3.1. DTC per-satellite excess

Figure 5 shows the cumulative distribution function of the flux density for the DTC and Ku-only comparison populations under four equivalent reductions. The raw per-detection comparison (panel a) is dominated by the range distribution of each population and is therefore not a measurement of intrinsic brightness. The range-corrected per-detection comparison (panel b) removes the geometric factor and yields a DTC/Ku median ratio that is quoted in the panel title. The per-satellite statistics in panels (c) and (d) give equal weight to each satellite, and avoid biasing the comparison in favour of satellites with unusually high detection counts. The range-corrected per-satellite median (panel d) is adopted as the main summary statistic for the per-satellite excess.

The main summary statistic yields a DTC/Ku median ratio of $1.449$ with a 95% bootstrap confidence interval of $[1.27, 1.67]$ drawn from $2\,000$ resamplings. A two-sided Mann–Whitney U test rejects the null hypothesis of identical distributions at $p = 4.1\times10^{-11}$ . The effect size quantified by Cliff’s $\delta$ is $+0.303$ . The direction of the effect is consistent across the four reductions shown in Figure 5(a)–(d), and the ratio is largest for the per-satellite range-corrected reduction, which is the reduction least affected by the geometry of the observations. Cross-checking against the LOFAR sample of Bassa et al. (Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024), the six DTC satellites observed there all appear in our 175-satellite DTC sample with the EDA2 detection counts listed in Table 2, showing sample overlap across the two instruments rather than independently confirming the EDA2 per-satellite excess.

The two datasets sample the same UEMR phenomenon at different points in the time–frequency plane. The LOFAR observations of Bassa et al. (Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024) consist of two 1-h coherent COBALT tied-array-beam acquisitions on 2024 July 19, with a native resolution of $12.2$ kHz $\times$ $41.94$ ms across the low-band antenna (LBA) (10–88 MHz) and high-band antenna (HBA) (110–188 MHz) bands. The present EDA2 reanalysis instead draws on the multi-night Grigg et al. (Reference Grigg, Tingay and Sokolowski2025) catalogue across 21 discrete coarse channels spanning $72.685$ – $234.375$ MHz, so the two observations overlap only in the 110–188 MHz window.

Within that window, all six v2-Mini DTC satellites in the HBA sample of Bassa et al. (Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024) (NORAD 60039–60048) appear in our 175-satellite DTC selection, with per-satellite EDA2 detection counts ranging from 16 to 103. The dispersion arises from a combination of pass geometry, the number of EDA2 transits per satellite, and intrinsic UEMR variability, whose relative contributions are not separable in the present sample. These counts establish EDA2 coverage of the six objects. We do not compare the EDA2 detection counts with the LOFAR flux densities reported for those satellites because they are different observables measured under different selection functions.

The two methodologies are complementary. The $12.2$ kHz LOFAR resolution is fine enough to resolve the comb spacings of 50 and 150 kHz reported by Bassa et al. (Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024) in the 116–124 MHz band for v2-Mini DTC satellites; the EDA2 coarse channelisation cannot. Conversely, the LOFAR HBA upper edge of 188 MHz lies below the $230.469$ MHz channel at which we report the largest polarisation anomaly (Section 3.2), and the spectral structure at that frequency awaits confirmation from independent broadband observations. The cross-check in Table 2 is therefore best read as a sample-membership consistency test; the per-satellite excess and the Bassa et al. (Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024) generation-over-generation factor of 32 are related in Section 4.

Figure 6 decomposes this comparison into each of the 21 coarse frequency channels. The majority of channels show a DTC/Ku ratio above unity, consistent with the per-satellite excess being broadband rather than concentrated in any single frequency band. A single channel at $130.469$ MHz shows the opposite sense with a particularly extreme separation from unity, where the DTC population is fainter than the Ku-only comparison by a factor of approximately six (median ratio $0.16$ ). Several other channels (e.g. 170, 180, 200, 225, 234 MHz; see Table 3) show DTC median ratios modestly below unity, but the magnitude of the $130.469$ MHz separation exceeds that of every other below-unity channel by an order of magnitude. No physical interpretation of the $130.469$ MHz exception is attempted here. The shaded band in Figure 6 marks the protected Radio Astronomy Service band between $150.05$ and 153 MHz; the one coarse channel that falls within this band is consistent with the broadband trend and is not anomalous.

3.2. Coarse-channel polarisation anomalies

Figure 7 shows the deviation of the XX polarisation fraction from the pooled v2-Mini baseline $f_\mathrm{XX,inst} = 0.4809$ in each of the 21 coarse channels. The baseline is the pooled-sample XX fraction across the $112\,377$ v2-Mini detections analysed here, and represents the XX fraction expected under the null hypothesis that the EDA2 orthogonal polarisation feeds have equal sensitivity and the sources emit unpolarised radiation; it is a pooled-data estimate rather than an independent instrumental calibration, and the leave-one-channel-out robustness check reported below confirms that the largest deviation is preserved under a null baseline that explicitly excludes the test channel. Positive bars indicate channels in which the XX polarisation is observed more frequently than the baseline, and negative bars indicate the opposite. A two-sided event-level binomial test is applied in each channel independently, and the Benjamini–Hochberg FDR procedure at $q = 0.05$ is used to control the false discovery rate across the 21 tests. After this correction, 11 of the 21 channels retain significance (coloured bars in Figure 7), and the remaining 10 are consistent with the null baseline at the $q = 0.05$ level.

