{"id":"55fbee4e-282f-4a9d-b579-23e880beb459","arxiv_id":"2505.13267","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Stacked radio spectra of XMM-LSS star-forming galaxies show no spectral index evolution from z=0.1 to 3.0, tentative low-frequency steepening, and a thermal fraction of 11 to 18 percent at 1.5 GHz.","lead":"By stacking radio images at eleven frequencies from 144 to 1500 MHz, this paper measures the average radio spectrum of about 129,000 star-forming galaxies in the XMM-LSS field across seven redshift bins from z=0.1 to 3.0. The average spectrum does not change much with redshift, appears to steepen at low frequencies, and has a thermal emission fraction of about 11 to 18 percent at 1.5 GHz.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed low-frequency steepening (q>0, steep alpha_low) may be an artifact of the coarser 144 MHz LOFAR beam: the boosting/null-stack tests in Appendix B2 and Section 3.3 do not account for source clustering or blending around real Ks-selected galaxies.","rationale":"The reader's weakest assumption is that the median-stacked flux densities are comparable across bands with different resolutions and noise after the boosting correction. I agree that this is the most load-bearing point, and it directly threatens the abstract's headline claim of consistent low-frequency steepening and q>0 in all bins. The paper does perform a careful null-stack test and a boosting simulation, which are genuine strengths, but neither test addresses astrophysical clustering and blending around the actual Ks-selected SFG positions. The residual maps used for injection in Appendix B2 are cleaned of detected sources, so the simulation cannot measure the effect of unresolved neighboring sources at the coarse 144 MHz resolution. A matched-resolution restacking experiment is the natural way to settle this. I also note that the two-point spectral indices in Table 5 do not show uniform low-frequency steepening across bins (z4, z6), so the abstract's phrasing is stronger than the data support even before considering systematics. However, this is an interpretive overstatement rather than a fatal flaw; the underlying stacking analysis is appropriate and the paper transparently lists the need for higher-resolution data. Therefore, the reader's CONDITIONAL verdict remains appropriate, and I do not recommend changing it.","tokens_in":34993,"tokens_out":6345,"duration_ms":66760,"concrete_test":"Convolve the uGMRT 320/370/420/460 MHz and VLA 1500 MHz images to the 7.5x8.5 arcsec LOFAR beam, re-stack the same SFG positions with the identical PASTA pipeline, and recompute alpha_144-320 and the CPL q. If the smoothed 320 MHz median flux rises so that alpha_144-320 becomes consistent with -0.8+-0.1 (and q drops to ~0), the low-frequency steepening is an angular-resolution artifact. As a complementary check, re-stack 144 MHz using only SFGs with no 144 MHz catalog source within 30 arcsec; if the low-frequency excess disappears, blending by neighboring sources is the cause.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central positive claim is that the median-stacked SEDs steepen at low frequency, with CPL curvature q>0 in all seven redshift bins (abstract, Section 5.0.3, Table 5). This conclusion rests heavily on the 144 MHz flux densities in Table 4. For example, in z5, S_144=24.25 uJy/beam while S_320=9.75 uJy/beam, giving alpha_low=-1.14; a canonical -0.8 spectrum extrapolated from 320 MHz predicts only ~18 uJy at 144 MHz. The 144 MHz map has a 7.5x8.5 arcsec beam and median RMS 400 uJy/beam, versus 3-8 arcsec and 10-133 uJy for the other bands (Table 1). The flux-boosting correction (Appendix B2, Figure B3) is calibrated by injecting unresolved Gaussian mock sources into residual maps; this cannot reproduce blending from real clustered SFGs or extended low-frequency emission. The null-stack test in Section 3.3 stacks random sky positions, which do not match the large-scale-structure environment of Ks-selected galaxies. If residual blending inflates the 144 MHz flux by even ~30%, it would bias alpha_low steep and drive q positive in exactly the manner claimed. Additionally, the overstatement is visible in Table 5: z4 has alpha_low=-0.79 and z6 has alpha_low=-0.90, so the 'consistent steepening across all redshift bins' is not actually seen in the two-point indices; it is largely an outcome of the model fits. The paper's own abstract concedes that higher-resolution observations are needed to deal with source blending and confusion noise, underscoring that this is the load-bearing uncertainty.