{"id":"95616c08-93e1-42dd-90c6-a1f2aaab670b","arxiv_id":"2504.20723","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Counts-in-cells of LoTSS-DR2 radio sources above 2 mJy are best described by a negative binomial distribution, and the variance scaling of those counts recovers the angular two-point correlation function measured directly.","lead":"Using the LOFAR radio survey's second data release, this paper studies how many radio sources fall in each patch of sky and tests which statistical model best describes those counts. It finds that a negative binomial distribution fits best and that counting sources in cells offers a cheap way to measure how clustered radio galaxies are.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Neglected angular clustering in Eq. (6) likely inflates the NB-vs-CP evidence and biases the inferred 1.27 component count.","rationale":"The paper is a careful analysis with correct moment derivations and a genuine model comparison; the Landy-Szalay cross-check and the value-added catalogue comparison provide independent support. The concern is not that the NB fits poorly (the KS test already rejects it for mask d), but that the strength of the preference and the physical interpretation rest on a generative model that omits the angular clustering the paper itself measures. This is an internal inconsistency rather than a disagreement with consensus: Eq. (6) has no clustering term, while Sect. 6.2 demonstrates clustering on the same scales. A reanalysis on decorrelated cells is the minimal check that would settle whether the Bayes factor and the inferred component count survive. The reader's conditional verdict already flags model misfit; this stress-test reinforces that condition without moving the verdict.","tokens_in":31011,"tokens_out":8241,"duration_ms":96493,"concrete_test":"Re-run the distribution analysis on a decorrelated subset of cells with centers separated by more than 2 degrees (so w(θ) is negligible), using the same mask d and S > 2 mJy. Recompute the χ²/dof, KS statistic with Monte Carlo critical values, the NB-vs-CP Bayes factor, and the method-of-moments estimate of p and the mean component count. If the Bayes factor falls below 10 or the inferred mean component count shifts by more than 0.1 relative to 1.27, the strong-evidence claim and the multi-component interpretation are not robust to the neglected angular clustering. This test reuses the existing public catalogue and mask, so it can be done without new observations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The generative model in Eq. (6) assumes O_i ~ Poisson(λ), i.e. no spatial correlation between physical objects. The same catalogue, however, shows significant angular clustering over the same scales: Sect. 6.2 measures w(θ) via Landy-Szalay and fits γ ≈ 2.08 at 2 mJy over 0.11–3.66 deg. A Poisson object field cannot produce this clustering. The negative binomial fit is performed on the one-point counts, and the entire excess variance above Poisson is attributed to the logarithmic component-count distribution, giving p = 0.37 and a mean component count of 1.27 at 2 mJy. If part of that excess variance comes from object clustering, p and the inferred component count are biased. Moreover, the Bayes factor of 25.4 (NB vs CP) is computed as a product over roughly 84,600 cells that are spatially correlated, so the effective sample size is smaller than assumed and the 'strong evidence' is likely overstated. The paper itself labels the component distribution 'an educated guess' (Sect. 2.2), so the physical interpretation is the least secure part of the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes the counts-in-cells statistics of the LoTSS-DR2 radio source catalogue (about 4.4 million sources over 5,635 square degrees) and compares three models for the counts: Poisson, compound Poisson, and negative binomial. Using a Cox-process construction, the authors argue that the negative binomial distribution is strongly preferred above a 2 mJy flux density threshold, with a fitted mean of 1.27 radio components per physical source. They also use the scaling of the normalised variance of counts with cell size (the reduced normalised variance Ψ2) to fit a power-law angular two-point correlation function, obtaining an exponent 1−γ between about −0.8 and −1.05 at 2 mJy and SNR 7.5, and they compare this with a direct Landy-Szalay estimate. The paper explicitly acknowledges that the distribution of the number of components per object is an 'educated guess' (Sect. 2.2) and that the KS test still rejects the negative binomial for the default mask (Sect. 6.1, Table 4).","tokens_in":31209,"tokens_out":4286,"duration_ms":42534,"significance":"If the central claims hold, the paper would (i) establish that a