Figure 5.

Cumulative distribution functions of the flux density for the DTC (orange) and Ku-only (purple) populations under four equivalent reductions: (a) raw per-detection; (b) range-corrected per-detection; (c) raw per-satellite median; (d) range-corrected per-satellite median. The ratio of medians is quoted in each panel title; the 95% bootstrap confidence interval is shown for panel (d).

The largest deviation, by a wide margin, occurs at $230.469$ MHz, where the XX fraction reaches $f_\mathrm{XX} = 0.811$ with a 95% Wilson score interval of $[0.795, 0.825]$ . A two-sided binomial test at this channel returns an event-level $p \sim 10^{-274}$ , well beyond the significance of any other channel in the dataset. The channel lies 80 MHz above the protected Radio Astronomy Service band at $150.05$ –153 MHz, and has no assigned Starlink service. The binomial test used here addresses only the population-level event-counting statistic; repeated detections within satellites mean that this is not a cluster-robust significance measure. As a sensitivity check, we compare the $230.469$ MHz XX fraction with a leave-one-channel-out pooled v2-Mini baseline using a satellite-cluster sandwich variance, allowing arbitrary dependence within each satellite. The deviation remains $+0.338$ (cluster standard error $0.025$ , $p=2.3\times10^{-40}$ ; 84 satellites contribute detections in the test channel). It does not address whether the emission itself is linearly polarised in a physical sense, but only whether EDA2 is more likely to detect the source in the X polarisation feed than in the Y feed.

Figure 8 characterises this channel in more detail. Panel (a) shows the distribution of per-satellite XX fractions for satellites with at least five detections at $230.469$ MHz; the distribution is asymmetric and centred well above the baseline, with most satellites showing $f_\mathrm{XX}$ between $0.7$ and $1.0$ . Panel (b) sorts the same satellites by their XX fraction and colour-codes each point by the mean observing elevation; no elevation dependence is apparent, which argues against a dominant beam-shape systematic. Panel (c) shows that both the DTC and Ku-only comparison populations contribute to the excess, with the DTC population exhibiting the larger deviation, consistent with the per-satellite excess reported in Section 3.1. Panel (d) compares the log-flux distributions in the two polarisations directly: the XX distribution contains substantially more events than the YY distribution, but the two distributions have similar shape and median, suggesting that the anomaly is in the number of detections rather than in the per-detection amplitude.

Table 2.

Cross-instrument check against the LOFAR sample of Bassa et al. (Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024). All six DTC satellites observed there, identified by their North American Aerospace Defense Command (NORAD) catalogue numbers, appear in our EDA2 sample with detection counts ranging from 16 to 103.

Table 2. Long description

A table with six rows and three columns. The columns are labeled NORAD identifier, Starlink name, and EDA2 detections. The rows list the following data: Row 1: NORAD identifier, 60039; Starlink name, STARLINK-11140; EDA2 detections, 54. Row 2: NORAD identifier, 60040; Starlink name, STARLINK-11122; EDA2 detections, 66. Row 3: NORAD identifier, 60041; Starlink name, STARLINK-11149; EDA2 detections, 103. Row 4: NORAD identifier, 60042; Starlink name, STARLINK-11120; EDA2 detections, 16. Row 5: NORAD identifier, 60043; Starlink name, STARLINK-11086; EDA2 detections, 18. Row 6: NORAD identifier, 60048; Starlink name, STARLINK-11135; EDA2 detections, 37.

Notes: Starlink names from the McDowell GCAT (McDowell Reference McDowell2020); detection counts from this work’s reanalysis of the Grigg et al. (Reference Grigg, Tingay and Sokolowski2025) catalogue.

Table 4 summarises the per-channel binomial test against the instrumental baseline and the Benjamini–Hochberg significance status for all 21 coarse channels in the dataset.

As a robustness check against the circularity of estimating the pooled-sample baseline from data that include the test channel, a leave-one-channel-out baseline is reported. For each channel c, a null XX fraction is computed from the other twenty channels combined, and the binomial test of channel c is repeated against this channel-specific null. Under the leave-one-channel-out baseline, 12 of the 21 channels survive Benjamini–Hochberg control at $q=0.05$ (versus 11 under the pooled-sample baseline), and the $230.469$ MHz channel returns a deviation of $+0.338$ against a leave-one-out baseline of $0.473$ with a binomial $p \sim 10^{-287}$ . The qualitative pattern of which channels deviate, and in particular the headline status of the $230.469$ MHz anomaly, is therefore preserved when the test channel is explicitly excluded from the null baseline calibration.

Figure 6.

Per-frequency ratio of DTC to Ku-only range-corrected median flux density, with 95% bootstrap confidence intervals. Squares filled in dark orange lie above unity (DTC-brighter channels); squares filled in lighter salmon lie below unity. The most extreme below-unity channel is at $130.469$ MHz, where the DTC/Ku ratio falls to $\approx$ $0.16$ . The shaded band indicates the Radio Astronomy Service band protected by the ITU between $150.05$ and 153 MHz.

3.3. Spectral structure of detections

Table 3.

Per-coarse-channel DTC/Ku range-corrected median flux ratio, complementing Figure 6. Channels with fewer than 20 DTC detections (four channels with $N_\mathrm{DTC}\in\{0,0,0,16\}$ ) are omitted because the per-channel median ratio is not reliably estimated; the $10\,164$ DTC detections tabulated here account for $10\,180-16=10\,164$ of the $10\,180$ DTC detections in the sample.

