What the authors report
The paper compares Prime 95B, Sona-11, Sona-6, Ximea xiJ and Orca-Quest cameras. For the Orca-Quest at −20 °C with air cooling, Table 2 gives standard-mode read noise of 0.37 ± 0.09 e⁻ and dark current of 0.0067 ± 0.0003 e⁻/s; ultra-quiet mode gives 0.22 ± 0.07 e⁻ and 0.0198 ± 0.0008 e⁻/s. The photon-number-resolving mode has a different calibrated output and no conventional read-noise value in that table. Sampling was not uniform across every camera, and Sona-11 lacked the same pixel-wise analysis. This is imaging characterisation, not EPRV qualification. [1]
Reading the evidence
| Observable | What I would retain | Why a single average is insufficient |
|---|---|---|
| Conversion gain | Per-pixel map and repeat measurements. | A gain error changes the electron-scale noise estimate. |
| Read noise | Distribution, tails and spatial correlations. | A small noisy population can affect extraction. |
| Dark response | Exposure and temperature dependence. | Long exposures weight unstable pixels differently. |
| Linearity | Residual versus signal in the intended mode. | A global score can conceal local response structure. |
The measurement chain
A conceptual reading aid, not measured data or a reproduction of a paper figure:
- Photons and readout mode
- Pixel-dependent response
- Calibrated extraction
- Flux / centroid validation
A detector noise budget needs an estimator
For independent noise terms in a simple aperture, a schematic variance is S + B + nₚ(Dt + σᵣ²), with source and sky electron counts S and B, nₚ pixels, dark rate D, exposure t and read noise σᵣ. This is a bookkeeping model, not a complete camera specification. Correlated noise and calibration uncertainty require additional terms.
Changing pixel size also changes sampling and the number of pixels used by an estimator. A lower per-pixel read noise may not give a lower extracted-spectrum variance if the same optical profile spans many more pixels. Comparisons should hold the intended measurement and sampling requirement fixed.
Why maps matter more than a brochure value
A wavelength centroid is a weighted combination of pixel signals. If pixel response errors vary across the line profile, their effect depends on where the profile lands. A detector can perform well for total flux while introducing a small position-dependent centroid bias.
I would preserve raw frames, acquisition settings, calibration maps and masking decisions together. Reprocessing then becomes possible when a noisy tail, unstable pixel population or amplifier pattern is found later. A single corrected image without that provenance is difficult to audit.
The boundary between low noise and accurate spectroscopy
Sub-electron read noise is valuable, but it does not establish intra-pixel uniformity, stable pixel geometry, timing accuracy or a negligible flux-dependent centroid shift. Those are separate observables with separate tests.
Nor should a proprietary integer-valued output be treated as an ordinary analogue frame without checking its transfer function. Quantisation, calibration and saturation can change the interpretation of variance. The data model must follow the actual output mode.
A spectroscopy-specific validation plan
Flux sweep: move an unresolved, stable line through a range of brightnesses while holding illumination geometry fixed. Inspect centroid and width residuals as well as integrated flux linearity.
Sub-pixel scan: translate the line slowly across pixel phase and repeat at several detector locations. Separate optical motion from detector response using an independent position reference.
Repeatability: repeat after power cycling, mode changes and temperature cycles. Use a fixed reference reduction first, then quantify how much recalibration is required.
Covariance: examine row, column and temporal residual correlations. A per-pixel standard deviation does not capture noise shared by many pixels.
Connection to instrumentation work
For a precision spectrograph I would write a detector acceptance test around the final scientific estimator, not around camera specifications alone. The useful deliverable is a reproducible link from pixel behaviour to extraction and wavelength error.
This also fits the wider control-system theme: stabilise the hardware, record its state, and verify the result out of loop. A quiet electronic readout does not remove the need to test the optical measurement.
Following the full technical argument
The experiment measures a population of pixels, not one camera number
A CMOS array contains many local electronic response chains. A global mean can look excellent even if a small subset of pixels is unstable or if row-level correlations remain. That is why gain, bias, read-noise and dark-response distributions need to be followed through the paper rather than reduced to a product ranking.
Conversion gain links the digital output to electron counts. In a shot-noise-dominated regime, variance and mean should obey a predictable relation once the convention for gain is fixed. An incorrect conversion changes every downstream electron-based claim. Mean–variance plots are consequently a calibration diagnostic as well as a convenient way to quote gain.
Repeated data sets are especially useful because they ask whether the calibration map transfers beyond one acquisition. A map that fits one exposure sequence can hide source instability or readout-state changes. The paper flags unequal sampling among cameras; I would not convert those comparisons into a precise vendor league table.
Read noise, dark current, quantum efficiency and linearity answer different questions
Read noise describes variation associated with reading out a pixel, whereas dark current accumulates with exposure time. Their relative importance changes with cadence and illumination. A slow low-noise mode is therefore not automatically optimal for a rapid measurement, and an attractive bias-frame result does not establish low-noise performance during a long exposure.
Quantum efficiency is a photon-conversion measurement with its own reference detector and optical path. The front window matters: a combined camera response must not be labelled a bare-sensor response without correcting for transmission. The reliable wavelength range of a revised measurement should also be respected rather than interpolated into unsupported ultraviolet performance.
Linearity must be read as residual versus signal, with saturation and output processing visible. An average linearity score can conceal a local response bend at precisely the signal level used by a scientific estimator. For photon-number-resolving output, the calibration algorithm changes the distribution being measured; it should not be treated as an ordinary analogue mode with a familiar noise statistic.
Why the appendix changes the strength of the conclusion
The appendix is not only extra camera plots. It contains warning signs that qualify comparisons: incomplete pixel-wise calibration in one case, unusual gain populations in another and a mean–variance inconsistency requiring further investigation. Those are reasons to inspect the measurement procedure, not reasons to infer a violation of photon statistics.