one-parameter logarithmic component-count distribution explains the counts-in-cells of a large radio survey, (ii) provide a physical interpretation of the overdispersion in terms of a mean component count of about 1.27, and (iii) validate a computationally cheap counts-in-cells route to the angular two-point correlation function that is linear in the number of sources. The paper's strengths include the use of a large, public dataset; clear derivations of the moments and generating functions for the three models; and cross-checks against the value-added catalogue and against a Landy-Szalay estimator on the same survey. The falsifiable predictions (e.g., the scaling exponent and the comparison with the value-added catalogue) are useful. However, the central model premise is an ad hoc parametric choice, and the statistical evidence is partly overstated given that the KS test rejects the preferred model and the power-law fits have large reduced chi-square values.","major_comments":[{"comment":"The generative model in Eq. (6) treats the number of physical objects O_i as Poisson with no spatial correlation, yet the same catalogue exhibits strong angular clustering measured in Sect. 6.2 with w(θ) fitted to a power law. Since p is estimated from the total variance (Eq. (23)), any variance contributed by object clustering is absorbed into the logarithmic component-count distribution, biasing the inferred p and the mean component count of 1.27 (Eq. (20)). The Bayes factor of 25.4 reported in Sect. 6.1 is computed as a product over roughly 84,625 cells that are spatially correlated, so the effective sample size is smaller than assumed and the 'strong evidence' is likely overstated. A concrete check would be to compare the variance implied by the fitted logarithmic distribution with the component-count variance estimated from the value-added catalogue, or to include a clustered object field (e.g., a Poisson cluster process) in the model; until such a test is performed, the physical interpretation of p as a component-count parameter is not secure.","section":"Sect. 2.2, Eq. (6)"},{"comment":"The KS test in Table 4 rejects the negative binomial distribution at all flux density thresholds for mask d (dn=0.0078 vs dα=0.0030 at 2 mJy), and the text in Sect. 7 acknowledges this, yet the abstract and conclusions state there is 'strong evidence in favour of the negative binomial distribution' and the text in Sect. 6.1 describes it as an 'excellent fit'. A model that is formally rejected at 99% confidence cannot simultaneously be claimed as the strongly favoured model without a careful statement of the approximate nature of the fit; the Bayes factor comparison should be reported with an effective sample size or as conditional on the assumed model family.","section":"Sect. 6.1, Table 4"},{"comment":"The power-law fits to Ψ2 yield reduced chi-square values as high as 8.4 (2 mJy, Nside 16-512) and 3.4 (2 mJy, Nside 16-256), well above unity, indicating that the single power-law ansatz is formally unacceptable for these data. The abstract's claim that the scaling is 'in good agreement with a power-law model' is not supported by these statistics; the authors should either provide a justification for the extra variance (e.g., cosmic variance, mask systematics, non-linear clustering), or restrict the quoted exponent range to angular scales where the fit is acceptable, or revise the claim.","section":"Sect. 6.2, Table 5"}],"minor_comments":[{"comment":"The comparison of the fitted mean component count (1.27) with the value-added catalogue value (1.13) is made at 2 mJy, while Sect. 5 states that the value-adding process is only complete above 4 mJy; the text should explicitly state that this comparison is an extrapolation and note the associated uncertainty.","section":"Sect. 7.1"},{"comment":"The reference list contains inconsistent entries: the in-text citation 'de Gasperin et al. 2023' in the Fig. 4 caption does not match the listed reference 'de Gasperin et al. 2021', and Shimwell et al. (2022) appears twice with slightly different formatting.","section":"References and Fig. 4 caption"},{"comment":"The notation for the central moments m3 and m4 is introduced in Eq. (2) but used later in Eq. (31)-(32) without reminding the reader that these are sample estimates; a brief restatement would improve readability.","section":"Sect. 2.1 and Sect. 6.