Pith. sign in

REVIEW 3 major objections 5 minor 12 references

Impact of common modes correlations and time sampling on the total noise of a H2RG near-IR detector

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read One linear correction cuts infrared detector read noise in half.

desk verdict Solid engineering result for H2RG common-mode correction, but the temporal 'predictive model' is in-sample interpolation, not a fixed law, so the abstract overclaims. read the letter →

arxiv 2506.04995 v1 pith:SDXGBGFN submitted 2025-06-05 astro-ph.IM

B. Kubik , R. Barbier , A. Castera , E. Chabanat , S. Ferriol , G. Smadja This is my paper
classification astro-ph.IM
keywords H2RGdetectorsHgCdTecommonmodesreferencepixels1/fnoisetemporalfrequencypowerspectruminfraredastronomy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish two things about a 2K×2K HgCdTe near-infrared detector array of the type planned for space use. First, that subtracting a combination of the array's built-in reference pixels — the average of four rows above and below, plus averages of four columns at the array edges — removes enough correlated readout noise to cut the standard CDS noise by 51% and the multi-sampled Fowler(16) noise by 65%. Second, that the remaining noise's frequency power spectrum has the simple form $A + B/\omega^\alpha$, so the Fowler noise at other frame and group sampling times can be predicted without measuring each sampling scheme separately. If both claims hold, a single reference-pixel correction plus a three-parameter spectrum gives a practical recipe for lowering readout noise on this class of infrared detectors and for choosing sampling schemes.

What carries the argument

The load-bearing object is the combined reference-pixel correction $c_{3mn}(x,y)$, defined in Equations (1)-(3): it forms sliding-window averages of the four reference rows above and below each active pixel ($U$ and $D$) and of the four reference columns left and right ($L$ and $R$), removes the up/down component from the left/right averages before summing, and applies interpolation to the up/down but not the left/right averages. With $x=64$ and $y=4$ it is the most effective of the corrections tested. The second mechanism is the variance formula (Equation 9), built on the autocorrelation-power-spectrum relation, which expresses Fowler group-to-group variance as an integral of $(1-\cos(\omega\Delta))$ times the assumed power spectrum $A + B/\omega^\alpha$ times a sampling kernel depending on $n$, $\delta$, and $\Delta$; fitting $A$, $B$, $\alpha$ to one exposure then yields predictions for other samplings.

What would settle it

Record long UTR exposures of the same detector with a frame delay and group spacing not used in the fits (for example $\delta = 0.5$ s and $\Delta = 80$ s), fit $A$, $B$, $\alpha$ on one subset, and predict the Fowler variance for the other; if the measured variance disagrees by more than the stated 10% or the fit requires an $\alpha$ outside 1.14-1.5, the single-power-spectrum model is falsified.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that the dominant readout noise in this H2RG array is a common-mode signal shared across pixels, and that a linear correction built from the reference pixels removes most of it. The best correction, $c_{3mn}(64,4)$, computes the average of the four up and four down reference rows over the full channel width ($x=64$), the average of the four left and four right reference columns over a 9-pixel vertical window ($y=4$), subtracts the up/down average from the left/right average first, then subtracts the sum from each active pixel. Applied to dark exposures it lowers CDS noise from 30.9 to 15.1 $e^-$ r.m.s. and Fowler(16) noise from 17.4 to 6.1 $e^-$ r.m.s., and it reduces the spatial inhomogeneity of the noise across the 32 outputs. The paper also claims that the group-to-group variance of Fowler sampling is governed by a power spectrum $A + B/\omega^\alpha$, extracted through the autocorrelation-power-spectrum relation, with fits giving $\alpha$ between 1.14 and 1.34 for the tested group periodicities and needing 1.5 for very low frequencies; using these parameters, predicted Fowler variances match measurements to about 3% for short group intervals and 10% for long ones, including the specific frame and group timings of a planned space instrument.

