Main

The first results from the James Webb Space Telescope (JWST) have raised more questions than answers about the prevalence, growth and impact of accreting supermassive black holes (SMBHs; namely active galactic nuclei or AGNs) in the first billion years of the Universe. Puzzlingly, a multitude of JWST spectroscopic surveys have found a significant overabundance of broad-line AGNs (BL AGNs) that is two orders of magnitude above predictions at z > 4. These AGNs were selected by their broadened Balmer emission indicative of gas moving at >1,000 km s−1 (refs. 1,2,3,4,5). Additionally, a subset of these z ≈ 5 JWST BL AGNs are characterized by their extreme compactness (<100 pc) and peculiar ‘V-shaped’ spectral energy distributions (SEDs) characterized by a puzzling ultraviolet (UV) excess alongside red colours recorded by the near-infrared camera (NIRCam) onboard JWST (m277 − m444 > 1.5, using magnitudes from NIRCam filters F277W, F444W)6. These are, therefore, dubbed ‘little red dots’ (LRDs)1,4,7,8,9,10,11,12. Although the full-width at half-maximum (FWHM) of their broadened components is generally narrower than the median FWHM found for bona fide AGNs at lower redshifts (FWHM Hα ≈ 2,500 km s−1 (ref. 13) versus 1,000 km s−1 < FWHM Hα < 2,000 km s−1), the relative dearth of conclusive star-formation signatures points to an accreting SMBH being the probable cause of the rapidly moving gas.

Several outstanding questions have arisen in attempts to rectify the tension between studies conducted pre- and post-JWST launch. A difficulty in placing JWST spectroscopically selected BL AGNs and photometrically selected LRDs is the surprising differences between the multi-wavelength observations of these sources and pre-JWST AGNs at similar epochs. Despite expectations, many of the BL AGNs have not been detected in X-rays14,15,16, have overall red UV to optical colours with a puzzling blue excess8,11,12 or fall short of predictions for their AGN or host-galaxy flux assuming obscuration in the rest-frame near-infrared (NIR) to mid-infrared and submillimetre12,17,18,19. This has led to numerous works studying how differences in either AGN properties or host-galaxy properties could explain why observations of these newly discovered sources lack certain telltale signatures of SMBH growth. Some recent studies propose that different dust distributions within these AGN host galaxies could explain the lack of infrared emission20. Some works indicate that different cloud properties in the broad-line region (BLR)14 or different accretion properties16,21,22 could explain the lack of X-ray detection. Some works have even claimed that unconstrained contributions to the wavelengths from star formation could explain these sources without the need for an AGN at all12,17,23.

Additionally, over 70% of z ≈ 5 JWST BL AGNs, including spectroscopically confirmed LRDs, have black hole (BH) masses estimated via local scaling relations24 that are at least an order of magnitude above their predicted BH masses, with many being over two orders of magnitude larger2,11. The implied extremely rapid growth of BHs is challenging to understand. The high ratios of the estimated BH-to-stellar mass of JWST-discovered AGNs alongside the surprisingly massive BHs of very luminous AGNs discovered pre-JWST has led to an, at times, controversial emerging picture of early BH growth: the UV or optically luminous BHs that we are observing in the early Universe start out supermassive (‘massive seeds’) and eventually their host galaxy accretes enough material or undergoes a sufficient number of mergers to match local BH-to-stellar mass relations9,25. This has led to increased speculation that these overabundant and over-massive BHs are evidence of efficient direct-collapse BH formation in the early Universe25,26,27. The masses of these BHs at the time of their observation are large enough, whereas smaller BH seeds would simply not have enough time to accrete sufficient material via moderate to low levels of accretion (λEdd < 1, where λEdd is the Eddington ratio).

This then leads to the question: what if they were accreting at much higher rates? This scenario, although a common assumption or feature in largely pre-JWST theoretical and simulated studies of early BH growth, was assumed not to play a large role for these newly discovered JWST BL AGNs28. This is largely because the estimated bolometric luminosities and BH masses, particularly from the deepest spectroscopic surveys, infer Eddington ratios below unity, thus implying that the mode of accretion would be described well by sub-Eddington models2. Although JWST probes the direct emission from the accretion disk via the rest-frame UV–optical emission, under the conventional models of thin-disk accretion, these sources should also be powerful X-ray emitters14,15,16,29. The connection between the UV–optical accretion-disk luminosity and the X-ray-emitting region is a powerful probe of the physics governing the entire central engine30,31. Pre-JWST studies found a notably tight correlation between the X-ray-emitting corona and UV–optical accretion-disk emission (αOX) for most UV- and X-ray-selected AGNs (for example, ref. 30). Yet, there has, at present, been no robust detection of any X-ray emission for any z > 5 spectroscopically confirmed JWST-selected BL AGN with anomalous SED shapes1,2,4.