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid application of standard techniques to a valuable public dataset, and its computational-efficiency claim for the counts-in-cells approach is well motivated. The main concern is that the central physical interpretation (the mean component count and the NB preference) rests on an ad hoc single-parameter model that is formally rejected by the authors' own KS test, and the neglect of angular clustering in the generative model is likely to bias the inferred parameters. I would like to see the authors either address the clustering bias directly or substantially soften the physical claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Lukas,\n\nQuick take on the LoTSS-DR2 counts-in-cells paper. The variance-scaling route to w(θ) is validated against the Landy-Szalay estimator, and the mean component count of 1.27 checks out against the value-added catalogue. That external consistency is the best part of the paper, and it should count in its favor.\n\nWhat's new is the claim that a negative binomial beats a compound Poisson for the one-point counts above 2 mJy, with a Bayes factor of 25.4. But the generative model in Eq. (6) assumes the physical objects are Poisson-distributed, meaning no angular clustering, and then assigns the entire excess variance to the component-count distribution. The same paper measures significant w(θ) over the same scales, with γ≈2.08 at 2 mJy. A Poisson object field cannot produce that clustering. So the excess variance is a mix of component multiplicity and clustering, and the fitted p and the 1.27 component count are biased by this misspecification. The paper itself labels the component distribution 'an educated guess,' so the physical interpretation is the least secure part of the argument.\n\nThe KS test also still rejects the NB for the default mask (d_n=0.0078 vs d_α=0.0030 at 2 mJy), so the 'excellent fit' wording is too strong. The Bayes factor is computed over roughly 85,000 spatially correlated cells, so the effective sample size is smaller than assumed and the 'strong evidence' is overstated. The power-law fits to Ψ2 have reduced chi-square up to 8.4, so those good-agreement claims are shaky too.\n\nNone of this is fatal. The method for getting w(θ) from variance scaling is genuinely cheap and matches the direct estimator, and the paper is careful about masking and completeness. It deserves a serious referee. But the authors should either fit a clustered object field or substantially soften the causal claim about multi-component sources. The central comparative claim—NB is preferred over CP—is probably right, but the physical takeaway as written goes beyond what the model actually tests.\n\nI'd send it to review with the expectation of major revision, mainly on the interpretation. If they fix the model or reframe the conclusion, this becomes a solid paper. Good reading for anyone working on radio surveys or counts-in-cells methods.\n\nBest","headline":"A solid counts-in-cells analysis of LoTSS-DR2 that overreaches when it attributes the full variance excess to multi-component sources, since the same data show angular clustering that is ignored in the generative model.","tokens_in":31838,"tokens_out":5544,"would_cite":false,"duration_ms":51971,"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":"Radio sources in LOFAR's sky survey are not Poisson-distributed: above 2 mJy their counts-in-cells follow a negative binomial distribution, whose cell-size scaling recovers the two-point correlation function cheaply.","keywords":["counts-in-cells statistics","negative binomial distribution","Cox process","two-point correlation function","LOFAR","LoTSS-DR2","large-scale structure"],"falsifier":"The empirical component-count distribution from the LoTSS-DR2 value-added catalogue, which is complete above 4 mJy, can be compared directly with the logarithmic law the negative binomial model assumes: if the measured distribution of components per physical object is not logarithmic, or if its fitted parameter changes with flux density, sky position, or source morphology, then the reported preference for the negative binomial is an artifact of the assumed family rather than a property of the radio-source population. A second check uses the model's own moment predictions: the negative binomial fixes skewness as $g_1 = (2n_c - 1)/(\\mu^{1/2} n_c^{1/2})$, so comparing the empirical third and fourth moments of the counts at 2 mJy with these predictions would falsify the model if they disagree within the quoted uncertainties.","tokens_in":30734,"feed_emoji":"📡","tokens_out":14403,"duration_ms":115372,"temperature":0.7,"pith_summary":"This paper tries to establish which probability distribution governs the number of radio sources in equal-sized cells of sky, using the 4.4-million-source LOFAR Two-Metre Sky Survey Data Release 2. Above a 2 mJy flux-density threshold, the counts are shown to follow a negative binomial distribution far better than a Poisson or compound Poisson distribution, and the paper traces that shape to the multi-component nature of radio objects: cores, lobes, and resolved spirals are each counted as several catalogue