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper uses median stacking of Ks-selected star-forming galaxies (SFGs) in the XMM-LSS field, in seven redshift bins from z=0.1 to z=3.0, across observer-frame radio frequencies 144-1500 MHz from LOFAR, uGMRT, MIGHTEE/MeerKAT, and VLA data. After removing quiescent galaxies and AGN candidates via UVJ colors and Donley et al. (2012) selection, the authors measure stacked median flux densities, correct them for median flux boosting, and fit power law, curved power law, and synchrotron-plus-free-free (double power law) models. They report no significant correlation between radio spectral index and redshift, a consistent low-frequency steepening with positive curvature q in all redshift bins, and a thermal free-free fraction at 1500 MHz of about 11-18 per cent.","tokens_in":35371,"tokens_out":5083,"duration_ms":49356,"significance":"If the results hold, the paper provides one of the few direct, below-threshold constraints on the average radio SED shape of faint star-forming galaxies at 144-1500 MHz, including the physically interesting claim of convex (low-frequency steepened) spectra and a roughly constant thermal fraction over cosmic time. The study is strengthened by the use of a public stacking code, null-stack tests, flux-boosting simulations with realistic noise, photometric-redshift quality metrics, and explicit AICc/BIC model comparisons. The data-driven measurement of a lack of alpha-z correlation is a useful negative result. However, the central positive claim about low-frequency steepening rests on one band (144 MHz) with a coarser beam and much higher noise than the other bands, and the quoted statistical significances in Table 5 are weaker than the abstract suggests.","major_comments":[{"comment":"The central claim of low-frequency steepening depends almost entirely on the 144 MHz stacked flux densities (e.g., S144=24.25 uJy/beam in z5 versus 9.75 uJy/beam at 320 MHz), but the flux-boosting correction in Appendix B2 is calibrated by injecting unresolved, unclustered mock sources into residual maps. This cannot reproduce the effect of blending around real Ks-selected galaxies in the 7.5x8.5 arcsec LOFAR beam, nor extended low-frequency emission, and the null-stack test in Section 3.3 uses random sky positions that do not match the large-scale-structure environment of the galaxies. Given that the 144 MHz median RMS is 400 uJy/beam while the stacked signals are only 14-28 uJy/beam, a modest unresolved blending residual could easily mimic the reported steep alpha_low values. A concrete remedy would be to inject mock sources at the actual positions of the Ks-selected SFGs (with realistic clustering and source sizes) into the 144 MHz maps, or to repeat the analysis with the 144 MHz data convolved to a common, coarser resolution and test sensitivity to the assumed source size distribution.","section":"Section 3.3, Appendix B2, Table 4"},{"comment":"The statement that the data show a 'consistent steepening of the low-frequency spectral index across all redshift bins' is not supported by the two-point spectral indices in Table 5: alpha_low = -0.79+/-0.27 (z4) and -0.90+/-0.41 (z6) are statistically consistent with the canonical -0.8 synchrotron value, and only z1, z3, z5, z7 are more than 1 sigma steeper. Likewise, the claim that q>0 in all redshift bins is based on best-fit values, but the CPL curvature is consistent with q=0 at less than 2 sigma in five of seven bins (e.g., q=0.04+/-0.15 in z2, q=0.06+/-0.07 in z3, q=0.08+/-0.07 in z4, q=0.11+/-0.08 in z5, q=0.15+/-0.11 in z6); only z1 (0.22+/-0.07) and z7 (0.35+/-0.12) individually require positive curvature. The