Table 3. Long description

A table with 11 rows and 6 columns comparing DTC/Ku flux ratios across different frequency channels. The columns are labeled as follows: Frequency (MHz), N DTC, N Ku, DTC/Ku, KS p, and Cliff δ. The table presents data for various frequency channels, with each row providing specific values for the listed parameters. Notable trends include variations in the DTC/Ku ratio and KS p values across different frequency channels.

Notes: KS denotes the Kolmogorov–Smirnov test. $N_\mathrm{DTC}$ and $N_\mathrm{Ku}$ are the detection counts contributing to the per-channel median ratio. Cliff’s $\delta$ is signed such that positive values indicate the DTC population is brighter; the only channel at which DTC is fainter at $|\delta|\gt0.5$ is $130.469$ MHz (bold).

Figure 7.

Deviation of the XX polarisation fraction from the pooled v2-Mini baseline $f_\mathrm{XX,inst}$ in each coarse channel. Bars coloured in dark red survive the Benjamini–Hochberg false discovery rate (BH-FDR) control at $q = 0.05$ ; grey bars do not. The annotation at $230.469$ MHz marks the channel with the largest deviation.

Two complementary spectral statistics summarise the per-channel distribution of v2-Mini UEMR in the dataset. The first is the detection-count column N in Table 4, which quantifies where the catalogue threshold is exceeded most often. Detections cluster strongly at $150.78$ , $153.12$ , $161.72$ , and $170.31$ MHz, which together account for $\sim$ $75\,000$ of the $\sim$ $112\,500$ stacked detections in the analysed sample (about two-thirds of the total). A secondary peak lies near 200 MHz, and the 73–110 MHz lower band is sparsely populated. The second statistic is the per-channel DTC/Ku ratio plotted in Figure 6: most channels show a ratio above unity, with the largest single-channel exception at $130.469$ MHz, where the ratio falls to $\approx 0.16$ and the DTC population is fainter than the Ku-only comparison.

The dataset can be sub-divided into three frequency regions for discussion. The low-band region between 73 and 110 MHz contains only three coarse channels in the analysed sample, each with $N \lt 350$ detections, and none of the three survives BH-FDR correction; this region therefore contributes weakly to the catalogue-level statistics and is consistent with sparse, near-threshold emission. The mid-band region between 110 and 190 MHz contains eleven coarse channels and most of the catalogue detections; it overlaps the LOFAR HBA window of Di Vruno et al. (Reference Di Vruno, Winkel, Bassa, Józsa, Brentjens, Jessner and Garrington2023) and Bassa et al. (Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024), in which v2-Mini DTC satellites were observed to emit broadband UEMR plus narrow comb structure at the 50 and 150 kHz scales. The high-band region between 200 and $234.4$ MHz contains four coarse channels absent from the LOFAR HBA coverage; three of these four channels ( $225.781$ , $230.469$ , and $234.375$ MHz) survive BH-FDR correction and together contain the largest XX-fraction excursions in the dataset. The $230.469$ MHz channel in particular sets the most extreme polarisation anomaly observed in any v2-Mini sub-population reported to date.

The per-channel asymmetry distribution above 200 MHz is therefore not a simple continuation of the lower-frequency pattern. At $225.781$ and $230.469$ MHz the XX-fraction deviations have the same positive sign ( $+0.091$ and $+0.330$ respectively) but differ by a factor of $\sim$ 4 in amplitude; at $234.375$ MHz the deviation reverses sign and is the largest negative deviation in the sample ( $\Delta=-0.109$ ). No published Starlink Direct-to-Cell band allocation overlaps the EDA2 frequency range, and the per-channel structure cannot be attributed to direct in-band emission from the cellular payload. Plausible origins include harmonic and intermodulation products of on-board oscillators and direct-current to direct-current (DC–DC) converters, conducted emission from power-conditioning subsystems, and out-of-band spectral leakage from communications hardware. These mechanisms are not distinguishable from the EDA2 catalogue alone, because the EDA2 coarse-channel width does not resolve the comb spacings of tens to hundreds of kHz reported by Bassa et al. (Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024) at LOFAR resolution. The mechanism discussion is deferred to Section 4.

3.4. Sub-MHz fine-channel structure of the $230.469$ MHz anomaly

The $230.469$ MHz coarse-channel anomaly reported in Section 3.2 is integrated over the full $\sim$ $0.78$ MHz coarse channel and is therefore agnostic to any sub-MHz spectral structure. Because Bassa et al. (Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024) resolved 50 and 150 kHz comb spacings within the 116–124 MHz band at LOFAR’s $12.2$ kHz resolution, and because the $230.469$ MHz channel lies outside their HBA coverage, the natural follow-up is to inspect the $\sim$ 24 kHz fine-channel data within the EDA2 coarse band.

Figure 8.

Characterisation of the $230.469$ MHz XX fraction anomaly. (a) Per-satellite XX fraction distribution for satellites with at least five detections in this channel. (b) The same, sorted by XX fraction and colour-coded by mean elevation. (c) XX fraction by population, with the number of detections quoted. (d) Range-corrected flux density distributions in XX and YY, showing that the XX excess is not accompanied by a commensurate increase in the per-detection amplitude.

