Pith. sign in

REVIEW 4 major objections 7 minor 18 references

X-Sifter: detecting transients in X-ray data using the optimal Poisson matched filter

T0 review · 4 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read X-Sifter detects 30% more X-ray sources than the standard Chandra catalog.

desk verdict A worthwhile pipeline paper whose headline sensitivity gain over CSC2.1 is not yet quantitatively established because the comparison does not control for threshold differences. read the letter →

arxiv 2412.07858 v1 pith:NMXTK44K submitted 2024-12-10 astro-ph.IM astro-ph.HE

Maayane T. Soumagnac , Eran O. Ofek , Shachar S. Israeli , Guy Nir , Imri A. Dickstein This is my paper · ORCID
classification astro-ph.IMastro-ph.HE
keywords PoissonmatchedfilterX-raysourcedetectiontransientChandrapointspreadfunctionbackgroundestimationsurveyspeed
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

This paper presents X-Sifter, a software pipeline that implements the Poisson noise matched filter for detecting sources in photon-counting images. The central claim is that, when applied to single-observation Chandra fields, this filter recovers about 30 percent more real sources near the detection limit than the Chandra Source Catalog, and reports signal-to-noise ratios roughly 1.3 times higher at that limit, equivalent to a 1.8-fold increase in survey speed. The work matters because it shows that a principled likelihood-ratio statistic can outperform commonly used heuristic filters in the low-count regime, and because the pipeline includes a temporal subdivision option that can reveal short-lived transients that are diluted in stacked exposures.

What carries the argument

The load-bearing object is the Poisson noise matched filter kernel $K_{\mathrm{PMF}}(F)=\ln(1+\frac{F}{B}P)$, where $P$ is the normalized point spread function and $B$ the background, applied by cross-correlating the image with this kernel. Because the kernel depends on the unknown source flux $F$, the pipeline imposes a thresholding relation (Equation 14) that fixes the filter flux $F_{\mathrm{th}}$ to equal the source flux at the detection limit, making the test approximately optimal for the faintest sources of interest. The implementation handles real data by partitioning each CCD into $16\times16$ sectors with buffers, modeling or rotating PSF stamps per sector and per energy channel, estimating a constant background per sector from the median of gridded tiles, and combining independent energy channels in quadrature. Precomputed libraries of PSF stamps and of the noise distribution $P(S|H_0)$ for a grid of background values keep runtime practical.

What would settle it

Run X-Sifter on simulated images with a known background gradient inside a single sector, such as a linear ramp across the sector with the same mean as a flat field, and compare the recovered flux and signal-to-noise of injected sources to the flat case; if the detection threshold shifts or completeness drops by more than the claimed margin, the sector-constant assumption is falsified. Alternatively, apply the code to a real field with strong structured background, such as the Galactic ridge, and compare against a deeper reference catalog: if the 30 percent gain vanishes or false positives appear, the assumption fails.

Watch

Extended reading notes

Core claim

The central claim is that X-Sifter, a pipeline implementing the Poisson noise matched filter of Ofek & Zackay, recovers real Chandra sources with significantly higher sensitivity than the catalog standard. In a comparison restricted to fields observed only once with Chandra, X-Sifter detected about 30 percent more real sources than CSC2.1, all near the detection threshold. For sources detected by both methods, X-Sifter's signal-to-noise ratio was on average 1.3 times higher near the threshold, which the authors equate to a factor of 1.8 in survey speed. The paper further argues that temporal subdivision of exposures is essential for transient detection, since short bursts are buried in the background of long stacked exposures; in a check on fields observed many times, all quiescent sources found by X-Sifter had CSC2.1 counterparts, while the unmatched detections were variable on timescales shorter than the stack.

Load-bearing premise

The pipeline assumes the background is constant within each 16 by 16 sector and estimates it as the median of gridded tiles after outlier rejection; if the true background varies inside a sector, the filter kernel and threshold are miscalibrated and the claimed sensitivity gain may not be realized.

Editorial extensions

If this is right

  • Single-observation Chandra archival searches gain roughly 30 percent more detections near the threshold at the same false-alarm rate.
  • The 1.3-fold signal-to-noise gain near threshold translates into a 1.8-fold increase in survey speed, allowing surveys to reach the same depth in less exposure time.
  • Temporal subdivision of exposures, when enabled, can expose short transients that are missed by catalogs built from stacked images, at the cost of running the pipeline multiple times.
  • The kernel and thresholding relation are general for any Poisson imaging instrument; the authors state the pipeline can be adapted to XMM-Newton and similar data.