abstract and conclusions should be reworded to report this more limited evidence, and the fitted DPL alpha_NT values (ranging from -0.80+/-0.50 to -1.96+/-0.41) should be discussed with their uncertainties rather than as a uniform steepening.","section":"Table 5, Section 5.0.3, Figure 10"},{"comment":"The thermal fractions reported in Figure 12 are not independent measurements but are derived from the same fitted models: Eq. (10) takes the fitted non-thermal spectral index as input, and for the DPL model the thermal fraction is essentially the ratio of the fitted normalizations A/(A+B) at the reference frequency. Presenting these values as a separate 'result' therefore adds little beyond the model fits, and the 11-18 per cent range inherits all the assumptions of the chosen models, including the fixed free-free index alpha_FF=-0.1. The discussion should clearly label these as model-dependent estimates, not as direct measurements of the thermal component.","section":"Section 5.0.4, Eq. (10)"}],"minor_comments":[{"comment":"The text says 'a 10 per cent contribution at 1400 GHz' but should read 'at 1.4 GHz' (or 1400 MHz).","section":"Section 5.0.4"},{"comment":"The Table 4 column heading 'S1280' is inconsistent with the text, which gives the MIGHTEE frequency as 1284 MHz; use one value consistently.","section":"Table 4"},{"comment":"The sentence 'Given that we do not see evidence of a curvature in the stacked spectra... we fit a generic curved power law model' is internally contradictory; either rephrase to state that a curvature term is included to test for deviations, or drop the introductory clause.","section":"Section 4.2.2"},{"comment":"The claim that the surveys 'have comparable angular resolution and similar rms noise sensitivity' is difficult to reconcile with Table 1, which shows beam sizes from 2.78 to 8.5 arcsec and median RMS values from 10 to 400 uJy/beam; please clarify what comparison is intended.","section":"Section 2"},{"comment":"The AICc and DeltaBIC table maps are shown without uncertainties; given that the reported differences between models are often smaller than 2, some indication of the stability of these values (e.g., via bootstrap) would help the reader judge the model-comparison claims.","section":"Figure 11"},{"comment":"There is a typo in 'e estimate the stacked background noise' (should be 'we estimate').","section":"Section 3.3"}],"recommendation":"major_revision","confidential_remarks":"This is a solid observational stacking paper with good methodological hygiene, and the no-evolution result is likely robust. My main concern is that the abstract and conclusions overstate the statistical significance of the low-frequency steepening, which is driven largely by the 144 MHz band. The 144 MHz blending/confusion issue is a genuine systematic that needs a more realistic simulation or a matched-resolution cross-check before the central claim can be published. I would encourage the editor to ask for a revised version that either adds such a test or substantially softens the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the headline. This is a careful stacking analysis of ~129,000 Ks-selected SFGs in XMM-LSS, using LOFAR 144 MHz, uGMRT sub-bands, MeerKAT 1.28 GHz and VLA 1.5 GHz. The new thing is the field and the sub-band images; the method is established. The paper does several things well: the sample construction is thorough, the null stacks are there, and they simulate flux boosting rather than ignoring it. The null result — no significant evolution of radio spectral index with redshift — is probably right and agrees with earlier work.\n\nThe soft spot is the claim that the SEDs consistently steepen at low frequency. The two-point α_low values in Table 5 are −1.26, −1.29, −0.79, −1.14, −0.90, −1.91; two of the six bins are consistent with the canonical −0.8, and the z2 bin has no 144 MHz data at all. The CPL curvature q is positive in all bins, but most values are within 2σ of zero; only z1 (0.22±0.07) and z7 (0.35±0.12) really stand out. So the abstract overstates what the data show.