Some recent studies have postulated that the lack of X-ray emission is due either to (1) these sources not being powered by SMBH accretion at all12,15,25 or (2) the physical BLR clouds themselves obscuring the X-ray corona14. However, other recent studies have begun to challenge the robustness of the BH mass estimates used to determine λEdd in the first place32. Thus, more nuanced approaches are being employed that challenge our assumptions on early BH growth by implicitly testing whether we are under-accounting for accretion, regardless of the reported BH masses and bolometric luminosities. Reference 33 employs a semi-analytic model to show how higher accretion rates and lower BH masses are preferred, and they were even able to recover the rest-optical continuum region and broad Hα properties of a few candidate JWST BL AGNs. This is complemented by simulation-based approaches, which have found that colour-selected LRDs are potentially characterized by high Eddington ratios34. By using X-ray stacking, other studies have challenged the published λEdd values of these new JWST AGNs. In ref. 16, the significantly weak X-ray detections associated with stacking a diverse sample of the LRD subpopulation was explained by associating the relative weakness of X-ray emission predicted in some models of super-Eddington accretion.

In this study, we aim to concretely determine whether these sources are accreting above the Eddington limit by self-consistently connecting the inconsistencies between predictions and observations across the entirety of the X-ray to NIR regime. We begin by assembling a sample of spectroscopically confirmed, serendipitously discovered, BL AGN candidates that have been identified in the literature to have anomalous rest-UV–optical SEDs and are covered by deep X-ray observations. We select from public data a sample of 14 BL AGNs with 4 < z < 7 that have been previously published and have complete JWST spectral coverage from the rest UV to optical. Further details on source provenance and selection criteria are given in Methods.

A surprising lack of X-ray continuum and UV line emission

We first determined by what margin these sources should have been detected in their respective X-ray surveys assuming the standard sub-Eddington accretion prescription. Assuming that these JWST sources are sub-Eddington accretors (as has been done in the discovery papers and subsequent studies on these sources1,2), we employed the widely used empirically derived relation between the 2-keV and 2,500-Å luminosity emission, parameterized as αOX (ref. 30). Intriguingly, we measured X-ray upper limits that are at least 5σ below the predicted αOX relation (median ΔαOX = −0.64). As seen in both panels of Fig. 1, the upper limits are well below the scatter of both the local and high-redshift AGN population. Further details on the αOX parameterization and X-ray, UV and optical data reduction are given in Methods and Extended Data Table 1.

Fig. 1: Significant X-ray weakness.
Fig. 1: Significant X-ray weakness.
Full size image

Left: the upper limits of αOX for the JWST z ≈ 5 BL AGN sample (purple triangles). The blue contours are the spectroscopically selected BL AGN sample derived from the Sloan Digital Sky Survey and confirmed by ref. 30. The blue solid line shows the αOX relation parameterized by ref. 30. The dashed-dotted, dashed and dotted lines show the 1σ, 3σ and 5σ scatter, respectively. The blue points are the high-z type 1 BL AGN sample (z > 3) from ref. 92. The orange points are candidate super-Eddington sources from the WISSH QSO sample93. Right: the offset from the ref. 30 relation (ΔαOX). The colours are consistent with those in the left-hand panel. The red dashed line is the canonical X-ray weakness threshold (ΔαOX < 0.3, as shown in ref. 38).

Steeper X-ray power-law slopes are needed to explain non-detections

The link between accretion rate (\(\dot{M}\)) and the power-law slope of the 0.5–10-keV X-ray spectrum (Γ) has been well studied over the past decades35,36,37. The AGN population with the strongest evidence of a significant relation between Γ and λEdd are those with high accretion rates22,37. The accretion rate is tightly connected to λEdd, where λEdd = LBol/LEdd, and LBol and LEdd are the bolometric and Eddington luminosity, respectively. Thus, λEdd is proportional to \(\dot{M}\)/\({M}_{\mathrm{BH}}\propto \dot{M}\)/\({\dot{M}}_{\mathrm{Edd}}\), where MBH is the BH mass and ṀEdd is the Eddington mass accretion rate. Numerous studies have found evidence that sources with high λEdd tend to have steeper X-ray slopes (Γ > 2)13. The most standard explanation that connects sources with a higher accretion rate with steeper Γ is related to the different accretion flows predicted for these sources. The lack of measured X-ray emission is usually ascribed to an intrinsic weakness in the X-ray emission (for example, photon-trapping when the diffuse timescale for photons to escape from the surface of a thick disk may be longer than the timescale for photons to be advected into the central BH) or due to shielding, both of which have been found to be applicable mechanisms in super-Eddington accretion models. For the epochs being probed in this study, the observed-frame 2-keV energies correspond to rest-frame X-ray energies that are >12 keV. Thus, any steepening of the power law from canonical assumptions severely curtails the detectability of these sources in X-rays.

Although we do not have any X-ray detections to measure Γ directly, we calculated the smallest Γ required such that the X-ray emission is not detected within the given X-ray observation of each source, assuming that the predicted αOX value is correct. We note that the empirical αOX versus L2500Å relation is contextualized implicitly under models of sub-Eddington accretion and may be intrinsically different for super-Eddington sources38. We limited this impact by using L2500Å values that are derived redwards of 2,500 Å, and thus, in a portion of the SED that is predicted to be more like sub-Eddington accretion models as opposed to the rest-UV and bluer39. We discuss this more in section ‘Estimating the UV luminosity’. From the measured L2500Å and the canonical αOX relation30, we used the corresponding monochromatic 2-keV luminosity L2keV predicted for these sources to find the minimum Γ needed to correspond with the upper-limit X-ray flux of each detection. Further details are given in section ‘The connection between Γ and M’. We found a median Γ > 3, and using these values, we estimated the lower limit on the corresponding λEdd as parameterized in ref. 37, yielding a median λEdd > 2.6, which surpasses the Eddington limit. As detailed in section ‘The connection between Γ and MBH’ and shown in Fig. 2, we highlight how these new limits reduce the tension between BH and stellar masses.