Reading between the lines

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

  • The sector-constant background assumption is likely the limiting factor in crowded fields or regions with strong background gradients; testing on such fields would show whether the claimed gain holds there.
  • The exponential extrapolation of the gamma-S curve used for high-significance conversion is acknowledged in the paper to lose accuracy, so X-Sifter's signal-to-noise values above about 7 sigma may be less reliable than the catalog's.
  • The same filtering formalism could be applied to count images in other Poisson-dominated regimes, such as gamma-ray or neutrino telescopes, with appropriate PSF and background models.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper presents X-Sifter, a source-detection pipeline for X-ray images based on the Poisson matched filter of Ofek & Zackay (2018b). The pipeline partitions each CCD into 16x16 sectors and several energy channels, models the PSF with MARX, estimates a single background value per sector, precomputes the null-hypothesis distribution of the detection statistic to set thresholds, and filters each sector to produce a combined S/N map. The authors validate with injected-source simulations and compare detections with the Chandra Source Catalog 2.1 in ten single-observation fields. They claim ~30% more sources detected, a 1.3x higher S/N near the detection limit, and a factor 1.8 increase in survey speed. They also report a small false-positive check on five stacked fields.

Significance. If the sensitivity gains are real, X-Sifter would be a valuable community tool: it implements the optimal Poisson matched filter in a practical way, with publicly released code, precomputed PSF libraries, and handling of energy-dependent PSF and background variations. The injection tests demonstrate internal consistency of the implementation. However, the headline quantitative claims are not yet convincingly supported: the comparison with CSC2.1 is not performed at matched false-alarm probabilities, the S/N conversion relies on an approximate extrapolation acknowledged by the authors, and the false-positive check uses a different sample than the main comparison. The background estimator is a single median per sector, which may be biased in structured fields. These issues are addressable, but the current paper does not establish the advertised factors.

major comments (4)
  1. [Section 4.2] The claimed ~30% increase in detected sources and the 1.3x higher S/N are not supported by a matched-threshold comparison. The paper does not state the false-alarm probability gamma or the corresponding S/N threshold used by X-Sifter for this comparison, nor does it give the effective significance threshold used by CSC2.1/wavdetect. If X-Sifter ran at a lower threshold, the 55 extra sources near the detection limit would be a threshold artifact. The mean S/N ratio is also not apples-to-apples because X-Sifter converts its S statistic to Gaussian sigma via the gamma-S relation, which the paper itself notes in Sections 2.2.3 and 4.2 is approximate and becomes inaccurate at high sigma. The authors should present the comparison at equal expected false detections per field, report the thresholds used by both pipelines, and provide uncertainties on the ratio (55/179 carries a binomial error of roughly +/-6%).
  2. [Section 4.3] The false-positive check does not validate the reality of the 55 extra sources in the single-observation fields. It uses five different fields with stacked CSC detections, where only 10 X-Sifter sources lack CSC counterparts, all attributed to transients or variability. This sample is much smaller and drawn from different data. To support the claim that the 55 extra sources are real, the authors should examine a random subset of those sources directly, for example by using any other observations of the same regions, or by injecting sources into pure simulated background images and running the full pipeline with the same threshold to estimate the empirical false-positive rate.
  3. [Section 3.1, step 4(c)] The background for a 16x16 sector is a single median value computed from a 4x4 grid of tiles after iterative outlier rejection. If the true background is spatially varying within a sector, or if the outlier rejection is biased by bright sources or flares, the single value of B enters both the PMF kernel (Equation 2) and the threshold (Equation 14), potentially miscalibrating the detection significance. The injection test in Section 4.1 uses the same background estimator and is therefore insensitive to this bias. The authors should test the pipeline on fields with structured background, or add a simulation with a known background gradient, to quantify the effect on completeness and false-alarm rate.
  4. [Section 4.1] The injection test is partially circular. The injected flux Fth is derived from Equation 14 using the same background estimate and PSF that the pipeline will use, so the recovered S/N peaking at the chosen sigma is largely by construction. This demonstrates numerical correctness but not absolute sensitivity. The external comparison to CSC2.1 is meant to provide the missing calibration, but the threshold-matching issues in the first major comment prevent it from doing so. A test with sources injected at random fluxes over a range (not only Fth) would provide a more meaningful completeness curve.
minor comments (7)
  1. [Abstract] The symbol '∼=' should be '≈' (an approximate equality sign, not a tilde-equals).
  2. [Section 2.2.3] The notation mixes 'Sthresh' and 'Sth'; please use one consistent symbol throughout.
  3. [Section 2.2.3] The phrase 'slope ranging from ∼ −1/2 and ∼ −3/2' should read 'between ∼ −1/2 and ∼ −3/2'.
  4. [Section 3.1, step 4(c)] The background estimation description should clarify that the final representative background is the median of the tile means, not a single global fit to the sector.
  5. [Section 3.4] 'X-swifter' is a typo for 'X-Sifter'.
  6. [Section 4.2] 'Theses two sources' should be 'These two sources'; the paper would also benefit from listing the ten OBSIDs used for the comparison, either in a table or in an appendix.
  7. [Section 4.3] 'X=Sifter' is a typo for 'X-Sifter'.