Load-bearing premise

The predictions rest on the assumption that a single three-parameter curve describes the noise's frequency dependence at every sampling scheme; the paper's own fits require the curve's slope exponent to change from 1.14 to 1.34 as the group spacing changes, and to be set to 1.5 for very long intervals, so the assumption is not actually satisfied by the data.

Editorial extensions

If this is right

  • Applying $c_{3mn}(64,4)$ to dark exposures of this detector lowers CDS noise by 51% and Fowler(16) noise by 65%, so a space instrument with this array can gain nearly a factor of two in read-noise-limited sensitivity without hardware changes.
  • The same correction reduces the reset-level drift with bias voltage by 95% (from 70 to 1.6 ADU/mV) and cuts temperature-induced reset-level variations by a factor of five.
  • Because a single power spectrum $A + B/\omega^\alpha$ predicts Fowler noise to about 3% for group intervals under 15 seconds and about 10% up to roughly 50 seconds, individual noise measurements for each sampling scheme can be replaced by interpolation.
  • The measured noise does not decrease as $1/\sqrt{n}$ with repeated sampling; the residual time correlations show that 1/f-type noise from readout electronics must be included in noise budgets.
  • For the timing parameters proposed for the planned space instrument, the predicted total variances match measured values to better than 10%.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same $c_{3mn}$-style correction should transfer to any H2RG-like array with four edge rows and columns, but the optimal $x$ and $y$ window sizes may shift with channel count and pixel rate; this paper only tests one configuration.
  • Editorial inference: the need to raise $\alpha$ to 1.5 for group intervals above about 50 seconds suggests a second low-frequency process with a steeper spectrum; a two-component model fitted jointly to all group spacings would test whether a single functional form can cover the full range.
  • Editorial inference: the predictive model could be inverted into a sampler optimizer: given a fixed exposure time, the pair $(\delta, \Delta)$ that minimizes total variance under the fitted spectrum can be found by grid search; the paper does not carry out that optimization.
  • Editorial inference: since the temporal analysis used spatial frame averaging rather than the $c_{3mn}$ reference-pixel correction, combining both corrections might push the low-frequency noise floor lower than either alone; that combination is not tested here.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper characterizes the readout noise of an H2RG HgCdTe near-IR detector similar to those planned for the Euclid NISP instrument. The authors define and compare common-mode corrections built from reference-pixel sliding-window averages, reporting that the combined up/down and left/right correction c3mn(64,4) lowers CDS noise from 30.9 to 15.1 e- (51%) and Fowler(16) noise from 17.4 to 6.1 e- (65%). They also develop a temporal noise model based on a power spectrum |f|^2 = A + B/omega^alpha and use it to predict Fowler-noise variances for other group and frame samplings, claiming agreement within about 3% for short intervals and 10% for long intervals, with the stated goal of avoiding ad-hoc noise measurements for each sampling scheme.

Significance. The common-mode correction study is a solid, practically relevant contribution: it is internally consistent, quantitatively detailed, and directly applicable to Euclid NISP and similar H2RG-based instruments. The temporal model, if it were a true out-of-sample prediction, would be valuable as an interpolation formula replacing per-scheme noise calibrations. However, as presented the prediction claim is not supported by the evidence because the model parameters vary with the group periodicity and the comparisons are in-sample. The paper also benefits from transparent algorithm definitions and realistic acknowledgement that only one engineering-grade detector was used, but the absence of error bars limits the strength of the quantitative claims.

