Undermassive Hosts of $z = 4-6$ AGN from JWST/NIRCam Image Decomposition with CONGRESS, FRESCO, and JADES

Abstract: In the local Universe, supermassive black hole (SMBH) masses strongly correlate with their host-galaxies' stellar masses ($M_{<em>}$), but galaxies hosting faint AGN recently found by JWST may deviate from this relation. To constrain the M$<em>{\text{BH}}$-M$</em>{</em>}$ relation at high redshift, we performed AGN-host image decomposition for 17 low-luminosity AGN galaxies at $z$\,$\sim$\,4--6 using NIRCam images in the JADES GOODS-N field. These sources are identified as AGNs from broad H$α$ emission lines detected by the CONGRESS and FRESCO surveys. We used \textsc{galfit+MCMC} to fit spatial profiles in 7 wide-band images and detected extended emission in 9 sources out of 17. The close spatial alignment between the extended components and the AGN centers indicates that this emission likely originates from the host galaxies. These sources are extended at 0.9--2.0~$μ$m, suggesting significant host-galaxy light in the rest-frame UV. For the sources with the host detection, the stellar mass inferred based on image decomposition result can be 1-2 dex lower than the results without image decomposition. The BH-to-stellar mass ratio spans $M_{\text{BH}}/M_\ast$\,$\sim$\,0.01--1.48, placing them well above the local $M_{\text{BH}}$--$M_\ast$ relation. In contrast, the host-galaxy size--mass relation broadly agrees with previous measurements. Our results suggest that the host galaxies of these faint AGN are either genuinely under-massive compared to their black hole masses, or too compact to be spatially resolved.

• The paper utilizes NIRCam images from the JADES survey to perform AGN–host image decomposition for a sample of broad H$\a emitters at z \(\sim 4-6\), revealing significantly lower stellar-mass estimates of host galaxies by up to 1–2 orders of magnitude.
• Using GALFIT and MCMC, the study found that 9 out of 17 targets had confident extended emission, confirming robust AGN–host decomposition of faint AGN hosts at early epochs.
• The decomposed SEDs showed how photometry-based stellar mass estimates could be inflated by misattributing AGN light to stellar content, highlighting the importance of imaging-based decomposition for accurate stellar mass measurements.

Motivation and context

The local $M_{\rm BH}$--$M_*$ relation is widely interpreted as evidence of co-evolution between supermassive black holes (SMBHs) and their host galaxies, but its origin remains contested between AGN-feedback regulation and merger-driven statistical convergence. Testing these scenarios requires measuring the relation at early epochs, when black holes were in their growth phase. JWST has revealed a large population of UV-faint, broad-line AGN at $z > 5$—including the "little red dots" (LRDs)—whose number density at $z \sim 5$ exceeds quasar luminosity-function extrapolations by one to two orders of magnitude. These objects host moderate-mass black holes ($10^{6-8}\,M_\odot$), roughly two orders of magnitude below luminous quasars at comparable redshifts, making them well suited for constraining host-galaxy properties: their low AGN luminosities reduce host–AGN contrast relative to bright quasars.

This paper performs AGN–host image decomposition for a sample of broad H$\alpha$ emitters at $z \sim 4$–6 in GOODS-N, using JADES NIRCam imaging. The central question is whether the hosts are genuinely undermassive relative to their black holes, or whether previous stellar-mass estimates—based on total photometry without spatial decomposition—have been inflated by misattributing AGN light to stars.

Sample and data

The sample derives from Zhang et al.'s search for broad H$\alpha$ emitters (FWHM $> 1000\ \mathrm{km\ s^{-1}}$) in GOODS-N using NIRCam slitless spectroscopy from FRESCO and CONGRESS, which yielded 19 Type-I AGN with black hole masses of $\log(M_{\rm BH}/M_\odot) = 6.65$–$M_*$0. Two targets were excluded due to image artifacts (overlap with a foreground galaxy; artificial stripping), leaving 17 galaxies spanning $M_*$1–5.54. Imaging consists of up to seven NIRCam wide bands (F090W–F444W), drizzled to 0.03″/pixel, from the JADES DR5 release.

Decomposition methodology

The analysis uses a two-stage approach. First, GALFIT fits each target with a point-spread function (PSF) model for the AGN plus a single Sérsic profile for the host, with the PSF center fixed and the Sérsic center free; structural parameters ($M_*$2, Sérsic index $M_*$3, axis ratio) are taken from a representative short-wavelength (SW) band where resolution is highest and AGN dominance lowest. Second, an affine-invariant MCMC sampler (emcee) determines per-band scaling coefficients for the fixed PSF and PSF-convolved Sérsic models, yielding component fluxes in each band. Host fluxes come from aperture photometry on AGN-subtracted images.

Two methodological choices deserve emphasis. The authors deliberately avoid simultaneous multi-band fitting packages such as GALFITM, because they find GALFITM converges to unphysical solutions at long wavelengths (effective radii ~10% of the PSF FWHM, Sérsic indices hitting parameter boundaries) where the AGN dominates; in such cases the inferred F444W Sérsic flux can increase by more than 100% relative to fixing SW-derived parameters. Where GALFITM behaves physically, the two approaches agree—for GN1014406, GALFITM gives $M_*$4 versus $M_*$5 from GALFIT+MCMC. Validation against an independent source (GN53757 from Juodžbalis et al.) shows agreement within 0.02 dex against a Forcepho-based estimate. Host detection significance is assessed via the Bayesian Information Criterion difference between composite and PS-only models.