entries. Viewed as a function of cell size, the same variance statistic recovers the angular two-point correlation function with an exponent $1-\\gamma$ between $-1.05$ and $-0.8$, in agreement with direct pair-counting measurements. The counts-in-cells route estimates clustering in time proportional to the number of sources rather than the number of pairs, an advantage for the much larger surveys now being planned.","feed_headline":"LOFAR's radio-source counts are negative binomial, not Poisson","feed_subtitle":"The cell-count variance of 827,000 sources reproduces the two-point correlation function at linear cost.","key_machinery":"The central object is the Cox process $N_i = \\sum_{j=1}^{O_i} C_{ji}$, in which the number of physical objects $O_i$ in a cell is Poisson with intensity $\\lambda$ and each object contributes $C_{ji}$ catalogue entries. Choosing $C_{ji}$ to follow a logarithmic distribution with parameter $p$ — the 'educated guess' of Section 2.2 — makes the generating function of $N_i$ collapse exactly to that of a negative binomial distribution with $r = -\\lambda/\\ln(1-p)$, so one parameter $p$ carries the whole overdispersion. The clustering measurement then runs through the reduced normalised variance $\\Psi_2$, which for a power-law angular correlation $w(\\vartheta) = A_0(\\vartheta/\\vartheta_0)^{1-\\gamma}$ evaluates to $\\Psi_2(\\Theta) = A_0 C_\\gamma (\\Theta/\\Theta_0)^{1-\\gamma}$, with a numerically computed coefficient $C_\\gamma$ of order unity.","core_discovery":"The central claim is that the counts-in-cells distribution of LoTSS-DR2 radio sources above a 2 mJy flux-density threshold is strongly better described by a negative binomial distribution than by a Poisson or compound Poisson distribution. The argument builds on a Cox process: physical objects are distributed as a Poisson process, and each object contributes a random number of detected radio components drawn from a logarithmic distribution; composing those generating functions yields exactly the negative binomial family, with a fitted mean of 1.27 components per source at 2 mJy. The paper supports this preference with a reduced chi-square test, a Monte-Carlo-calibrated Kolmogorov-Smirnov test, and Bayes factors of 25.4, 17.2, and 15.9 at 2, 4, and 8 mJy, and it notes that even the preferred model is formally rejected by the KS test for the default mask, indicating residual effects beyond the model. The paper further claims that the scaling of the reduced normalised variance of counts with cell size follows a single power law whose exponent $1-\\gamma$ lies between $-1.05$ and $-0.8$ at 2 mJy and SNR 7.5, consistent with the angular two-point correlation function measured directly with the Landy-Szalay estimator and with earlier optical, infrared, and radio surveys.","pith_inferences":["If the logarithmic component-count law is universal, the fitted parameter $p$ becomes a portable descriptor of how radio morphology maps into catalogue entries; measuring it as a function of flux density and source type across surveys would test whether the law is truly constant or itself a function of the source population.","Because the variance-based and Landy-Szalay exponents agree, the cheap variance route could double as a systematics check for pair-counting pipelines: a disagreement between the two at a given angular scale would point to masking or completeness problems rather than to cosmology.","The paper's admitted 'educated guess' on the component-count distribution is directly testable: the value-added catalogue, complete above 4 mJy, provides the empirical component-count distribution, so one can check whether it is really logarithmic or a mixture that varies with morphology — the Bayes factor in favour of the negative binomial depends on this.","Below roughly 0.1 degrees, multi-component associations dominate the variance, so the usable angular range of the method is set by how well the component law is modelled; pushing to sub-mJy thresholds would require modelling the star-forming galaxy population separately."],"forward_implications":["The fitted negative binomial mean of 1.27 components per source at 2 mJy is consistent with independent component-association counts (the value-added catalogue gives 1.13 above 4 mJy; a LOFAR survey cross-match gives about 1.33), so the model's parameter has a direct physical reading.","The variance-of-counts method estimates the two-point correlation function with linear rather than quadratic scaling in the number of sources, making it a practical clustering probe for future surveys with tens of millions of detections.","Higher moments of the same counts extend the machinery to higher-order correlation functions at the same linear cost, opening a route to non-Gaussianity tests on large