Circularity Check

1 steps flagged · score 2.0 of 10

Central sensitivity claim is externally benchmarked; only the injected-source validation is circular by construction.

  1. fitted input called prediction [Section 4.1 (simulation validation), using Eq. 14 from §2.2.1]
    "For a false-alarm probability γsim corresponding to a 4σ detection, and using the value of B measured in step (1) and the stamp P simulated in step (2), we calculate Fth by solving Equation 14; ... We generate a simulated image by injecting a source modeled as M(q − q0) = Poisson(B + FthP(q − q0)) at an arbitrary location q0 in the real image. ... We repeat this procedure 10,000 times and find that the distribution of signal-to-noise (S/N) levels peaks near 4σ, consistent with the γsim chosen."

    Fth is not an independent test flux: it is solved from Eq. 14, which was derived (Eqs. 8–12) by imposing Sth = E(S_H1(Fs)) at Fs = Fth, with Sth the γsim-percentile of P(S|H0). Injecting Poisson(B+FthP) and filtering with K(Fth) therefore gives E(S)=Sth by construction, which converts to γsim (≈4σ). The measured 'peak near 4σ' is thus the defining equation read backward; the test can only check numerical consistency (solver/FFT/interpolation), not validate the thresholding relation or the detection statistic.

full rationale

The central derivation is otherwise self-contained: the Neyman-Pearson lemma justifies the likelihood-ratio test; Eq. 1–3 define the PMF statistic; Eqs. 8–14 derive the thresholding relation from the detection-limit definition; and P(S|H0) is obtained by Monte Carlo over the stated Poisson model. The headline ~30% sensitivity gain is benchmarked against the external Chandra Source Catalog, not against X-Sifter's own fitted quantities, so that claim has independent content. The only circular step is the §4.1 injection validation, where Fth is chosen by solving Eq. 14 and therefore reproduces the chosen 4σ peak by construction; this is a minor, non-load-bearing self-consistency check. The paper's unspecific γ threshold in the CSC comparison is a correctness risk, but it is not circularity. Self-citation of Ofek & Zackay (2018a,b) is not flagged because the cited PMF derivation is a published mathematical result with stated assumptions and is not fitted to the present data.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The pipeline's sensitivity claims rest on the Poisson likelihood model, MARX PSF fidelity, per-sector constancy of background and PSF, energy-channel independence, and an exponential extrapolation for the gamma-S tail. The free parameters are mostly user-chosen settings, such as sector size, energy bands, background grid, false-alarm rate, and PSF crop fraction, plus the threshold flux Fth that fixes the kernel. No new physical entities are introduced.

