with coefficients 3.2 and 2.8. For amplified sensors, gain suppresses read noise as 1/M^2 but leaves the glow variance (scaled by F_g c^2) untouched, so the floor improves only as M^{−1/2}. The paper confirms this by measuring glow independently, fitting the noise-versus-N curve, and finding agre","pith_inferences":["If glow is the dominant plateau mechanism, then detectors designed with glow-blocking layers or low-glow ROICs should show a lower noise floor that scales as the one-fourth power of the residual glow—an experimentally testable prediction the authors do not carry out.","The same cumulative-noise mechanism may apply to any integrating detector that injects a per-sample charge, such as some CMOS or skipper-CCD readout modes; revisiting their averaging behavior with a per-sample noise term might resolve similar plateaus.","The ratio of the fitted glow variance to the directly measured mean glow gives cF_g (Eq. 29), a quantity that could be used to identify the physical photon-generation mechanism inside the source follower.","For mission planning, the floor implies that future space observatories cannot rely on averaging alone to reach sub-electron noise; glow mitigation at the ROIC level becomes a mission-critical requirement."],"forward_implications":["Any non-destructive readout scheme has an optimal read count; beyond it, additional reads make the measurement noisier.","Reducing glow by a factor of ten cuts the noise floor only in half (one-fourth power), so meaningful gains require large glow reductions or blocking layers.","In linear-mode avalanche photodiodes, glow—not 1/f drift—is what stops averaging from reaching deep sub-electron noise; the floor sits near 0.5 e− at current glow levels.","Fowler and SUTR estimators are biased by glow (e.g., Fowler-N adds a bias of N G electrons), which must be subtracted for precision photometry.","The noise-versus-N curve shape can be used to separate glow from other noise sources in real detectors."],"fun_headline_variants":["Readout glow, not 1/f drift, sets IR detector noise floor","Averaging infrared reads hits a wall: readout glow","Glow photons from electronics limit infrared detector sensitivity","IR sensor noise floor: readout glow beats averaging","Readout glow caps averaging benefit in infrared detectors"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The formulas assume that the glow photoelectrons added on each read are independent Poisson draws with the same mean and variance; if glow has temporal correlations (for instance a 1/f component from the source-follower), or if its per-read variance is not Poisson, the floor scaling and the conclusion that glow rather than 1/f drift causes the plateau would not follow.","fun_headline_variants_meta":{"raw":{"variants":["Readout glow, not 1/f drift, sets IR detector noise floor","Averaging infrared reads hits a wall: readout glow","Glow photons from electronics limit infrared detector sensitivity","IR sensor noise floor: readout glow beats averaging","Readout glow caps averaging benefit in infrared detectors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000264,"raw_usage":{"total_tokens":1495,"prompt_tokens":851,"completion_tokens":644,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":563}},"tokens_in":595,"tokens_out":644,"duration_ms":6260,"temperature":1.0,"reasoning_tokens":563,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T00:38:22.147403+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the per-read glow contribution as a time series by reading out fast with a reset between every frame and differencing; then estimate the autocorrelation of consecutive glow estimates. If the autocorrelation at any nonzero lag is clearly nonzero, the independent-Poisson premise fails. Alternatively, suppress glow with a blocking layer and check whether the noise-versus-N floor moves exactly as σ_RN^{1/2} G^{1/4} with the measured change in G.","supporting_citations":[],"review_version":1}