Detection results

Of the 17 targets, 15 show lower BIC with the composite model, but only 9 have $M_*$6, interpreted as confident extended-emission detections (~53% of the sample). Six have marginal evidence ($M_*$7) without clear visual confirmation, and two are consistent with a pure point source across all bands. Detected hosts have effective radii of 0.10–0.87 kpc and modest Sérsic indices ($M_*$8–4.9). Critically, the centroid offsets between the PSF and Sérsic components are small, typically ~0.12 kpc (all $M_*$9 kpc), arguing against spatially offset nebular emission as the origin of the extended light. The detection fraction is consistent with the ~50% of LRDs reported to host extended components in prior work.

The decomposed SEDs show that the point-source component dominates at long wavelengths: over ~50% and ~70% of the F356W and F444W light, respectively, comes from the PSF component. This directly contradicts photometry-based SED fitting by Zhang et al., which identified clear AGN components in only about half of the same sources—a discrepancy with direct consequences for stellar masses, discussed next.

Stellar masses and comparison with photometry-based fitting

SED fitting with PROSPECTOR (with a Parrot ANN emulator) on the decomposed host fluxes yields stellar masses of $z > 5$0–9.9. The comparison with Zhang et al.'s total-photometry SED fits splits into two regimes:

Regime	Mass difference vs. decomposition
AGN evident in photometry-based fit	Consistent, typically $z > 5$1 dex
No AGN evidence in photometry-based fit	Decomposition masses 0.9–1.8 dex lower

The implication is stark: when SED fitting based on total photometry fails to recognize the AGN contribution, a substantial fraction of nuclear light is assigned to the stellar component, inflating $z > 5$2 by 1–2 dex. A secondary effect compounds this: fitting with HST+NIRCam data alone (without MIRI/ALMA constraints) can itself bias stellar masses high by ~0.6 dex even under a pure-stellar assumption. Similar behavior appears in Pérez-González et al.'s prospector-AGN+ experiments, where sources lacking AGN evidence yield masses more than 1 dex higher than constrained fits. These results argue that imaging-based decomposition—or at minimum, mid-infrared constraints—is essential for reliable stellar masses in faint AGN hosts.

Size–mass relation and the $z > 5$3–$z > 5$4 plane

The detected hosts fall within ±1σ scatter of the size–mass relations of Allen et al. and Danhaive et al., broadly consistent with Chen et al.'s LRD measurements. The authors take this agreement as supporting the interpretation that the extended emission traces the host galaxy rather than nebular gas or scattered AGN light.

In the $z > 5$5–$z > 5$6 plane, the picture is markedly different. For the nine host detections, the inferred $z > 5$7 ratios span 0.01–1.48, placing all targets well above the local Reines & Volonteri relation; stellar masses are typically a factor of ~100–1000 below what the local relation would predict. Ratios lie closer to—but still above—the Greene et al. relation favored by Sun et al. as less galaxy-type dependent. The range matches Chen et al.'s decomposition-based LRD measurements (0.02–1.38) but is systematically higher than Juodžbalis et al.'s prism+photometry ratios (0.002–0.22), again traceable to the mass-attribution issue. Several individual systems exceed $z > 5$8, comparable to UHZ1, A2744-QSO1, and GN1001830, supporting the view that rapid early SMBH growth relative to hosts—and possibly heavy seeding via direct collapse—is common among faint high-redshift AGN.

Limitations and open questions

Several caveats bound these conclusions. First, the nature of the extended emission is not uniquely determined: while the lack of centroid offsets and consistency with the size–mass relation favor a stellar origin, nebular emission cannot be excluded with current imaging, and deep NIRSpec IFU observations would be needed to separate continuum from line emission. The authors document two cases (GN1033320, GN1029154) where foreground interlopers mimicked companion or off-centered nebular features, demonstrating that imaging alone can be misleading. Scattered AGN emission is disfavored by non-detections of C IV and He II combined with high equivalent widths of C III] and He I, but this argument relies on those specific line diagnostics.

Second, the decomposition cannot rule out a highly concentrated central stellar component that mimics point-like emission even at NIRCam resolution. Mock recovery tests quantify this risk: even with no injected AGN, forcing a PS+Sérsic split biases recovered host flux by ~20% (≈0.2 dex in $z > 5$9); in F444W, recovered flux reaches only 84.4% of truth at zero AGN contribution. Reliable host-flux recovery holds in F090W–F200W for AGN/host ratios up to ~1000, degrades beyond ratio ~40 in F356W, and is intrinsically biased in F444W. All detected hosts in the sample have AGN/host ratios below 2, within the regime where Sérsic parameters are stably recovered. Provided host SNR ≳ 3 and AGN/host ratio ≲ 1000, stellar masses are recoverable to within 0.5 dex.

Third, the black hole masses rest on single-epoch virial estimators calibrated locally, assuming pure Doppler broadening of H$z \sim 5$0. If Rusakov et al.'s scenario of electron-scattering broadening in dense ionized gas applies, virial masses could be overestimated by up to ~2 dex, substantially alleviating the tension with local relations. However, a direct dynamical measurement in one LRD returned a mass consistent with the single-epoch value, leaving the question unresolved. Finally, the assumed common Sérsic profile across all bands neglects known wavelength dependence of galaxy structure, though recent work finds no significant stellar-continuum size variation from UV to optical.

Conclusion

This work demonstrates robust AGN–host decomposition for faint broad-line AGN at $z \sim 5$1–6, detecting extended host emission in 9 of 17 targets (~53%). The principal quantitative result is that imaging-based decomposition reduces stellar-mass estimates by 1–2 dex relative to total-photometry SED fitting whenever the AGN component goes unrecognized, yielding $z \sim 5$2 ratios of 0.01–1.48 that place these systems far above the local scaling relation. Whether these hosts are genuinely undermassive, too compact to fully resolve, or their black hole masses inflated by non-virial line broadening remains open; resolving it will require spatially resolved spectroscopy and dynamical mass measurements in larger samples.