Fig. 2: Lower than predicted MBH.
Fig. 2: Lower than predicted MBH.
Full size image

The dark purple points are the values from ref. 2. The MBH errors include both the propagated uncertainty of the measured variable and the scatter of the virial MBH scaling relations. The light orange and brown lines correspond to the relation between BH and stellar mass for a sample of bulge-dominated AGNs and the total sample of local AGNs, respectively. The width corresponds to the scatter of each relation from ref. 24. The green line is the BH-to-stellar mass relation derived from ref. 94. The light purple triangles are the upper limits of the BH mass of our sample that overlaps with ref. 2, derived from the upper limits of the X-ray power slope Γ. These BH mass upper limits are, on average, an order of magnitude below those in ref. 2.

Intrinsically missing rest-UV ionization lines

To concretely test whether these sources are candidate super-Eddington accretors, we must also inspect other portions of their SEDs that would be impacted by the intrinsic reduction in X-ray photons, namely, tracers of high ionization in the rest UV. Although all these sources are well detected (>5σ) in their respective rest-UV imaging (<2 μm), the spectral rest-UV features will be the most discriminating between competing AGN accretion models. On examining the G140M microshutter-assembly spectra recorded by the near-infrared spectrograph (NIRSpec) onboard JWST for the subset of our sample (seven sources) covered by the JWST Advanced Deep Extragalactic Survey (JADES; Figs. 3 and 4 and Extended Data Tables 2 and 3), we did not find a single source with any detectable UV line emission, including C iv and C iii], which, aside from Lyα, comprise some of the brightest lines in the rest UV at these epochs. In particular, some studies have indicated that the rest-UV SEDs of these sources could be wholly explained through star formation10. The lack of C iii], a prominent rest-UV line in low-metallicity star-forming systems40,41,42,43, disfavours a star-formation-only interpretation. Furthermore, if obscuration is a factor in the dearth of rest-UV lines, explaining how the much fainter continuum is detected—while the brighter emission lines are not—is difficult. If the continuum is being driven by AGN emission from the accretion disk, the high-ionization rest-UV lines, which are excited in regions beyond the accretion disk, would not be obscured whereas the rest-UV accretion-disk emission would also not be obscured. We further explore the impact of obscuration on line detectability in section ‘Missing X-ray and UV emission under a super-Eddington accretion paradigm’.

Fig. 3: JWST/NIRSpec medium-resolution (R ≈ 1,000) spectra for the seven sources with G140M coverage.
Fig. 3: JWST/NIRSpec medium-resolution (R ≈ 1,000) spectra for the seven sources with G140M coverage.
Full size image

For each galaxy, the three subpanels show zoomed-in regions centred on the expected locations of the C iv, Hβ + [O iii] and Hα + [N ii] emission lines, respectively. The high-ionization C iv emission line was not detected in any source that has wavelength coverage. The spectra are expressed in flux density per unit frequency (Fν). Details can be found in Methods.

Fig. 4: Fits to the Hα + [N ii] emission lines for eight sources with NIRSpec/G395M observations in order of increasing redshift.
Fig. 4: Fits to the Hα + [N ii] emission lines for eight sources with NIRSpec/G395M observations in order of increasing redshift.
Full size image

a–h, MPT ID 11836 (a), MPT ID 8083 (b), MPT ID 20621 (c), MPT ID 62309 (d), MPT ID 77652 (e), MPT ID 3608 (f), MPT ID 1093 (g), MPT ID 61888 (h). Each is fitted with a combination of four Gaussians: narrow components to Hα (light green) and both [N ii] lines (blue and red), plus a broad component to the Hα line (dark green). The best-fitting parameters for each individual Gaussian are shown on the top right of each plot, and the combined fit (purple) values are shown on the left of each plot. The spectra are expressed in units of flux density per unit wavelength (Fλ). We also annotate on each panel the model continuum flux (Cont.) and Bayesian information criterion (BIC). Details are described in Methods, and values can be found in Extended Data Table 3.

Assuming that obscuration is the unlikely culprit of the missing UV lines, we turn to explanations that can intrinsically impact the overall properties of the ionizing continuum. In accretion-disk models that predict a radiatively inefficient hot corona, the reduced availability of seed photons has been found to have an effect on the nebular regions surrounding the central engine, namely, the higher ionization potential lines in the BLR44,45. This indicates that the broad-line-emitting gas is incident with an unusually weak photoionizing continuum. Some studies have found that the shape of the ionizing continua in these sources will lead to lower-ionization potential lines, such as the Balmer series, so that there is a less obvious discrepancy from standard accretion models compared with the higher-ionization potential lines, such as C iv λ1549 Å. Perhaps most relevant to the JWST BL AGNs of this study is a similar class of AGN accretors known as weak-line quasars (WLQs)39,45,46,47. Pre-JWST, WLQs were a small fraction of the type 1 quasi-stellar object (QSO) population. They are defined by their weak UV high-ionization emission lines; for example, the rest-frame equivalent width of C iv <10 Å (refs. 48,49). A substantial fraction of WLQs (~50%) are X-ray-weak compared with their typical luminous AGN counterparts (≤6%). In fact, the empirical αOX of these sources indicates that a significant fraction of the X-ray WLQ population have X-ray fluxes that are mainly undetected or at least a factor of 6 below their expected values50. Also note that the X-ray weakness was similarly derived with an assumed Γ ≈ 2.