major comments (3)
  1. [Section 5, Eq. (10) and Table 4] The power-spectrum model of Eq. (10) assumes a fixed spectral index alpha, but the fits in Table 4 give alpha = 1.14, 1.15, and 1.34 for group periodicities Delta = 0.46, 3.56, and 14.24 s at the same frame period delta = 7.12 ms. Since Delta is an analysis grouping choice rather than a detector property, a stationary detector noise spectrum cannot depend on it; the variation of alpha with Delta indicates that Eq. (10) is a local fit parameter, not a stable predictive law. This directly undermines the claim in the abstract that ad-hoc readout noises for different samplings can be avoided.
  2. [Section 5.1, Table 5] The 'predicted' Fowler variances in Table 5 are not out-of-sample predictions. For Delta < 15 s the parameters are taken from set 1 or set 2 fitted at those same group periodicities; for 15 < Delta < 50 s the set 3 parameters are used; and for Delta > 50 s alpha is manually set to 1.5. No independent exposure or cross-validation is presented. The 3-10% agreement therefore demonstrates in-sample interpolation with regime-dependent parameter selection, not predictive power. The abstract and conclusions should be rewritten to present the model as an interpolator over the tested (delta, Delta) range, with an explicitly stated validity domain.
  3. [Section 5, Tables 4 and 5] No uncertainties are quoted for the fitted parameters A, B, and alpha, nor for the measured variances in Tables 4 and 5. Given that the central quantitative claims are the 3-10% agreement and the dependence of alpha on Delta, the absence of error bars makes it impossible to judge whether the alpha variation is statistically significant or whether the 3-10% residuals are meaningful. The authors should provide uncertainties on all fitted and measured quantities, and ideally validate the model on an independent dataset not used for the fits.
minor comments (5)
  1. [Section 5, paragraph after Eq. (9)] The text states 'delta = 7.12 µsec'; this should be 'msec' (the correct value 7.12 msec appears elsewhere in the paper).
  2. [Section 3.3] The sentence 'the optimal box for the up and down pixels is the full channel average: x = 64 with the CDS noise reduced by 0.4 e− r.m.s' is inconsistent with Table 1, which shows a CDS reduction of 9.9 e− for c1n(64); please clarify what the 0.4 e− reduction refers to.
  3. [Table 5] The table formatting is hard to parse: rows such as 'set 1,27.40 12.36 12.66' and '0.00712 57.07 26.10 26.17' lack clear column separation, and the entry 'α = 1.51' appears to contain a stray digit. Please reformat the table for readability.
  4. [Eq. (11)] The variable D_I is introduced without a clear definition or subscript formatting; please define it explicitly (e.g., as the signal difference between consecutive groups) and state its units.
  5. [Throughout] There are several typographical errors, including 'Anther way' in Section 3.1, 'form the mean value' in Section 3.3, and 'an 2K × 2K' in the abstract; these should be corrected.

Circularity Check

2 steps flagged · score 6.0 of 10

The temporal 'predictive model' is in-sample: A, B, α are fitted to the same UTR exposures, switched per Δ range, and α is hand-set to 1.5 for slow readouts, so the quoted 3-10% agreement is interpolation rather than independent prediction.

  1. fitted input called prediction [Section 5.1, Eq. (10), Tables 4 and 5]
    "We can also check how accurately the evolution of Fowler noise with the group and frame period is predicted using the frequency power spectrum with the parameters found in our analysis. ... In the range ∆ < 15 sec the parameters in set 1 and set 2 describe the noise with an accuracy better than 3%. ... In the very low frequency regime, ∆ > 50 sec, in order to predict accurately the variance the power α must be increased up to 1.5."

    The parameters A, B, and α in Eq. (10) are fitted to σ²_F(n, δ, Δ) from the same 25 UTR exposures (Table 4, δ = 7.12 ms, Δ = 0.46, 3.56, 14.24 s). Table 5 then compares Eq. (9)-(10) with 'measured' variances from those same exposures, while switching parameter sets per Δ range and hand-setting α = 1.5 for Δ > 50 s. The 3-10% agreement is therefore an in-sample interpolation with a flexible three-parameter curve and a regime-dependent spectral index, not an independent out-of-sample prediction. This does not support the abstract's claim that ad-hoc readout noises for different samplings can be avoided.