We isolate the $80\,894$ v2-Mini fine-bin observations (fine-channel index 0–30 in the Grigg et al. (Reference Grigg, Tingay and Sokolowski2025) catalogue, $\sim$ $24.4$ kHz per bin) at the $230.469$ MHz coarse channel and recompute per-fine-bin statistics independently of the stacked detection summary already used in Section 3.2. As shown in Figure 9, fine channel 22 (corresponding to $\approx$ $230.627$ MHz, an offset of $+159$ kHz from the coarse-channel centre) carries an empirical inter-bin standardised contrast $z_\mathrm{inter}=11.0$ in the mean flux relative to the other 30 fine bins (top panel: $72.86$ Jy in fine channel 22 versus $61.92 \pm 0.99$ Jy elsewhere; per-bin sample sizes range from $N=2\,525$ in fine channel 22 to $N=2\,641$ elsewhere, with the $\sim$ $80\,894$ total v2-Mini fine rows reflecting the per-bin post-filter rather than $2\,704 \times 31$ raw) and a corresponding $+55\%$ excess in the 95th-percentile flux (middle panel). Because the same detections contribute to every fine-bin mean, $z_\mathrm{inter}$ is a descriptive inter-bin contrast rather than a standard-normal test statistic. We therefore do not attach a normal-theory or Bonferroni p-value to it; inference rests on the magnitude of the contrast and the cross-channel controls. The median flux in fine channel 22 is not elevated, which identifies the feature as a tail-driven phenomenon contributed by a sub-population of bright detections rather than a uniform shift in the bulk of the distribution.

The cross-coarse-channel control shown in Figure 10 addresses a known concern: EDA2 fine channels can in principle carry a fixed bandpass artefact that would inflate flux at the same index across every coarse channel. We re-run the per-fine-bin mean-flux estimator at the five highest-occupancy control channels in the catalogue ( $150.78$ , $153.12$ , $161.72$ , $170.31$ , and $200.00$ MHz) and recover z-scores between $-0.4$ and $+0.9$ , all within the $\pm 3\sigma$ band. The control therefore does not reproduce a fixed fine-channel-index enhancement in the tested control channels, which argues against such an artefact and supports, but does not prove, a narrowband emission process specific to the $230.469$ MHz coarse channel rather than to a per-fine-index instrumental response; adjacent high-band coarse channels ( $225.781$ and $234.375$ MHz) are not used as controls because they themselves carry the largest XX-fraction excursions in the catalogue (Table 4) and cannot be assumed feature-free.

Expressed as a ratio rather than absolute mean, the DTC subset shows an excess in fine channel 22 of $83.93 / 70.1 = 1.197\times$ over its non-target bins, compared with the Ku-only ratio of $\approx$ $50.2/46.8 = 1.073\times$ , that is, the DTC fine-channel narrowband excess is approximately $0.12$ in fractional units larger than the Ku-only excess at the same fine bin. The XX-feed fraction in fine channel 22 ( $0.834$ ) is nominally elevated above the pooled per-bin value of $0.810$ , but this comparison is not independent: the Grigg et al. (Reference Grigg, Tingay and Sokolowski2025) catalogue records a single per-detection polarisation index that is propagated across all 31 fine-channel rows of the same detection, so per-fine-bin XX/YY values are not separate measurements and the nominal significance is not interpretable in the usual binomial sense. The narrow flux excess therefore implies a narrowband emission process at $\approx$ $230.627$ MHz that contributes preferentially to detections recorded in the XX feed; a definitive mechanism-level interpretation is deferred to Section 4.1, where this fine-channel constraint is combined with the eclipse-state result. Before that, the next subsection (Section 3.5) presents three mechanism-oriented observational checks that further constrain the physical origin of the $230.627$ MHz feature without invoking a phenomenological model.

Table 4.

Per-channel event-level binomial test of the XX detection fraction against the pooled v2-Mini baseline $f_\mathrm{XX,inst}=0.4809$ . Eleven of the 21 coarse channels survive Benjamini–Hochberg false discovery rate control at $q=0.05$ . The $230.469$ MHz channel is the largest deviation in the dataset.

Table 4. Long description

A table with 11 rows and 5 columns. The columns are labeled as follows: f (MHz), N, fxx, Δ, p-value, and BH. The table presents the results of a binomial test for different frequency channels, including the frequency in MHz, the number of events (N), the detection fraction (fxx), the difference (Δ), the p-value, and the Benjamini-Hochberg (BH) significance status. Each row provides the specific values for these parameters across different channels. Notable values include the highest deviation in the dataset at 150.78 MHz with a p-value of 6 x 10^-61 and the largest detection fraction at 225.78 MHz with a value of 0.572.

Notes: N: total detections in the channel; $f_\mathrm{XX}$ : XX-feed detection fraction; $\Delta$ : deviation from the pooled baseline; BH column marks channels surviving false discovery rate control at $q=0.05$ ( $11/21$ ).

3.5. Mechanism-oriented checks of the fine-channel feature

The fine-channel narrowband excess of Section 3.4 admits several physical interpretations – harmonic emission from on-board clock fundamentals, a switch-mode payload mode, an intermodulation product between oscillators, uniform thermal scaling across the population, or a hardware fault localised to a small sub-population. We use three numerical checks to assess which simple interpretations are consistent with the observed patterns. The checks address: (T1) whether the $230.627$ MHz feature is consistent with an integer harmonic of any clock fundamental resolved by Bassa et al. (Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024) at LOFAR’s $12.2$ kHz resolution; (T2) whether the fine-channel excess is confined to a single $\sim$ $24.4$ kHz bin or smears into adjacent fine bins; and (T3) whether the per-satellite distribution of fine-channel excess is dominated by a few permanently-bright satellites or expressed uniformly across the fleet. These alternatives are descriptive; the catalogue does not identify an operational duty cycle. The combined panel results are shown in Figure 11.