Missing X-ray and UV emission under a super-Eddington accretion paradigm

Observational evidence of the Eddington limit being exceeded is critical in informing the demographics of early Universe BHs. For one, a wide variety of analytical, theoretical and simulation studies have found that bursts of super-Eddington accretion are able to reduce the need for efficient massive-BH seed production or have found that the conditions required to produce super-Eddington accretion are achievable in the early Universe28,51,52,53,54,55,56,57. The need for different accretion models to describe high-Eddington fraction sources has been known for over 30 years, with many models based on a slim accretion disk or an optically thick, geometrically thicker-than-standard disk with a quasi-Keplerian accretion flow58,59,60. These models predict super-Eddington flows that are much less radiatively efficient than standard sub-Eddington disks, as the emitted flux should saturate at the local Eddington limit and the remaining power would be lost through radial advection or winds.

On considering the lack of X-ray detection and the lack of high-ionization UV lines but the significant broadening of the Hα line component, we constructed a model to self-consistently explain these observations under the framework of super-Eddington accretion. For WLQs, some physical interpretations include their BLR gas being shielded by a column of gas between the inner-accretion disk or X-ray corona and the BLR or an intrinsic X-ray weakness because the corona fails to fully form44,48,61. Both of these scenarios are usually attributed to the changes in accretion flow that are expected at higher accretion rates. As previously mentioned, these models account for the inner edge of an optically and geometrically thick ‘slim’ disk59,62. This configuration is expected to shield or not intrinsically produce the X-ray and extreme-UV radiation incident on the BLR gas, thus softening the incident ionizing continuum.

Owing to the lack of detection in both UV line emission and X-ray, which precludes a fit of the sources to models of super-Eddington accretion, we instead consider whether the relative levels of X-ray emission and NIR emission are consistent with a super-Eddington SED shape. For our approach, we used agnslim, the slim-disk model available through XSPEC60, which adopts a slim-disk emissivity so that the surface luminosity is kept at the local Eddington limit within a critical radius. For high-Eddington accretion, the advected flux yields lower fluxes emitted in the UV continuum, and thus, there is an intrinsic reddening of the UV–optical slope. For comparison, we also used the available model in XSPEC, agnsed, to represent a canonical sub-Eddington source with a 45° inclination of the disk63. For the sub-Eddington case, we constructed the SED to reflect how these JWST BL AGNs have been previously interpreted in the literature, namely, as AGNs with low to modest accretion rates and MBH ≈ 107 M⊙ (refs. 1,2). We limited our model comparison to the observed data for the subset of sources with G140W observations, and thus, we set the parameters of the sub-Eddington SED model with the median MBH and LBol reported in ref. 2.

As we cannot directly estimate LBol for these sources under the super-Eddington prescription due to lack of detections, we illustrate the relative difference in the predicted X-ray–UV properties between sub- and super-Eddington models. For the super-Eddington case, we normalized the slim-disk model to the sub-Eddington model at 4,000 Å, a region of the wavelength space less impacted by the softer ionizing continuum than the rest UV and still easily comparable with other wavelength bolometric tracers in the literature64. Additionally, we configured the parameters of the super-Eddington model with values for the electron temperatures, the radii for warm and hot comptonization, and the outer disk radii so that they are consistent with other WLQs with X-ray weakness described in the literature45. We set the mass transfer rate relative to λEdd = 1, MBH = 106 M⊙ and maximal spin. The value of MBH was driven by the upper limits found in the X-ray non-detection (section ‘The connection between Γ and MBH’). The general assumption that for these given sources to be accreting at higher rates with the same inferred bolometric luminosity derived from the measured Hα flux necessitates a lower MBH than reported by refs. 1,2. Maximal spin is expected for sources accreting at rates of Eddington or above65. Figure 5 shows the models as a function of energy.

Fig. 5: Input SED models for Cloudy normalized by \({{\mathcal{Q}}}_{{\rm{H}}}\).
Fig. 5: Input SED models for Cloudy normalized by $${{\mathcal{Q}}}_{{\rm{H}}}$$ Q H .
Full size image

The solid blue line is the slim-disk prescription and the dashed line is a radiatively efficient, sub-Eddington prescription. The light blue box corresponds to the UV–extreme-UV regime, and the darker blue box the X-ray regime.

We used these two models as input to Cloudy v17.0266 (section ‘Signatures of Sub-Eddington and Super-Eddington accretion from Cloudy simulations’) to assess the impact of these different accretion models on the simulated Hα flux. We found for both the super-Eddington and sub-Eddington cases values consistent with the range of measured Hα fluxes for these sources. The fiducial models for the super-Eddington and sub-Eddington cases yield a mean \({L}_{\mathrm{H\alpha }}\approx 5\times 1{0}^{41}\) erg s−1 and \({L}_{\mathrm{H\alpha }}{\approx }3\times 1{0}^{42}\) erg s−1, respectively, which are within the range of the observed broad \({L}_{\mathrm{H\alpha }}\) values reported for the sample2. As shown in Fig. 6, we found that the C iv, He ii and C iii] line emission is significantly suppressed for slim-disk accretion compared with the canonical sub-Eddington prescription. The line detection limits represent the median 3σ line depth at the given line of all the G140M observations with a dust law applied67. We used the median published values for the G140M observed subsample. Although some dust curves, including that for the Small Magellanic Cloud68, can account for the attenuation of UV line emission at these AV’s, those same curves would also preclude the UV-continuum detections observed for these samples2. Thus, even when accounting for obscuration, the sub-Eddington model predicts that we should be able to detect C iv, He ii and C iii] well above the detection threshold. Even for the maximum published AV derived for a source in our sample (AV ≈ 1), the difference between the predicted dust-free C iv line luminosity in the sub-Eddington case is over an order of magnitude above the upper limit of the line detection.