free parameters (6)
  • Fth = determined by Equation 14 for each sector, background, and gamma
    The PMF kernel depends on the unknown source flux Fs; a fixed Fth approximates it and is set self-consistently via the thresholding relation. Sensitivity claims depend on this choice.
  • Sector size and buffer = 16 by 16 sectors with a 160 pixel buffer
    This partition assumes PSF and background are approximately constant within each sector; the size is a user-chosen parameter affecting sensitivity and runtime.
  • Energy channels = nc = 3 logarithmic channels by default
    The number and boundaries of energy channels affect how well energy-dependent PSF and background variations are captured and how S/N values combine.
  • Background tile grid = n = 4 tiles per side by default
    The background estimate is the median of tile means after iterative outlier rejection; the grid size and clipping thresholds are assumed to yield an unbiased sector background.
  • False-alarm probability gamma = 4 sigma used in injection tests; pipeline threshold configurable
    The detection threshold and the derived Fth and Sth all depend on the chosen false-alarm probability.
  • PSF crop fraction = 99 percent by default
    Cropping the PSF stamp to save runtime discards 1 percent of the PSF energy, a tradeoff between accuracy and speed.
assumptions (6)
  • domain assumption Pixel counts follow Poisson statistics with constant expected background B within a sector (Section 2.1, Equation 2).
    The entire PMF formalism and thresholding relation assume M(q) is Poisson with mean B plus source flux; real backgrounds and detector responses vary, so this is an idealization.
  • standard math Neyman-Pearson lemma establishes the likelihood-ratio test as most powerful at fixed false-alarm probability (Section 2.1).
    The optimality claim relies on Neyman-Pearson, but the alternative hypothesis H1 depends on the unknown source flux Fs, which the paper addresses by fixing Fth.
  • domain assumption MARX simulations accurately represent the Chandra PSF for all sectors and energy channels (Section 3.2.1).
    PSF stamps are generated with MARX; errors in MARX propagate directly into the filter kernel and source flux estimates.
  • domain assumption PSF and background are approximately constant within each 16 by 16 sector (Section 3.1, step 3).
    The pipeline filters each sector with one kernel; violations of this assumption cause loss of optimality near sector boundaries or in steep background gradients.
  • domain assumption Energy channels provide independent information, so S/N values combine in quadrature (Equation 16, Section 2.2.5).
    This requires the energy range in which PSF and background vary to be larger than the photon energy uncertainty; otherwise, channels are correlated.
  • ad hoc to paper The gamma-S relation can be extrapolated exponentially to low false-alarm probabilities (Section 2.2.3).
    The paper itself finds slopes varying from about -1/2 to -3/2 and non-exponential, step-like structures at low background, so a universal exponential extrapolation is an ad hoc modeling assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of X-Sifter: detecting transients in X-ray data using the optimal Poisson matched filter." pith.science (2026). https://pith.science/paper/NMXTK44K

@misc{pith2026241207858,
  author       = {Pith},
  title        = {Pith review of: X-Sifter: detecting transients in X-ray data using the optimal Poisson matched filter},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NMXTK44K}},
  note         = {Machine review of arXiv:2412.07858}
}
read the original abstract

We present X-sifter, a software package designed for near-optimal detection of sources in X-ray images and other forms of photon images in the Poisson-noise regime. The code is based on the Poisson-noise-matched filter (Ofek & Zackay), which provides an efficient method for calculating the delta log-likelihood function for source detection. The software accounts for several complexities inherent in real data, including variations in both the instrumental Point Spread Function (PSF) and background across the detector and as a function of energy. We validate the pipeline using real data with simulated source injections, as well as actual Chandra images. A comparison between the sources detected by our pipeline and those in the Chandra Source Catalog (CSC) suggests an approximate ~30% increase in the number of detected (real) sources. Near the detection limit, the reported S/N of our pipeline is approximately 1.3x higher than that of the CSC. This corresponds to a factor of 1.8 increase in survey speed.

Figures

Figures reproduced from arXiv: 2412.07858 by the authors.