\n\nThe stress-test concern about the 144 MHz beam is legitimate and I take it seriously. That band has a ~8 arcsec beam and 400 µJy/beam noise, while the other bands are finer and much deeper. The boosting correction is calibrated with unresolved mock sources injected into residual maps; that will not reproduce blended or extended emission around real galaxies in clusters. The null-stack test on random sky positions does not rule this out. A 30% inflation of the 144 MHz flux would plausibly produce the steep α_low and positive q they report. So the low-frequency steepening should be treated as tentative until higher-resolution data confirm it.\n\nThe thermal fractions (11–18%) come from the fitted α via Eq. 10, so they are a restatement of the model, not an independent measurement. That is standard, but it should be labelled as such. Data are only available on request; that is a minor issue.\n\nBottom line: this is a solid reference measurement for XMM-LSS and the no-evolution result is useful. The positive claim needs to be toned down or defended with matched-beam tests. I would send it to peer review with a request for heavy revision, not desk-reject it.","headline":"Careful XMM-LSS stacking that delivers a likely-robust null result (no α–z evolution), but the claimed low-frequency steepening/q>0 is overstated relative to the error bars.","tokens_in":35977,"tokens_out":4294,"would_cite":false,"duration_ms":37859,"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":"Median-stacked radio spectra of 128,967 star-forming galaxies in the XMM-LSS field steepen toward low frequencies in every redshift bin from z = 0.1 to z = 3.0, with no spectral-index evolution.","keywords":["radio continuum","star-forming galaxies","median stacking","radio spectral index","XMM-LSS field","spectral energy distribution","synchrotron emission","redshift evolution"],"falsifier":"Re-derive the stacked SED from 144 MHz images restored to the 4.5-arcsec resolution of the VLA 1500 MHz map, recomputing the flux-boosting correction by injecting mock sources into the convolved maps; if the corrected 144 MHz median flux density drops by more than its measurement uncertainty (the 144 MHz stacks have signal-to-noise ratios of roughly 4-7), the low-frequency steepening and the $q>0$ curvature are resolution artifacts rather than properties of the galaxies. A complementary check: stack only the sources individually detected at $\\geq 5\\sigma$ at 1284 MHz and ask whether the convex shape survives in a radio-selected subsample.","tokens_in":34765,"feed_emoji":"📡","tokens_out":30771,"duration_ms":251422,"temperature":0.7,"pith_summary":"Most star-forming galaxies in deep radio surveys are too faint to detect individually, so their average radio spectrum must be reconstructed statistically. This paper median-stacks 128,967 $K_S$-selected star-forming galaxies in the XMM-LSS field in seven redshift bins from $z = 0.1$ to $z = 3.0$ and measures the average flux density at eleven frequencies between 144 and 1500 MHz using LOFAR, uGMRT, MeerKAT, and VLA images. It claims the median spectrum is convex — steeper at low frequencies, with a low-frequency spectral index $\\alpha_{\\rm low}$ (the slope, $S\\propto\\nu^\\alpha$) around $-0.8$ to $-1.9$ — and flatter at high frequencies, with a positive curvature parameter $q$ in every redshift bin. The paper also reports no correlation between spectral index and redshift, and a thermal free-free fraction of 11-18 percent at 1500 MHz. The shape of the spectrum is exactly what k-corrections assume when converting observed radio fluxes into rest-frame luminosities and star-formation rates at high redshift, so a convex, non-evolving average spectrum changes those corrections, and it runs against recent claims that low-frequency spectra flatten.","feed_headline":"Galaxy radio spectra steepen, not flatten, at low frequencies","feed_subtitle":"The shape holds from z = 0.1 to z = 3 across 128,967 stacked galaxies, contradicting recent flattening reports.","key_machinery":"The carrying object is the median-stacked radio SED, built by cutting $36\\times36$ arcsec thumbnails from each radio map at the near-infrared positions of