Fig. 6: Predicted line luminosities L(line) for C iv, He ii and C iii].
Fig. 6: Predicted line luminosities L(line) for C iv, He ii and C iii].
Full size image

Dark blue symbols show the sub-Eddington prescription (agnsed), and light purple symbols show the slim-disk prescription (agnslim). Orange triangles are median 3σ upper limits from the JWST/NIRSpec G140M/G235M spectra.

The intrinsic reddening of the X-ray to optical SED of super-Eddington sources as predicted by this model (section ‘Signatures of Sub-Eddington and Super-Eddington accretion from Cloudy simulations’) has profound implications for the interpretation of the rest-UV to optical spectra and photometry as measured by JWST. Studies that assumed these sources were sub-Eddington accretors via the measured λEdd would find that the spectrophotometric fitting of these sources with canonical QSO templates would attribute the dearth of UV emission to dust obscuration. Additionally, JWST BL AGNs that would also be photometrically selected as LRDs have Balmer decrement values that, under the assumption of case B recombination, would indicate substantial dust attenuation9. As a proof of concept, we measured the Balmer decrement between our fiducial sub-Eddington and super-Eddington accretion models. We found for the sub-Eddington model Hα/Hβ ≈ 3.5 and for the super-Eddington model Hα/Hβ ≈ 8.5. The sub-Eddington model is within the range of normal Balmer decrements of unobscured sources69, yet in the super-Eddington case, it is well above the expected ratio. It is generally known that the much higher densities of the BLR (compared with the narrow-line or H II regions) can lead to different Balmer decrements due to line optical depths and collisional effects that are not included in conventional case B recombination calculations70,71. This is potentially exacerbated in the super-Eddington case due to the intrinsically weaker ionizing continuum. Thus, although the models do not include dust extinction, a natural consequence of an intrinsically reddened AGN SED would be higher observed Balmer decrements.

If, indeed, these sources are all super-Eddington accretors, we can estimate the duty cycle. Using the seven sources with rest-UV coverage, we computed the number density over the JADES area and compared the number of non-AGN galaxies with similar MUV ≈ −18.85[AB] at the median redshift of this sample (z ≈ 5)72. We found an upper limit of <5% for the duty cycle. Interestingly, ref. 73 uses a semi-analytic model to attempt to recover the observed luminosity functions of JWST-discovered AGNs at z > 4, finding that they required a super-Eddington phase that corresponds to a duty cycle between 0.5% and 4%, which is consistent with our results.

Conclusion

In summary, models of accretion that generate a relative decrease in the extreme-UV ionizing continuum, as is predicted by some models of super-Eddington accretion, decrease the tension between the multi-wavelength properties of high-z JWST BL AGNs. These models are able to self-consistently account for (1) the lack of X-ray detection despite sufficiently deep observations and (2) the lack of UV high-ionization line detection, such as C iv, without the need for high attenuation. Thus, if the true accretion rates of these JWST BL AGNs are higher than what is being inferred via the measured λEdd, then that potentially implies that the BH masses of these sources measured via the FWHM of the broad component of the Hα emission are being overestimated (as λEdd ∝ LBol/LEdd, where LEdd ∝ MBH). The outstanding question is what fraction of the total LRD or serendipitously selected JWST BL AGN population show these hallmarks of super-Eddington accretion? Thus, applying rest-UV colour thresholds and obtaining a deep rest-UV spectroscopic follow-up is critical for a more extensive study of the entire JWST-discovered AGN population.

We also note that in ref. 2 and ref. 1, the bolometric luminosities of the JWST BL AGNs were determined by applying a bolometric correction to the luminosity of the broadened Hα line. As shown in Methods, normalizing the super-Eddington and sub-Eddington accretion models to produce equivalent amounts of Hα flux would yield significantly different bolometric corrections across the entirety of the multi-wavelength SED. This would critically impact attempts to accurately predict the amount of rest-frame NIR and mid-infrared flux expected for these sources. For instance, almost every LRD observed in the rest-frame 2–5-μm range has significantly less flux than predicted from either energy balance arguments or applying bolometric corrections12,18,20. Future work will apply these accretion models in the context of the measurements of these sources at redder wavelengths.

These results imply that high-Eddington accretors may be more common than presently assumed via observations. Each serendipitous BL AGN found in both JADES and the EIGER+FRESCO survey is consistent with higher Eddington accretion rates, as parameterized by our models. This has a fundamental impact on how we model the growth of the first SMBHs and leaves wide open many avenues of inquiry on connecting the early Universe environment to BH accretion.