T1 – Harmonic-coincidence test

We evaluated the integer-harmonic predictions $N \times f_0$ of each of the 14 Bassa+2024 clock fundamentals ( $27.5$ , $36.66$ , 55, 110, and 220 kHz for v2-Mini DTC LBA; $37.5$ , 50, 75 and 150 kHz for v2-Mini LBA; 50, 150 kHz for v2-Mini DTC HBA at 116–124 MHz; $48.8$ , 65 and $97.5$ kHz for v2-Mini HBA at 157–165 MHz) and selected, for each fundamental, the integer N that minimises the residual relative to the centroid of EDA2 fine channel 22 $230.627441$ MHz. A match was declared when $|N f_0 - 230.627441\,\mathrm{MHz}| \le 12.207$ kHz, that is, within one half of the EDA2 fine bin (Figure 11(a)). Five of the 14 fundamentals produce a match within tolerance. The expected number of chance coincidences under the null hypothesis of uniformly-distributed harmonics on the EDA2 fine grid is $\sum_i \min(1, 2 \Delta f / f_{0,i}) = 5.73$ with $\Delta f = 12.207$ kHz, so the observed count is at the chance level (observed/expected $= 0.87$ ) and the matching set is scattered across fundamentals rather than concentrated on a single one. The 50 and 150 kHz fundamentals each appear twice in the 14-entry list (once for the LBA non-DTC v2-Mini context and once for the HBA DTC v2-Mini context, in which Bassa et al. Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024 report the same numerical comb in distinct band/population settings); treating these as a single oscillator each reduces the expected count to $\approx$ $5.08$ and leaves the observed/expected ratio at $0.98$ – still at the chance level. Because harmonically related fundamentals are not statistically independent, the observed/expected ratio is reported as a rule-out criterion rather than as a formal p-value. A superset of 9 additional crystal-oscillator candidates (10, 13, 16, 20, 25, 27, 100 MHz and the $12.288$ MHz audio codec, plus a $32.768$ kHz real-time clock) yields one match (the $32.768$ kHz real-time clock (RTC) at harmonic $N = 7038$ ), also at chance ( $0.76$ expected). We conclude that the $230.627$ MHz feature is not uniquely tied to any LOFAR-resolved Starlink clock fundamental, disfavouring a common oscillator with the 116–165 MHz combs while not excluding clock-spurious emission in general from other internal references not characterised by the lower-band study.

T2 – Adjacent-bin coherence

The single-bin nature of the excess in fine channel 22 identified in Section 3.4 can be tested directly. Pivoting the v2-Mini detections at the $230.469$ MHz coarse channel into a wide table indexed by fine-channel (31 fine bins per detection, $N=2\,525$ detections with valid flux in fine channel 22 after the geometry filter), we restrict to detections in which the flux in fine channel 22 exceeds its sample 95th percentile ( $p_{95} = 241$ Jy, $N = 127$ detections in the bright subset) and compute the mean flux at each of the other 30 fine bins within this subset (Figure 11(b)). The mean of the 28 bins excluding fine channels 21, 22, and 23 defines the inter-bin baseline ( $\mu = 143.5$ Jy beam $^{-1}$ , $\sigma = 8.4$ Jy beam $^{-1}$ across the 28 bin means). The z-scores relative to this baseline are $z_{22} = +25.4$ (a sanity check confirming the bright cut), $z_{21} = -1.26$ , and $z_{23} = -0.96$ ; both adjacent bins therefore lie at or below baseline rather than above. Neither adjacent bin is elevated in this subset. The $230.627$ MHz feature is therefore unresolved beyond a single $24.4$ kHz bin at the EDA2 resolution; this check does not by itself establish the intrinsic line profile, and a sharper measurement requires higher-resolution follow-up. The negative adjacent-bin z-scores are descriptive and are not interpreted as evidence for suppressed neighbouring-bin power.

Figure 9.

Per-fine-channel statistics within the $230.469$ MHz coarse band ( $N=80\,894$ v2-Mini fine-bin observations from the Grigg et al. (Reference Grigg, Tingay and Sokolowski2025) catalogue, geometry-filtered). Top: mean flux, split DTC vs Ku-only. Middle: median and 95th-percentile flux, pooled. Bottom: XX-feed fraction with binomial error bars; the pooled-sample value $\hat p=0.810$ is marked. Fine channel 22 (catalogue index 22; $\approx$ $230.627$ MHz) is highlighted as the only bin with markedly elevated mean and upper-percentile flux; the XX-fraction variation across bins is constrained by the per-detection polarisation assignment in the catalogue and is informative only insofar as the mean shift reflects the larger-flux fraction in fine channel 22.