A variance below a naïve shot-noise expectation can arise from processing, correlations, gain conventions or an inadequate experimental reference. The correct next step is an independent illumination and acquisition check. It is not to present the camera as beating the fundamental randomness of incident photons.
For a spectrograph, the next validation layer is the movement of a line or spectrum over these calibrated pixels. A precise integrated flux does not guarantee an unbiased centroid. I would propagate measured response maps through realistic extraction weights and test brightness, pixel phase, detector position and readout state before claiming precision-RV suitability.
Reading based on the full manuscript, including methods and discussion—not only its abstract. These explanations are original commentary; they are not a reproduced analysis or new experiment. Original manuscript · expanded 21 October 2026.
Paper plots, explained in detail
Read the paper plots alongside the explanations below. Each discussion keeps its source figure number and links to the original caption and context. Openly licensed panels are shown here; other figures open at the original paper. The explanations distinguish the plotted evidence from what it cannot establish.
Gain distributions and the mean–variance calibration
- How to read the axes and panels
- Rows correspond to different Orca-Quest output modes. Distributions are paired with variance–signal behaviour for representative pixels, so this is more than a summary histogram.
- What the plot is telling me
- Inspect consistency across signal level and repeated acquisitions before trusting one electron-per-ADU conversion. Differences between the central and tail pixels show why a global gain can lose information. The high-signal turnover must be distinguished from a low-noise operating region.
- What it does not establish
- Saturation and proprietary processing alter the mean–variance relation. A falling variance near a digital limit is not evidence that photon shot noise has disappeared.
Source Figure 1 and original caption ↗ · Original author/publisher rights retained. Open the figure at the source alongside this commentary; no additional copy is distributed here.
Read-noise distributions with a display offset
- How to read the axes and panels
- The two-dimensional distribution combines bias and per-pixel scatter, with marginal histograms. One mode’s bias is deliberately displaced for visibility.
- What the plot is telling me
- The figure is useful for comparing the width and tails of the read-noise populations. It should not be used to infer a physical bias separation from the plotted horizontal displacement, since that visual offset is a presentation choice.
- What it does not establish
- A narrow marginal distribution does not test spatial or temporal covariance. Shared amplifier noise can matter to an extracted spectrum even when each pixel’s scatter is small.
Source Figure 2 and original caption ↗ · Original author/publisher rights retained. Open the figure at the source alongside this commentary; no additional copy is distributed here.
Dark response and the effect of output quantisation
- How to read the axes and panels
- The horizontal direction is dark-current estimate and the plotted populations correspond to different modes at the stated temperature. The photon-counting output has special structure.
- What the plot is telling me
- Look at tails as well as the central population. A small set of warm pixels can matter disproportionately in long integrations. Features introduced by integer-valued output should be interpreted through the acquisition duration and processing rather than immediately assigned to new detector physics.
- What it does not establish
- This distribution is conditional on the measurement setup and exposure length. It is not a temperature law, a persistence test or a guarantee for every unit of the same camera model.
Source Figure 3 and original caption ↗ · Original author/publisher rights retained. Open the figure at the source alongside this commentary; no additional copy is distributed here.
Camera response includes the front window
- How to read the axes and panels
- The curve shows a wavelength-dependent quantum-efficiency measurement for the sensor-plus-window system, over the part of the experiment considered reliable.
- What the plot is telling me
- Read the peak together with the wavelength range and optical reference. A camera can be well matched to one band and poorly matched to another, so a single peak percentage cannot stand in for the throughput of a whole observing programme.
- What it does not establish
- The measured combined response should not be silently substituted for bare-sensor QE. Unsupported portions of an older or unreliable measurement are not repaired by drawing a smooth continuation.
Source Figure 4 and original caption ↗ · Original author/publisher rights retained. Open the figure at the source alongside this commentary; no additional copy is distributed here.
Linearity needs the residual panels
- How to read the axes and panels
- The rows compare modes, and each row pairs a signal response with departures from a fitted linear relation. Representative pixels sample different gain locations.
- What the plot is telling me
- The residual panels expose departures that can be hard to see in an almost straight main curve. Compare those departures at the intended operating signal, not only a global score. Saturation and the calibrated counting output have different limits and require different interpretation.
- What it does not establish
- A global linearity metric is not a bound on every pixel or on wavelength-centroid error. The latter depends on spatial response and the estimator.
Source Figure 5 and original caption ↗ · Original author/publisher rights retained. Open the figure at the source alongside this commentary; no additional copy is distributed here.
An appendix anomaly that should not be oversold
- How to read the axes and panels
- The Ximea panel combines gain distributions with a mean–variance comparison and an explicit concern in the paper about the apparent noise relation.
- What the plot is telling me
- This is an example of a valuable unresolved diagnostic. If a variance curve appears inconsistent with the assumed photon-statistics reference, the assumptions and acquisition must be revisited. Possible processing or calibration effects deserve an independent measurement before a physical conclusion.
- What it does not establish
- Do not describe an apparent sub-shot-noise curve as a new camera capability. The authors identify it as something needing investigation; the reading note should retain that uncertainty.
Source Figure 17 and original caption ↗ · Original author/publisher rights retained. Open the figure at the source alongside this commentary; no additional copy is distributed here.
Sources and scope
- Khandelwal et al. (2024). Beyond CCDs: characterising sCMOS detectors for optical astronomy. Read the version used: arXiv:2409.16449v2 · Abstract and version history.
This is a selective literature-reading note. Source claims appear in “What the authors report”; the comparison table, conceptual chain and proposed tests are analytical commentary. No new observations, reanalysis or experimental validation are claimed. Expanded source review: 21 October 2026.