angular scales.","The fitted slope steepens with flux-density threshold and flattens with higher signal-to-noise cuts, which the paper reads as a population effect: AGN dominance above 2 mJy raises clustering, while brighter, more isolated sources cluster less."],"supporting_citations":[{"why":"The LoTSS-DR1 counts-in-cells analysis this work extends; it established that radio counts are non-Poissonian and that a compound Poisson model describes them well.","marker":"Siewert et al. 2020"},{"why":"Supplies the random mock catalogue, the 'mask d' geometry, and the Landy-Szalay two-point correlation measurements used as the direct benchmark.","marker":"Hale et al. 2024"},{"why":"The LoTSS-DR2 data release: the source catalogue itself and the injection-based completeness estimates that justify the 2 mJy threshold.","marker":"Shimwell et al. 2022"},{"why":"Provides the generating functions for the Poisson, logarithmic, and negative binomial distributions that carry the Cox-process derivation.","marker":"Johnson et al. 2005"},{"why":"The Cox process construction: a Poisson-distributed number of objects each contributing a random number of components.","marker":"Cox 1955"},{"why":"Establishes the link between the variance of counts-in-cells and the angular two-point correlation function that the power-law fit relies on.","marker":"Totsuji & Kihara 1969"},{"why":"The LoTSS-DR2 value-added catalogue whose mean component count of 1.13 above 4 mJy is compared with the negative binomial mean of 1.27.","marker":"Hardcastle et al. 2023"},{"why":"The direct two-point correlation estimator that validates the counts-in-cells power-law result.","marker":"Landy & Szalay 1993"},{"why":"An independent estimate of the mean association component count of roughly 1.33 that corroborates the negative binomial mean number of components.","marker":"Böhme et al. 2023"}],"fun_headline_variants":["LOFAR radio sources cluster: counts defy Poisson","Radio sky isn't random: LOFAR shows negative binomial","Negative binomial wins for LOFAR source counts","Counts-in-cells: LOFAR sources follow negative binomial"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every physical radio object generates a number of catalogue components drawn from one logarithmic distribution with a single parameter $p$ that is the same across the whole survey and across flux densities; the paper introduces this as 'an educated guess'.","fun_headline_variants_meta":{"raw":{"variants":["LOFAR radio sources cluster: counts defy Poisson","Radio sky isn't random: LOFAR shows negative binomial","Negative binomial wins for LOFAR source counts","Counts-in-cells: LOFAR sources follow negative binomial"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000296,"raw_usage":{"total_tokens":1827,"prompt_tokens":1163,"completion_tokens":664,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":779,"completion_tokens_details":{"reasoning_tokens":598}},"tokens_in":779,"tokens_out":664,"duration_ms":6549,"temperature":1.0,"reasoning_tokens":598,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:21:47.651184+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The empirical component-count distribution from the LoTSS-DR2 value-added catalogue, which is complete above 4 mJy, can be compared directly with the logarithmic law the negative binomial model assumes: if the measured distribution of components per physical object is not logarithmic, or if its fitted parameter changes with flux density, sky position, or source morphology, then the reported preference for the negative binomial is an artifact of the assumed family rather than a property of the radio-source population. A second check uses the model's own moment predictions: the negative binomial fixes skewness as $g_1 = (2n_c - 1)/(\\mu^{1/2} n_c^{1/2})$, so comparing the empirical third and fourth moments of the counts at 2 mJy with these predictions would falsify the model if they disagree within the quoted uncertainties.","supporting_citations":[{"cited_title":"M., Hale , C., Bhardwaj , N., et al","cited_arxiv_id":null,"evidence_quote":"The LoTSS-DR1 counts-in-cells analysis this work extends; it established that radio counts are non-Poissonian and that a compound Poisson model describes them well."},{"cited_title":"L., Schwarz , D","cited_arxiv_id":null,"evidence_quote":"Supplies the random mock catalogue, the 'mask d' geometry, and the Landy-Szalay two-point correlation measurements used as the direct benchmark."},{"cited_title":"2005, Univariate Discrete Distributions, Wiley Series in Probability and Statistics (Wiley)","cited_arxiv_id":null,"evidence_quote":"Provides the generating functions for the Poisson, logarithmic, and negative binomial distributions that carry the Cox-process derivation."}],"review_version":1}