Methods

Parent sample

We drew from published samples built from the extensive JWST observations that comprise JADES (JWST guaranteed time observation IDs 1180, 1181, 1210 and 1286, PIs Eisenstein and Luetzgendorf; refs. 74,75,76) and the First Reionization Epoch Spectroscopically Complete Observations (FRESCO) survey (JWST general observer ID 1895, PI Oesch; ref. 77), which are within the same observational fields of the deepest Chandra observations ever performed (GOODS-North: CDFN 2Ms78 and GOODS-South: CDFS 7Ms79). The parent AGN samples that we selected from are presented in ref. 2 and ref. 1. In addition to deep X-ray observations, we required these sources to be (1) at redshifts z > 4 to ensure complete photometric coverage from the rest-UV to optical wavelengths, (2) previously characterized as an AGN via a statistically significant (>5σ) broad component (>1,000 km s−1) in Hα and (3) photometrically characterized as compact with red colours in the JWST NIRCam >2-μm imaging bands and significant detections in the <2-μm JWST NIRCam imaging bands (for details, see refs. 1,2). These studies comprised some of the first discoveries of these enigmatic AGN candidates with the JWST/NIRSpec and NIRCam grism instruments. From these surveys, we applied our sample requirements and selected 14 rest-frame spectroscopically confirmed BL AGNs at 4 < z < 7, seven of which also have NIRSpec G140W medium-resolution spectroscopy, which we used to probe their rest-frame UV emission. We stress that these sources are most probably not representative of the total AGN demography at these epochs but are, rather, some of the most spectroscopically complete and pristine examples of a potentially new class of sources that required JWST for efficient discovery. We note that a more complete selection of AGNs at this epochs would require access to wavelengths that are less impacted by obscuration, which may be so significant it would preclude detection of the broad lines in the first place80,81.

α OX and X-ray upper limits

In the standard, sub-Eddington accretion paradigm, UV and optical photons from the accretion disk are linked to the X-ray continuum emission that arises via inverse Compton scattering, forming what is conventionally called a ‘corona’. The empirical αOX is the power-law slope that relates the X-ray-emitting corona (as probed by the monochromatic rest-frame 2-keV emission) to the accretion-disk luminosity (as probed by 2,500-Å emission) via αOX = \(-0.384\,\log\) L2keV/L2500Å. This relation has been found to hold for almost all unobscured or obscuration-corrected AGNs within 0.2-dex scatter, from local AGNs to the highest z QSOs discovered pre-JWST30,82,83. Importantly, at high-enough redshifts, the rest-frame energies being probed by most X-ray telescopes with sufficient sensitivity are much higher than 2 keV (at z = 5, Chandra coverage of 2-keV probes the rest-frame 12 keV). Thus, the X-ray power-law slope, nominally assumed to be Γ ≈ 2 for most AGNs at these epochs, is used to infer the rest-frame 2 keV in sources without sufficient counts to measure Γ directly.

W downloaded from the Chandra Source Catalog V2.1 the full-field combined event, arf and rmf stacks of every published Chandra observation that covers each JWST coordinate of the sources used in this sample. These full-field combined event files are stacked observation detections event files filtered by the appropriate science energy band. The event files used for the stacked detections event file have been reprocessed through acis_process_events to apply the latest instrument calibrations and the standard event status and event grade filters. Additionally, all observations within a given observation stack were aligned and reprojected such that they have a consistent coordinate system. The combined exposure maps were computed by applying the aspect histogram sampled at 0.5" resolution and were blocked by 1 in sky coordinates. Robust upper limits were estimated via the exposure-corrected 0.5–2-keV count rates within a 2" aperture. The monochromatic 2-keV flux upper limits were determined using the CIAO tools function aprates84 by assuming an X-ray power-law slope of Γ = 2 and the respective galactic absorbing column density (NGal = 8.8 × 1018 cm−2 or 1.1 × 1019 cm−2 for CDFN or CDFS, respectively). Using the redshifts listed in Extended Data Table 2, we converted these fluxes to the rest-frame monochromatic 2-keV luminosity. We report the 3σ upper limits in Extended Data Table 1. Finally, we stacked all the X-ray observations of our sources. Using the method outlined in ref. 16, which is based on the Chandra Stacking Tool (CSTACK) as described in ref. 85, we stacked them in three energy bins (0.5–2 keV, 2–8 keV and 0.5–8 keV) but did not find a detection in any bin. Specifically, we measured a stacked 0.5–2-keV upper limit of 2.14 × 10−18 erg s−1 cm−2 with 90% confidence.

The connection between Γ and M BH

As stated in the main text, empirical studies have found on average steeper Γ in sources accreting above the Eddington limit. To provide a lower limit on Γ, we must find how steep the power-law slope needs to be such that the monochromatic 2-keV luminosity predicted via αOX would not be detected within the depth of the corresponding X-ray surveys. Using αOX = \(-0.384\,\log\) L2keV/L2500Å and the estimated L2500Å, we found relatively bright L2keV emission expected for these sources under the conventional assumptions of sub-Eddington accretion. To convert these monochromatic luminosities to their respective predicted Chandra 0.5–2-keV fluxes, we simply repeated the method used to determine the X-ray upper limits but in reverse. Thus, using the conventional power law as our spectral input, we solved for the minimum Γ needed to reproduce the measured upper-limit Chandra 0.5–2-keV fluxes for each source.