T3 – Per-satellite ratio distribution

For each of the 82 v2-Mini satellites with $\ge$ 5 detections at the $230.469$ MHz coarse channel, we computed the per-satellite ratio $R_i = \overline{F_{22,i}}\, /\, \overline{F_{\mathrm{other},i}}$ in which the numerator is the satellite’s mean flux in fine channel 22 and the denominator is its mean over the 28 non-adjacent fine bins. The distribution (Figure 11(c)) has median $R = 1.07$ , mean $R = 1.18$ , $p_{95} = 1.79$ , and maximum $R = 3.75$ . We use the distribution for two descriptive robustness checks. Three of the 82 satellites ( $3.7\%$ ) have $R_i \gt 2$ . Re-running the same inter-bin z-score machinery applied in Section 3.4 on the 79 remaining satellites (i.e. removing all detections from the three $R_i \gt 2$ satellites and recomputing the per-fine-bin means and inter-bin baseline) yields $z_{22} = +12.95$ – the excess remains after the brightest three are removed, so those three satellites alone do not account for the signal. The mean of the top decile of $R_i$ is $2.13$ , whereas the mean of the bottom half is $0.94$ (Figure 11(d)); the $2.26\times$ ratio shows substantial rank-defined dispersion, and the $45/82$ ( $54.9\%$ ) satellites with $R_i \gt 1.05$ show that the excess is not limited to a few units. This spread could reflect intrinsic satellite-to-satellite heterogeneity, intermittent emission, sampling variation, or measurement noise; the catalogue alone does not distinguish among them.

Combined interpretation

The three checks narrow the observational description without identifying the mechanism. The $230.627$ MHz feature (T1) is not coincident with any uniquely-identified LOFAR-resolved clock fundamental, (T2) is not resolved beyond one $24.4$ kHz fine bin at the EDA2 resolution, and (T3) is not solely driven by the three largest per-satellite ratios. These observations are compatible with, but do not select, several electronic-emission scenarios, including a payload-active, communications-active, or power-management regime, a switch-mode power-supply harmonic that activates with service pattern, or an intermodulation product between on-board oscillators not catalogued by Bassa et al. (Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024) – and these constraints apply to the v2-Mini population as a whole rather than only to the DTC subset, since the $230.627$ MHz fine-channel feature is present in both the DTC and Ku-only subpopulations (Section 3.4). However, the present data cannot uniquely identify the emitting subsystem. A mechanism-level confirmation would require telemetry-correlated observation cadence or direct payload characterisation, neither of which is available in the public domain at present. The orthogonal eclipse-state evidence considered next (Section 3.6) provides an independent handle on the same question.

Figure 10.

Cross-coarse-channel control for the feature in fine channel 22. The z-score is the deviation of the mean flux in fine channel 22 from the mean over the other 30 fine bins, in units of the inter-bin standard deviation. The $230.469$ MHz channel (red star) shows $z = +11.0$ , whereas five high-occupancy control channels ( $150.78$ , $153.12$ , $161.72$ , $170.31$ , and $200.00$ MHz; blue circles) all lie within $\pm 1\sigma$ , arguing against a fixed same-index response across the tested controls.

3.6. Eclipse-state dependence

The pooled-population form of the eclipse-state comparison is tested first, in which the range-corrected flux density of all v2-Mini satellites is compared between illuminated and eclipsed states without regard to the payload variant. The pooled test returns an illuminated/eclipsed ratio of medians of $1.164\,[1.151,1.177]$ , formally rejecting the null of no dependence but by a small margin. Taken on its own, this result would suggest a weak positive correlation between illumination and UEMR, consistent with a solar-array-driven power-conditioning mechanism. As shown below, however, this pooled result is a pooling artefact of the unequal sub-population sizes combined with opposite-signed within-population trends, and its physical interpretation after sub-population separation is misleading.

When the same test is applied separately to the DTC and Ku-only comparison populations, the picture changes qualitatively. The Ku-only comparison population yields an illuminated/eclipsed ratio of $1.184\,[1.172, 1.197]$ (95% bootstrap confidence interval), consistent in direction with the pooled result and in the expected sense: illuminated Ku-only satellites emit more UEMR than eclipsed ones. The DTC population, by contrast, yields an illuminated/eclipsed ratio of $0.465\,[0.439, 0.492]$ (95% bootstrap confidence interval), corresponding to DTC the pooled detected-event sample being roughly $2.1\times$ brighter when eclipsed than when illuminated. The two population ratios have opposite signs, and their 95% satellite-cluster bootstrap intervals reported below are disjoint for this pooled estimator. Because the Ku-only comparison contributes an order of magnitude more detections than the DTC population, the Ku-only sign dominates the pooled median and masks the DTC reversal.

Figure 12 examines the robustness of the DTC eclipse-state reversal against the most obvious systematics. Panel (a) shows the distribution of per-satellite illuminated/eclipsed ratios for DTC satellites with at least five detections in each state. The distribution extends predominantly below unity; the solid vertical line in the panel marks the population ratio of medians, $0.47$ , which is the same summary statistic adopted in the primary test above and is not the same quantity as the median of the per-satellite ratios, which is closer to unity because individual satellites with modest detection counts scatter around the population trend. Panel (b) shows the ratio as a function of satellite altitude in bins of 56 km; the DTC reversal is preserved across the 300–412 km altitude range with non-trivial detection counts, while the Ku-only comparison ratio remains above unity in every altitude bin. The numerical content of the populated DTC altitude bins is summarised in Table 5.

Figure 11.