Figure 1
Figure 1. The effect of choosing a constant flux Fth in the PMF, studied in 1-D. A source with a flux Fs is simulated and filtered with the approximated PMF Ker￾nel of Equation 6, with various values of Fth. The y-axis shows the completeness of the detection for each choice of Fth. Although the completeness is maximal when Fth = Fs, it is affected by less than 1% when Fth varies by an or￾der of magnitude. See also Ofek & Zack… view at source ↗
Figure 2
Figure 2. The thresholding relation is equivalent to [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The effect of various background values and various PSF on P(S | H0) and the γ − S relation. The top panel shows two PSF stamps, labeled PSF1 and PSF2, modeled with MARX at two different detector location of CCD0 of the ACIS Chandra camera. The middle panels show the γ)−S relation (left) and the P(S | H0) distribution computed with PSF1 (in red) and with PSF2 (in blue), for a background level bck = 1.13 × 10−5 count… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: For each background values, images Bsim where simulated (see §2.2.2) and used to compute P(S | H0) and derive Sth (see Equation 15). The bottom panel shows the maximum photon count value recorded in the Bsim images. All quantities were computed using the stamps PSF1 (r…
Figure 5
Figure 5. Figure 5: Description of all the steps of the X-Sifter algorithm. Yellow boxes designate actions, blue diamonds designate data outputs/inputs, red diamonds signal loops and the green boxes are for the input and output of the pipeline [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: shows an example of one of the library “sur￾rogate” PSF stamps after (1) modeling the PSF with MARX, (2) shifting the PSF to ensure it is centered, and (3) cropping an area that contains 99% of the PSF. The PSF library made available with the first release of X-Sifter …
Figure 7
Figure 7. Figure 7: Description of the simulations made to validate the pipeline [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: A comparison of the source detection by X-sifter and by the Chandra Source Catalog CSC2.1. Left panel: For sources detected in both catalogs, and in fields observed only once by Chandra (to ensure the sources were not detected in stacks), we show the S/N given by X-Sif…
Figure 9
Figure 9. Figure 9: The Search for False Positives. Left panel: an example of a sky region covered by seven observations, including OBSID = 21996. The cross and circle mark the specific area used for this analysis. Right panel: a region covered by six OBSIDs, including OBSID = 7187. In th…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 11 canonical work pages

  1. [1]

    2020, ApJ, 896, 39

    Alp, D., & Larsson, J. 2020, ApJ, 896, 39

    work page 2020

  2. [2]

    H., Lott, B., & The Fermi-LAT collaboration

    Ballet, J., Bruel, P., Burnett, T. H., Lott, B., & The Fermi-LAT collaboration. 2023, arXiv e-prints, arXiv:2307.12546

    arXiv 2023

  3. [3]

    1996, A&AS, 117, 393

    Bertin, E., & Arnouts, S. 1996, A&AS, 117, 393

    1996

  4. [4]

    1979, ApJ, 228, 939 De Luca, A., Salvaterra, R., Belfiore, A., et al

    Cash, W. 1979, ApJ, 228, 939 De Luca, A., Salvaterra, R., Belfiore, A., et al. 2021, A&A, 650, A167

    work page 1979

  5. [5]

    N., Primini, F

    Evans, I. N., Primini, F. A., Glotfelty, K. J., et al. 2010, ApJS, 189, 37

    2010

  6. [6]

    N., Evans, J

    Evans, I. N., Evans, J. D., Mart ´ ınez-Galarza, J. R., et al. 2024, arXiv e-prints, arXiv:2407.10799

    arXiv 2024

  7. [7]

    E., Kashyap, V., Rosner, R., & Lamb, D

    Freeman, P. E., Kashyap, V., Rosner, R., & Lamb, D. Q. 2002, ApJS, 138, 185

    work page 2002

  8. [8]

    1986, ApJ, 303, 336

    Gehrels, N. 1986, ApJ, 303, 336

    1986

Show all 18 references
  1. [9]

    1984, SAO Special Report, 393

    Schwarz, J. 1984, SAO Special Report, 393

    1984

  2. [10]

    M., Rutledge, R

    Law, N. M., Rutledge, R. E., & Kulkarni, S. R. 2004, MNRAS, 350, 1079

    2004

  3. [11]

    M., et al

    Macias, O., Gordon, C., Crocker, R. M., et al. 2018, Nature Astronomy, 2, 387

    2018

  4. [12]

    2012, MNRAS, 422, 1674

    Masias, M., Freixenet, J., Llad´ o, X., & Peracaula, M. 2012, MNRAS, 422, 1674

    2012

  5. [13]

    R., Bertsch, D

    Mattox, J. R., Bertsch, D. L., Chiang, J., et al. 1996, ApJ, 461, 396

    1996

  6. [14]

    Neyman, J., & Pearson, E. S. 1933, Philosophical Transactions of the Royal Society of London Series A, 231, 289

    1933

  7. [15]

    O., Shvartzvald, Y., Sharon, A., et al

    Ofek, E. O., Shvartzvald, Y., Sharon, A., et al. 2023, PASP, 135, 124502

    2023

  8. [16]

    Stetson, P. B. 1987, PASP, 99, 191

    1987

  9. [17]

    Stewart, I. M. 2006, A&A, 454, 997

    2006

  10. [18]

    Zackay, B., & Ofek, E. O. 2017, ApJ, 836, 187

    2017

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.