We then used these Γ values to estimate the lower limit on λEdd using equation (8) in ref. 37. Finally, we estimated an upper limit for the BH mass that is constrained by our X-ray non-detections via λEdd ∝ LBol/MBH for the G140M subsample with confirmed broad-line features. As shown in Fig. 2, using the inferred bolometric luminosities of the sources in our sample1,2, our MBH upper limits are at least an order of magnitude below the literature BH masses for our sources. Extended Data Table 2 lists the associated limits on Γ and MBH, and Fig. 2 shows the difference between the previously published MBH versus the adjusted MBH when accounting for the estimated λEdd derived in this work. We note that the bolometric correction via Hα would be intrinsically different if the sources were accreting above super-Eddington, but in this exercise, we simply aimed to highlight that the assumption of uncertainty in robustly measuring MBH at these epochs coupled with anomalously weak X-ray emission can correspond to higher accretion rates, as constrained by the data.

Estimating the UV luminosity

The rest-frame L2500Å luminosities were estimated via the inferred bolometric luminosities derived from the flux of the broadened Hα component. For the JADES subsample, we used our rederived fluxes and compared the results with the values published in ref. 2. We found our values to be within the measurement error reported for each source, except for MPT ID 3608. We did not confirm a statistically significant broad component, and thus, we omitted it from the median values used when comparing with the photoionization model predictions. For the EIGER+FRESCO sample, we used values published by ref. 1. To be consistent with the literature referenced in our comparisons, we chose to infer the intrinsic L2500Å emission by probing the reddest available AGN power indicator. Bluer parts of the spectrum, which probe rest-frame UV–optical wavelengths, are more attenuated by levels of obscuration that are unaccounted for or have more extreme differences in their ionizing continuum under models of super-Eddington accretion. Using the bolometric luminosities inferred from the flux of the broadened Hα component, we applied a bolometric correction via the relation given by ref. 64 to estimate the 4000 Å luminosity (L4000Å) luminosity. Then assuming a canonical f = ν0.44 relation that describes the intrinsic continuum shape of powerful AGNs at these wavelengths, we inferred the L2500Å luminosity86.

JWST NIRSpec spectral extraction and line fitting

The subsample of our sources (eight objects) with G140M microshutter-assembly spectral coverage are within the JADES survey. We downloaded the fully reduced two-dimensional (2D) spectra of G140M, G235M and G395M published by the JADES collaborations and detailed in ref. 76. These 2D frames were visually inspected for artefacts, which were manually masked. We applied a further sigma clipping with a threshold of 10σ to the 2D spectra to remove any other spurious pixels. The one-dimensional (1D) spectrum of each source was obtained via an optimal extraction87 using a spatial weight profile derived from the trace of the source in the masked 2D spectrum, such that the pixels near the peak of the trace are maximally weighted. To create this extraction profile, we collapsed the 2D signal-to-noise (SNR) spectrum in the spectral direction, took the median value at each spectral pixel and fitted a Gaussian to the positive trace. For source MPT ID 1093, there was a second lower-redshift source dispersed into the G140M + G235M spectra due to nearby failed-open shutters during observation. This source prevented a conclusive fit to the profile of the target object in these filters. Therefore, we used the profile fit obtained from the G395M spectrum to effectively remove this contamination from our final spectrum. We then combined the spectra from each filter to produce a single spectrum per source covering the full wavelength range of observations. In the regions of wavelength overlap between gratings, we replaced the data from the filter with the lower resolution with the higher resolution data.

As this process removes the neighbouring contamination and uses a different extraction method, these 1D spectra differ from those released by the JADES team76 and from those used to identify these sources as BL AGNs by ref. 2. We, thus, performed our own measurements of the emission lines for consistency in the analysis for this paper. To do so, we ran the 1D spectra through an automated line-fitting routine originally detailed by ref. 88 and modified for JWST/NIRSpec spectra, as described by ref. 5. We redetermined the redshift of each source via the Hα line and identified the Hβ + [O iii] emission lines and the expected location of C iv. Figure 3 shows snippets of our extracted 1D spectra for each source in the sample. The locations of these lines of interest are marked. For each of the sources, our redshift measurement is in agreement with the published JADES redshifts2. We note that the JADES public NIRSpec catalogue76 includes contamination in the spectrum of MPT ID 1093, and emission lines from the interfering galaxy were misidentified as C iv emission from the galaxy of interest.

For each of our target sources we fitted the Hα + [N ii] emission line complex, as shown in Fig. 4, with a combination of four Gaussians: narrow components to Hα (light green) and both [N ii] lines (blue and red), plus a broad component to the Hα line (dark green). The best-fitting parameters for each individual Gaussian are shown on the top right of each plot and reported in Extended Data Table 3, and the combined fitting values are shown on the left of each plot (purple). The narrow component of the fits was restricted to within 30 km s−1 of the measured FWHM of the [O iii] λ4960 + 5008 lines (~300 km s−1). The same FWHM was used for all three narrow lines. In these fits, we fixed the ratio [N ii] λ6585/[N ii] λ6550 = 2.8 and restricted the peaks of each to be within 1 px (~18 Å) of the redshifted separation from the peak of the Hα emission line. The broad component of the Hα line was not fixed to the same wavelength as the narrow component, allowing for a velocity offset (Δv), and the FWHM was restricted to >3 times the narrow FWHM. In most cases, our measured broad FWHM matches that of those reported by ref. 2 with two notable exceptions. For MPT ID 3608, the broad component fit of the Hα line was inconclusive (Fig. 4b), having an FWHM of 1,761 ± 1,443 km s−1 and SNR = 1.56. The broad component fit of the Hα line for MPT ID 62309 (Fig. 4d) was not significant with SNR = 1.35 and a FWHM of 935 ±285 km s−1. Although MPT ID 3608 did make the pre-spectroscopic inspection selection criteria, as there was no conclusive measurement of a broad component, it was removed from our BL AGN sample for this paper (and, subsequently, not shown in Fig. 3). We have indicated so in Extended Data Table 3.