Three mechanism-oriented observational checks for the $230.627$ MHz fine-channel feature reported in Section 3.4. (a) T1, harmonic-coincidence test: predicted frequencies $N \times f_0$ of all 14 fundamentals reported by Bassa et al. (Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024) for v2-Mini Starlink satellites (top row) and a superset of 9 plausible crystal-oscillator candidates (bottom row), evaluated at the integer N minimising the residual relative to the target centroid $230.627441$ MHz; the shaded band is the $\pm 12.207$ kHz EDA2 half-fine-bin tolerance. Filled circles denote matches within tolerance and crosses denote non-matches. Five of the 14 Bassa fundamentals produce a match within tolerance, compared with $5.73$ expected by chance under the null hypothesis $\sum_i \min(1, 2\Delta f/f_{0,i})$ over the tested fundamentals. (b) T2, adjacent-bin coherence within the bright subset (top 5% by flux in fine channel 22, $N=127$ detections): the mean flux per fine bin is shown against the inter-bin baseline (shaded band, $\mu = 143.5$ Jy beam $^{-1}$ , $\sigma = 8.4$ Jy beam $^{-1}$ over the 28 bins excluding fine channels 21, 22, and 23). The point for fine channel 22 at $z=+25.4$ is off-scale relative to the inter-bin spread; fine channel 21 ( $z=-1.26$ ) and fine channel 23 ( $z=-0.96$ ) lie at or below baseline rather than above, showing no adjacent-bin elevation in this subset. (c) T3, distribution of per-satellite ratios $R_i = \overline{F_{22}} / \overline{F_\mathrm{other}}$ across the 82 v2-Mini satellites with $\ge 5$ detections at the $230.469$ MHz coarse channel. (d) Rank-ordered $R_i$ on a logarithmic axis; horizontal references mark the top-decile mean $R = 2.13$ and the bottom-half mean $R = 0.94$ .

Figure 12.

Robustness of the DTC eclipse-state reversal. (a) Distribution of per-satellite illuminated/eclipsed ratios for DTC satellites with at least five detections in each state; the solid vertical line marks the population ratio of medians (the same statistic adopted in the primary test in the text), and the dashed line marks unity. (b) Illuminated/eclipsed ratio as a function of satellite altitude, for DTC (orange squares) and Ku-only (purple circles) populations. (c) Ratio as a function of sub-satellite latitude. (d) Ratio as a function of frequency, for DTC only. Channels with fewer than 30 illuminated or 30 eclipsed detections are omitted; in particular, $230.469$ MHz is absent because no eclipsed detections were recorded in this channel during the observing campaign. Error bars show 95% bootstrap confidence intervals; the horizontal line at unity marks the null hypothesis.

Panel (c) shows the same comparison as a function of sub-satellite latitude. The analysed sample is concentrated in a narrow latitude strip, but the DTC ratio remains below unity in every latitude bin, which argues against a geographic selection effect coupled to the day/night cycle through the local solar time. The bin centred at $-24.3^\circ$ contains the smallest number of eclipsed DTC detections ( $N_\mathrm{ecl}=286$ , against 414–583 in the other bins) and is correspondingly the noisiest, with a ratio of approximately $0.85$ ; the four other bins lie between $0.45$ and $0.55$ . Panel (d) shows the DTC ratio as a function of the coarse frequency channel; the reversal is preserved in the majority of channels where the sample size permits a measurement, although a small number of channels show ratios consistent with unity.

A complementary robustness check resamples at the satellite level rather than at the detection level: bootstrapping over the 175 DTC satellites with replacement and re-evaluating the population ratio of medians on each resample yields a 95% confidence interval of $[0.338, 0.627]$ , wider than the detection-level interval $[0.439, 0.492]$ as expected from the smaller effective sample size, but still entirely below unity. Table 6 summarises the eclipse-state ratios across all populations and sample variants considered in this section.

The stratified comparisons in Figure 12(b)–(d) are reported as descriptive robustness checks rather than independent hypothesis tests, and we do not apply additional multiple-comparisons control to them; the primary uncertainty summary is the 95% satellite-cluster bootstrap interval for each pooled eclipse ratio and their ratio of ratios.

A potential confound is that the DTC and Ku-only populations were deployed in different launch epochs, so the reversal could in principle reflect satellite age or commissioning state rather than the DTC payload. To test this, we restrict the Ku-only comparison to the 998 satellites whose launch dates fall within the 2024 January–October window spanned by the 175 DTC satellites, discarding the 625 earlier-launched Ku-only satellites that have no DTC counterpart. The eclipse-state ratio of the matched Ku-only sub-sample is $1.188\,[1.169, 1.209]$ , statistically indistinguishable from the full Ku-only ratio of $1.184\,[1.172, 1.197]$ , while the DTC ratio is unchanged at $0.465\,[0.438, 0.494]$ by construction. The two matched-sample confidence intervals remain disjoint. The unchanged matched Ku-only result argues against launch epoch as the sole explanation, although matching cannot remove every residual population-composition difference.

The reversal also survives a shared-systematics argument. The DTC and Ku-only populations are processed through identical EDA2 calibration, ephemeris, and flux-density pipelines. The opposite pooled trends are therefore difficult to attribute to a simple pipeline artefact common to both classes, but this argument does not remove differences in population composition, observing schedule, or detection selection.