  2. self citation load bearing [Section 5, Eq. (10), refs. [5] and [6]]
    "Following the authors of 5 the power spectrum is assumed to be of the form |f(ω)|² = A + B/ω^α. As the authors of 6 pointed out, this form of the parametrization is suggested by the general frequency behavior of circuit components."

    The central temporal predictive claim rests on the power-spectrum ansatz Eq. (10), which is not derived in this paper but imported from refs. [5] and [6], both by the same first author (Smadja) who is also an author here. The cited papers do not provide an independent, parameter-free justification: the ansatz is the very assumption being tested. The later finding that α must be changed from 1.14/1.15/1.34 to 1.5 depending on Δ shows that this self-cited form is not a stable predictive law, so the model has no fixed content outside the fitted regimes.

full rationale

The common-mode correction analysis (Section 3, Tables 1-3) is internally coherent and not circular: the c3mn(64,4) correction is defined by reference-pixel averages and compared against raw or differently corrected noise on the same data, which is a legitimate data-reduction comparison. The circularity is concentrated in the temporal 1/f predictive model. Its parameters A, B, and α are fitted on the very UTR exposures whose derived variances are later called 'measured' in Table 5, and the comparison uses different parameter sets per frequency regime, including a hand-set α = 1.5 for slow readouts. Thus the quoted 3-10% agreement is an interpolation, not an independent validation. The self-cited functional form of Eq. (10) is also load-bearing and unverified. The paper contains useful independent measurements (reference-pixel corrections, reset noise, temperature drift), so the overall circularity is partial rather than total; score 6 reflects that the central predictive claim reduces largely to in-sample fitting with regime-tuned parameters.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. Its central dependence is on fitted correction-window sizes and on a three-parameter noise power spectrum that is assumed rather than derived; the temporal model also leans on the authors' own earlier papers.

free parameters (4)
  • Common-mode box sizes x and y = x = 64, y = 4
    Chosen to minimize CDS and Fowler noise in Figures 2; the central c3mn correction uses these values.
  • Power-spectrum amplitude A = 0.388 to 0.412 uV^2/Hz
    Fitted jointly via Eq. (9) to measured Fowler variance for each group periodicity in Table 4.
  • Power-spectrum 1/f coefficient B = 119.04 to 139.12 uV^2/Hz
    Fitted jointly with A and alpha for each Delta value in Table 4.
  • Power-law exponent alpha = 1.14, 1.15, 1.34, and 1.5 by hand for slow sampling
    Not constant across group periods, which undermines the assumption of a fixed power-spectrum form.
assumptions (4)
  • standard math Wiener-Khinchin theorem relates autocorrelation to power spectrum
    Invoked in Eq. (8) as the basis for the variance formula in Eq. (9).
  • domain assumption Reference pixels mimic the common modes of active pixels
    All corrections c1-c3 subtract reference-pixel averages from science pixels, assuming the two share the same bias, clock, and temperature fluctuations.
  • ad hoc to paper Power spectrum has the fixed form A + B/w^alpha
    Assumed in Eq. (10) based on refs 5 and 6; the paper's own fits show alpha varying with sampling, so the form is not universal.
  • domain assumption Group differences are stationary and reset transients are negligible after 50 frames
    Used in Section 5 before fitting; residual reset transients could contaminate the low-frequency part of the power spectrum.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Impact of common modes correlations and time sampling on the total noise of a H2RG near-IR detector." pith.science (2026). https://pith.science/paper/SDXGBGFN

@misc{pith2026250604995,
  author       = {Pith},
  title        = {Pith review of: Impact of common modes correlations and time sampling on the total noise of a H2RG near-IR detector},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SDXGBGFN}},
  note         = {Machine review of arXiv:2506.04995}
}
read the original abstract

We present the readout noise reduction methods and the 1/f noise response of an 2Kx2K HgCdTe detector similar to the detectors that will be used in the Near Infrared Spectrometer Photometer - one of the instruments of the future ESA mission named Euclid. Various algorithms of common modes subtraction are defined and compared. We show that the readout noise can be lowered by 60% using properly the references provided within the array. A predictive model of the 1/f noise with a given frequency power spectrum is defined and compared to data taken in a wide range of sampling frequencies. In view of this model the definition of ad-hoc readout noises for different sampling can be avoided

Figures

Figures reproduced from arXiv: 2506.04995 by the authors.