Once the redshift had been established, each spectrum was inspected at the expected location of C iv λ1548 for any significant emission features. With the exception of MPT ID 61888, where these wavelengths unfortunately fell in a gap in the data, none of the sources exhibited any significant (>3σ) features within 10 px (~60 Å) of the location of C iv. To get an accurate measurement of the upper limit, we forced a single Gaussian fit at the line location and used the error on that fit as the 1σ limit. Reported upper limits on the Civ emission are ×3 the median measured limits. This same method was used to obtain the 3σ upper limits for He ii λ1640 and C iii] λ1907,1909 for each source. The medians of these values for the full sample are shown in Fig. 6 as red triangles and listed in Extended Data Table 3.

Signatures of sub-Eddington and super-Eddington accretion from Cloudy simulations

For the XSPEC agnslim and agnsed models, we use radii, temperatures and power-law slopes of all components (hot and warm), as estimated for objects in the literature (for example, refs. 45,60,63) with similar bolometric luminosities (LBol ≈ 1044.5 erg s−1) BH masses (MBH = 106–107 M⊙) and accretion rates (log \(\dot{m}\) = 0 to −0.1) for the super- and sub-Eddington prescriptions, respectively. As already mentioned, we conservatively chose the maximal spin for agnslim. We note that assuming a higher BH spin for this model increases the relative amount of X-ray photons compared with a lower BH spin. Thus, as assuming that maximal spin is a conservative choice, any decrease in spin would further decrease the expected amount of X-ray photons. Finally, we note that these selections for the relative shape and normalization ensured that agnslim is also consistent with X-ray and UV-continuum upper limits, assuming negligible obscuration. We note that it is important to ensure that all parameters in the agnslim model are adjusted to be self-consistent with one another. As is seen in Extended Data Table 1 in ref. 60, the relative difference between the fitted values of the super-Eddington candidate RXJ 0439.6-5311 differ significantly from the default values of agnslim. Figure 5 shows the optical–X-ray SED of these physics models for both accretion-disk prescriptions normalized by \({{\mathcal{Q}}}_{{\rm{H}}}\). We note that the intrinsic shape of the slim-disk prescription for our model parameters yields a significant difference from UV to X-ray wavelengths.

We performed photoionization simulations using the agnslim and agnsed SEDs as input to Cloudy v17.0266. In setting the physical conditions in the cloud, we assumed values appropriate to the BLR (for example, ref. 89). Namely, we assumed an open geometry, a fixed inner radius R ≈ 5 × 1016 cm and a density nH = 1010 cm−3 at the illuminated face of the cloud. Constant gas pressure was assumed until reaching an effective hydrogen column density NH = 1023 cm−2, at which point the simulations stopped. We set the gas-phase abundance to 0.1 Z⊙ and used the abundance patterns and scalings from ref. 90, where solar is 12 + log(O/H) = 8.76 for the adopted abundance pattern. Although most of the powerful AGNs before JWST were, indeed, not extremely metal poor (generally, >30% of Z⊙) at these epochs, these peculiar sources are consistently below that. For the G140M subsample, all sources are shifted on the BPT diagram and overlap with metallicity models that range between 10% and 20% of Z⊙ (ref. 14). To aid with the use cases of these models, we opted for 10% to cover a broader range of similar sources at the redshifts of the sample and higher, as the average metallically drops. Finally, we note that dust grains were not included in the cloud.

For the agnsed model, we set the ionization parameter to reproduce the median inferred bolometric luminosity of the sample (~7 × 1044 erg s−1). This corresponds to the ionization parameter value of \(\log \,{\mathcal{U}}\) = −2.35, given the adopted radius and hydrogen density. Note that although we used the ionization parameter to define different ionized regions within the system, we did not base the run parameters of the models on a fixed ionizing flux. We took the radius into account by converting these values into the ionizing photon rate Q. We normalized the agnslim model to the agnsed model at 4,000 Å, such that the inferred bolometric luminosities of these sources via either a bolometric correction to the 4,000-Å continuum or the BLR Hα flux would yield a similar LBol from the observer’s perspective. In reality, the intrinsic bolometric luminosity of the slim-disk model would be significantly different, which, thus, highlights the need for bolometric corrections that are apt for high-accretion-rate systems. This relative normalization yields a slim-disk model with an ionizing photon rate \({{\mathcal{Q}}}_{{\rm{H}}}\) that is ~40 times lower than the ionizing photon rate for the agnsed model. As a result, the normalization for the agnslim model in Cloudy corresponds to \(\log \,{\mathcal{U}}\) = −4, which we note is smaller than the canonical BLR value of sub-Eddington sources.