The detection-level confidence intervals reported above treat individual detections as approximately independent, which is violated by the satellite/pass/frequency clustering of the EDA2 catalogue. A cluster-robust restatement resamples at the satellite level ( $N_\mathrm{sat,\,DTC} = 175$ , $N_\mathrm{sat,\,Ku} = 1\,623$ ), pools detections within each resample, and recomputes the illuminated/eclipsed median ratio. The cluster bootstrap yields a DTC ratio of $0.465\,[0.338, 0.627]$ and a Ku-only ratio of $1.184\,[1.115, 1.257]$ , both consistent in direction and central value with the detection-level estimates but with intervals widened to reflect the within-satellite correlation. A joint cluster bootstrap of the difference, drawing the DTC and Ku-only satellite samples independently, gives $(\mathrm{ratio}_\mathrm{DTC} - \mathrm{ratio}_\mathrm{Ku}) = -0.72\,[-0.87, -0.54]$ and a ratio of ratios of $0.39\,[0.28, 0.53]$ . These percentile intervals quantify uncertainty in the pooled detection-level estimator under satellite resampling; they are not a direct estimate of a universal within-satellite response. We use the cluster intervals as the primary uncertainty summary.

As a simple arithmetic illustration, combining the catalogue’s $75.5\%/24.5\%$ detected-event split with the pooled flux ratio gives a factor of approximately $1.28$ relative to an illuminated-only value. The catalogue does not contain the nondetection exposure needed to verify proportional sampling of the two states, so this number is neither a calibrated correction nor an operational prediction for SKA-Low.

The frequency-dependent decomposition in Figure 12(d) shows that the DTC reversal is not concentrated in any single channel. The reversal direction is preserved at the four highest-N channels in the sample ( $150.78$ , $153.12$ , $161.72$ , and $170.31$ MHz), all of which lie within the LOFAR HBA window where Bassa et al. (Reference Bassa, Di Vruno, Winkel, Jozsa, Brentjens and Zhang2024) reported strong broadband v2-Mini UEMR. The reversal is therefore a population-level property of the DTC payload class across the dominant emission band of the catalogue, rather than a feature of a single anomalous channel; in particular, it is not driven by the $230.469$ MHz polarisation excess discussed in Section 3.2.

Table 5.

Per-altitude-bin DTC illuminated/eclipsed range-corrected median flux ratio, complementing Figure 12(b). Bins above 412 km contain fewer than five detections in each state and are omitted.

Table 5. Long description

A table comparing illuminated and eclipsed flux ratios across two altitude bins. The table has two rows and four columns. The columns are labeled Altitude bin (km), N illum, N eclips, and Illum/eclips ratio. The row labels are 300-356 and 356-412. Row 1: 300-356, 4998, 2076, 0.542. Row 2: 356-412, 2545, 117, 0.173.

Notes: Both populated bins yield a ratio well below unity, consistent with the population-level DTC reversal reported in Section 3.6. Bins above 412 km are sparsely populated in the present sample.

Table 6.

Illuminated/eclipsed range-corrected flux density ratios for the four populations and sample variants considered in Section 3.6. Confidence intervals are detection-level bootstrap with $B=2\,000$ resamplings unless otherwise noted.

Table 6. Long description

A table with six rows and five columns comparing illuminated and eclipsed range-corrected flux density ratios. The columns are labeled Population, N sat, illum / N sat, ecl, N det, illum / N det, ecl, and Illum/ecl ratio. The rows are labeled Pooled V2-Mini, Full DTC, Full Ku-only, Matched Ku-only, DTC, satellite-bootstrap, and Ku-only, satellite-bootstrap. Each row provides specific values for the different populations and sample variants. The table includes confidence intervals and notes on resampling methods.

Notes. Matched Ku-only restricts the comparison to the 998 Ku-only satellites launched within the 2024-01-03–2024-10-18 window spanned by the DTC sample; discarding the 625 earlier-launched Ku-only satellites. The satellite-bootstrap row resamples DTC satellites with replacement rather than detections, yielding a wider but still entirely sub-unity interval.

Figure 13.

Summary of the ratio estimates reported in Section 3. The DTC/Ku point is the ratio of per-satellite median range-corrected flux densities, with a 95% bootstrap interval. The DTC and Ku-only illuminated/eclipsed ratios and their ratio of ratios use 95% satellite-cluster bootstrap intervals. The vertical line at unity denotes equal ratios.

A further internal consistency check examines the per-satellite distribution rather than the population aggregate. Among the 24 DTC satellites with at least five detections in each illumination state, the median per-satellite illuminated/eclipsed ratio is $0.886\,[0.441, 1.598]$ (95% cluster bootstrap, $N=24$ ); the interquartile range (IQR) $[0.42, 1.62]$ is asymmetric about unity, with the lower quartile below unity by a larger margin than the upper quartile is above. Fourteen of the 24 satellites (58%) lie below unity and ten (42%) below $0.75$ . The per-satellite distribution is directionally similar to the detection-level reversal but the per-satellite cluster bootstrap interval crosses unity, reflecting the modest sample size of DTC satellites with reliable two-state coverage; the central value is several times closer to unity than the detection-level $0.465\,[0.439,0.492]$ estimate. The paired subset therefore does not independently establish a within-satellite illumination effect, and the pooled ratio should not be read as a universal per-satellite response.

Taken together, Figure 12(a)–(d) and the matched-launch test above show that the population-level DTC eclipse-state contrast persists descriptively under stratification by altitude, latitude, frequency, and launch epoch. The paired per-satellite subset is directionally similar but inconclusive, and its ratios show substantial scatter ( $\mathrm{IQR}=[0.42, 1.62]$ for the 24 satellites with at least five detections in each illumination state). Figure 13 summarises the effect sizes and 95% confidence intervals for the comparisons reported in this section. The eclipse-state reversal is the only effect in which the DTC and Ku-only comparison populations depend on an external parameter with opposite signs.