Figure 1
Figure 1. The sliding windows parameters definition for a pixel ( [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. CDS noise in e− r.m.s. in the channel 20 after reference pixels corrections with different window size. Left: Corrections c (20) 1n (x) and c (20) 1mn(x) interpolated. Middle: Corrections c2m(y) and c2mn(y) interpolated. Right: Corrections c3mn(x, y) for different parameter sets. The most effective correction is c3mn(x, y) with x = 64, y = 4. 4. RESET NOISE AND TEMPERATURE DRIFTS MITIGATION WITH REFERENCE PIXELS For… view at source ↗
Figure 3
Figure 3. CDS noise in e− r.m.s. after reference pixels corrections. Left: c2m(4). Right: c (ch) 3mn(64, 4) for channels 0, 1, 15, 25, 30 and 31. Correction c2m(4), which uses only left and right reference pixels, reduces the noise in the channels 0 and 31 more effectively that in other channels. Correction c (ch) 3mn(64, 4) uses left/right and up/down reference pixels and reduces the noise to the same level in all the channe… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: On the top: reset level and reset noise per pixel as function of [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: On the top: reset level and reset noise per pixel as function of the temperature for different reference pixels [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Fowler noise σ 2 F (n, δ, ∆) as a function of n. The contribution of the dark current and stray photons has not been subtracted but their contribution is negligible. The noise is computed for raw data, then for pixels after reference channel subtraction (refsub data) a…
Figure 7
Figure 7. Figure 7: The measured group to group readout noise as a function of the number of repeated reads and the fitted [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

  1. [1]

    Teledyne Imaging Sensors HA WAII-2RG Technical ManualMAN-0001, Rev 4, 2009-06-05

    work page 2009

  2. [2]

    Moseley, S., H., et al., Proceedings of SPIE, 7742 [2010]

    work page 2010

  3. [3]

    J., et al., Proceedings of SPIE, 8155 [2011]

    Rauscher, B. J., et al., Proceedings of SPIE, 8155 [2011]

    work page 2011

  4. [4]

    J., et al

    Rauscher, B. J., et al. 2007, PASP, 119, 768-786 (erratum 122, 1254 [2010])

    work page 2007

  5. [5]

    Smadja, G., et al., Nuclear Instruments and Methods in Physics Research A 622 (2010) 288-294

    work page 2010

  6. [6]

    Smadja, G., et al., Nuclear Instruments and Methods in Physics Research A 694 (2012) 95–100

    work page 2012

  7. [7]

    Finger, G, et al., Proceedings of SPIE, 5499, p.47 [2004]

    work page 2004

  8. [8]

    Smith, R., et al., Proceedings of SPIE, 6276, p.62760R [2006]

    work page 2006

Show all 12 references
  1. [9]

    Schubnell, M., et al., Proceedings of SPIE, 6276, p.62760Q [2006]

    2006

  2. [10]

    Schottky, W., Phys. Rev. 28 (1926) 74

    1926

  3. [11]

    Ferriol, DAS v.1.0 IPNL SW deliverable PTFB2 CNES-IN2P3

    S. Ferriol, DAS v.1.0 IPNL SW deliverable PTFB2 CNES-IN2P3

  4. [12]

    Kubik, E

    B. Kubik, E. Chabanat, SELENE v.1.0 IPNL SW deliverable PTFB2 CNES-IN2P3

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.