the selected galaxies and co-adding them with a median rather than a mean, which suppresses interlopers and image artifacts. Because the stacked median is boosted relative to the true source median when map noise is comparable to source flux, a correction factor $F_{\\rm boost} = S_{\\rm recovered}/S_{\\rm input}$ is derived by injecting 5,000 mock sources into residual maps, re-stacking them, and comparing input to output medians; this factor is applied to every reported flux density. Three observed-frame models are then fitted to the corrected SEDs: a power law $S_\\nu = S_0(\\nu/\\nu_0)^{\\alpha}$; a curved power law $S_\\nu = S_0(\\nu/\\nu_0)^{\\alpha}\\exp\\!\\big(q\\,\\ln^2(\\nu/\\nu_0)\\big)$, whose sign of $q$ decides concave versus convex; and a double power law $S_\\nu = A(\\nu/\\nu_0)^{-0.1} + B(\\nu/\\nu_0)^{\\alpha_{\\rm NT}}$ summing free-free and synchrotron emission. The curvature term $q$, together with the two-point indices $\\alpha_{\\rm low}$, $\\alpha_{\\rm mid}$, and $\\alpha_{\\rm high}$, is the diagnostic that carries the central claim.","core_discovery":"This paper claims that the average radio continuum spectrum of star-forming galaxies below the survey detection threshold is convex rather than a simple power law: it steepens toward low frequencies and flattens toward high frequencies. The curved-power-law fit returns a positive curvature parameter in every redshift bin ($q \\approx 0.04$ at $z \\approx 0.7$ up to $q \\approx 0.35$ at $z \\approx 2.7$), and the double-power-law fits give steep synchrotron indices from $\\alpha_{\\rm DPL} \\approx -0.80$ to $\\alpha_{\\rm DPL} \\approx -1.96$, well below the canonical $\\alpha \\approx -0.8$. The two-point spectral index $\\alpha_{\\rm low}$ (144-320 MHz) is steeper than $\\alpha_{\\rm high}$ (656-1500 MHz) in every redshift bin where both could be measured. The paper further claims no significant inverse correlation between spectral index and redshift over $0.1 \\leq z \\leq 3.0$, so the average SFG radio spectrum does not evolve with cosmic time. At 1500 MHz, the thermal free-free fraction is 11-18 percent, leaving synchrotron emission dominant throughout the band; the models all fit the data adequately, with no strong preference among power law, curved power law, and double power law.","pith_inferences":["The tension with the low-frequency flattening reported in the two radio-selected studies the paper cites may be a selection effect: a radio detection requires brightness at the survey frequency, so radio-selected samples could favor flatter-spectrum objects, whereas the Ks-selected stack here is dominated by below-threshold galaxies. Splitting the XMM-LSS stack by whether galaxies are detected at ","If the convex spectrum is generic, rest-frame 144 MHz luminosities of faint high-redshift SFGs are lower than $\\alpha = -0.8$ extrapolations from 1.4 GHz imply, which should appear as a deficit of ultra-steep-spectrum sources in low-frequency-selected samples, a cross-check existing LOFAR source-count data could already perform.","The paper's own thermal-fraction formula predicts $f_T \\approx 0.5$ at 30 GHz, so its interpretation implies a quantitative target: extending the same stacking from 1.5 to 10 GHz should show the spectrum progressively flattening toward $\\alpha \\approx -0.1$. If instead the slope stays steep, the convexity is intrinsic synchrotron curvature rather than a free-free signature."],"forward_implications":["Radio k-corrections that assume one power-law slope of $\\alpha \\approx -0.8$ mis-estimate rest-frame luminosities of high-redshift star-forming galaxies, and the error grows with redshift because the true average spectrum is curved.","Because the spectral index does not correlate with redshift, luminosity functions and star-formation-rate densities built from radio surveys can adopt a redshift-independent average spectral shape rather than an evolving one.","With a thermal fraction of only 11-18 percent at 1500 MHz, synchrotron emission dominates the sub-mJy radio population, and the free-free component should become increasingly visible at frequencies above $\\sim$1.5 GHz, a concrete prediction for the 1-10 GHz stacking the paper outlines.","The consistent low-frequency steepening across all seven bins means the rest-frame low-frequency (around 144 MHz) emission of faint SFGs is weaker than a straight $\\alpha = -0.8$ extrapolation from 1.4 GHz would predict."],"supporting_citations":[{"why":"Supplies the LOFAR 144 MHz image whose stacked median flux anchors the low-frequency end of every SED.","marker":"Hale et al. 2019"},{"why":"The superMIGHTEE uGMRT pilot data provide the eight mid-frequency stacked points from 320 to 784 MHz.","marker":"Lal et al., submitted"},{"why":"Supplies the MIGHTEE 1284 MHz image serving as the mid-band anchor between the uGMRT and VLA frequencies.","marker":"Heywood et al. 2022"},{"why":"Supplies the VLA 1500 MHz image that fixes the high-frequency end and the thermal-fraction estimate.","marker":"Heywood et al. 2020"},{"why":"The SPLASH-SXDF catalogue defines the Ks-selected sample, photometric redshifts, and physical properties.","marker":"Mehta et al. 2018"},{"why":"Establishes that a stacked median is boosted relative to the true median, motivating the correction applied to all fluxes.","marker":"White et al. 2007"},{"why":"Provides the mock-source injection procedure from which the flux-boosting correction factors are derived.","marker":"Algera et al. 2022"},{"why":"The low-frequency flattening result that this paper's steepening result is explicitly contrasted against.","marker":"An et al. 2021"},{"why":"The rest-frame UVJ criterion used to separate star-forming from quiescent galaxies.","marker":"Schreiber et al. 2015"},{"why":"The mid-infrared color criterion used to remove AGN-contaminated sources from the SFG sample.","marker":"Donley et al. 2012"}],"fun_headline_variants":["Star-forming galaxies' radio spectra curve, steepening at low frequencies","Radio spectra of faint galaxies curve down, no redshift trend","Faint galaxies' radio spectra are convex, not flattened by redshift","Stacked SFG radio spectra steepen low, flatten high, steady over z","Convex radio spectra in faint galaxies, thermal fraction 11-18%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the median-stacked flux density measured at 144 MHz in LOFAR's coarse $7.5\\times8.5$ arcsec beam, after the simulation-based boosting correction, is directly comparable to the stacked fluxes from the finer-resolution uGMRT, MeerKAT, and VLA maps, so that residual source blending or an imperfect correction at that single frequency could not be the true origin of the reported low-frequency steepening.","fun_headline_variants_meta":{"raw":{"variants":["Star-forming galaxies' radio spectra curve, steepening at low frequencies","Radio spectra of faint galaxies curve down, no redshift trend","Faint galaxies' radio spectra are convex, not flattened by redshift","Stacked SFG radio spectra steepen low, flatten high, steady over z","Convex radio spectra in faint galaxies, thermal fraction 11-18%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000543,"raw_usage":{"total_tokens":2719,"prompt_tokens":1183,"completion_tokens":1536,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":799,"completion_tokens_details":{"reasoning_tokens":1442}},"tokens_in":799,"tokens_out":1536,"duration_ms":12316,"temperature":1.0,"reasoning_tokens":1442,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:16:17.395661+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-derive the stacked SED from 144 MHz images restored to the 4.5-arcsec resolution of the VLA 1500 MHz map, recomputing the flux-boosting correction by injecting mock sources into the convolved maps; if the corrected 144 MHz median flux density drops by more than its measurement uncertainty (the 144 MHz stacks have signal-to-noise ratios of roughly 4-7), the low-frequency steepening and the $q>0$ curvature are resolution artifacts rather than properties of the galaxies. A complementary check: stack only the sources individually detected at $\\geq 5\\sigma$ at 1284 MHz and ask whether the convex shape survives in a radio-selected subsample.","supporting_citations":[],"review_version":1}