Active Galactic Nuclei and Host Galaxies in COSMOS-Web. I. NIRCam Images, Point-spread-function Models and Initial Results on X-Ray-selected Broad-line AGNs at 0.35 ≲ z ≲ 3.5

We present detailed and comprehensive data reduction and point-spread-function (PSF) model construction for all public JWST NIRCam imaging data from the COSMOS-Web treasury program (up to 2023 June, totaling 0.28 deg2). We show that the NIRCam PSF has significant short-timescale temporal variations and random spatial variations in all four filters (F115W, F150W, F277W, and F444W). Combining NIRCam with archival Hubble Space Telescope imaging, we perform multiwavelength active galactic nucleus (AGN)+host image decomposition to study the properties of 143 X-ray-selected (L bol = 1043.6–47.2 erg s−1) broad-line AGNs at 0.35 ≲ z ≲ 3.5. Leveraging the superb resolution, wavelength coverage, and sensitivity of NIRCam, we successfully detect host stellar emission after decomposing the central AGN point source in 142 objects. ∼2/3 AGNs are in star-forming galaxies based on the UVJ diagram, suggesting that there is no instantaneous negative AGN feedback. X-ray-selected broad-line AGN hosts follow a similar stellar mass–size relation as inactive galaxies, albeit with slightly smaller galaxy sizes. We find that although major mergers are rare (∼7%–22%) among the sample, more subtle nonaxisymmetric features from stellar bars, spiral arms, and minor mergers are ubiquitous, highlighting the importance of secular processes and minor mergers in triggering AGN activity. For a subsample of 30 AGNs at 1 < z < 2.5 with black hole mass measurements from single epoch spectra, they follow a similar black hole mass-stellar mass relation as local inactive early-type galaxies but reside preferentially near the upper envelope of nearby AGNs. We caution that selection biases and the intrinsic differences of AGN populations at different redshifts may significantly affect their location on the black hole mass-stellar mass plane.


Introduction
The discovery of tight correlations between the masses of the supermassive black holes (BHs) and the properties of their host galaxies (e.g., stellar velocity dispersion and bulge/total stellar mass) in the local Universe suggests that BHs and galaxies may coevolve with each other (e.g., Magorrian et al. 1998;Gebhardt et al. 2000;Kormendy & Ho 2013).Popular scenarios propose that the feedback from active galactic nuclei (AGNs) plays an important role in regulating the growth of the BH and its host galaxy by injecting energy and momentum into their environment (e.g., McNamara & Nulsen 2007;Hopkins et al. 2008;King & Pounds 2015).However, the details of these feedback processes and their impact on host galaxies are still being debated.Investigating when and how these correlations are established, as well as obtaining robust properties of AGN-host galaxies (e.g., morphology, structure, environment, and stellar population), is key to understanding galaxy and BH evolution.
The close track of the cosmic BH accretion history to the cosmic star formation history suggests that star formation and black hole growth are closely connected across cosmic time (e.g., Boyle & Terlevich 1998;Silverman et al. 2008;Kormendy & Ho 2013;Madau & Dickinson 2014), at least in the global sense.At cosmic noon (z ≈ 2), when star formation and BH accretion reach their peak epoch, we would expect stronger AGN feedback at play, making it the ideal epoch to investigate the cosmic evolution of the BH-galaxy scaling relations.However, previous studies of BH mass (M BH )-host stellar mass (M * ) relations of AGNs during this epoch tend to produce contradictory results, with AGNs lying above, on, or below the location relation (e.g., Borys et al. 2005;Alexander et al. 2008;Jahnke et al. 2009;Merloni et al. 2010;Sun et al. 2015;Ding et al. 2020;Suh et al. 2020;Zhang et al. 2023).Various factors may account for the large apparent discrepancies among different works, including measurement uncertainties and limited dynamical ranges of M BH and M * , small sample statistics, and selection biases (e.g., Lauer et al. 2007;Shen & Kelly 2010;Schulze & Wisotzki 2011;Shankar et al. 2016;Li et al. 2021a).
No consensus has been reached in terms of whether major merger is the primary triggering mechanism of AGN activity.Based on different AGN samples with imaging data of various depths and resolutions, some found a low merger fraction (20%), while others found that merger fractions can be as high as 40%-50%, with tentative evidence that the merger fraction increases as AGN luminosity increases (e.g., Treister et al. 2012;Villforth et al. 2014;Hong et al. 2015;Villforth et al. 2017;Ellison et al. 2019;Gao et al. 2020;Marian et al. 2020; followed by derivation of host properties including host structural parameters from AGN-host image decomposition and stellar properties from fits to their spectral energy distributions (SEDs) in Section 5. Section 6 discusses the properties of AGN-host galaxies, AGN-host offset, and the M BH -M * relation.Our main conclusions are summarized in Section 7. We adopt a flat ΛCDM cosmology with H 0 = 70 km s −1 Mpc −1 , Ω m = 0.3, and Ω Λ = 0.7, and a Chabrier (2003) initial mass function (IMF) for stellar population analysis.The reduced NIRCam imaging data and PSF models are publicly available in Dataverse via doi:10.7910/DVN/1GDKDY.

Data
COSMOS-Web (Casey et al. 2023 PIs: Kartaltepe & Casey, ID = 1727) is a 255 hr JWST treasury program targeting the central area of the COSMOS field.It covers a contiguous area of 0.54 deg 2 with NIRCam imaging in four filters: two shortwavelength filters (SW), F115W and F150W; and two longwavelength filters (LW), F277W and F444W.At the same time, a 0.19 deg 2 area of MIRI imaging with a single filter (F770W) is observed in parallel with NIRCam observations.At the time of writing, 80 visits have been successfully executed (2023 January and 2023 April-May), covering roughly half of the entire COSMOS-Web field.The remaining visits are scheduled in 2023 December and 2024 January.Moreover, the entire COSMOS-Web field is covered by HST Advanced Camera for Surveys (ACS) Wide Field Channel (WFC) I band (F814W) observations (Koekemoer et al. 2007), which extends the wavelength coverage to rest-frame ultraviolet (UV)/optical for relatively low-redshift sources (∼0.4-0.8 μm at z  1).
In this paper, we make use of NIRCam imaging data of the COSMOS-Web survey and HST F814W mosaics (v2.0)4 from the COSMOS survey (Koekemoer et al. 2007;Massey et al. 2010).

Data Reduction
We retrieve uncalibrated NIRCam raw data of 80 visits from MAST. 5 We reduce the data using version 1.10.2 of the jwst6 pipeline (Bushouse et al. 2023) with the Calibration Reference Data System (CRDS) version of 11.17.0 (context file jwst_1089.pmap)and custom steps for NIRCam image data reduction.By the time of writing, a new pipeline version (1.11.2) and CRDS context file (1097) are available, but none of the updates affects the reduction and calibration of the NIRCam imaging data.All of the JWST NIRCam data used in this paper can be found in MAST:10.17909/0ktr-qn45.Our detailed reduction procedures are as follows: 1. We first reduce uncalibrated raw images of individual exposures using the Stage 1 pipeline Detector1Pipeline, which applies basic detection-level corrections.We adopt the default configuration with the exception of three parameters of the Jump Detection step.We turn on expand_large_events to flag large cosmic-ray events and fine tune min_jump_area=10 and sat_requir-ed_snowball=False for better identification and flagging snowballs.We then subtract 1/f noise (horizontal and vertical patterns) in the output countrate (slope) images of Stage 1 pipeline using scripts developed by the Cosmic Evolution Early Release Science Survey (CEERS; ERS 1345, PI: Steven Finkelstein) team (Bagley et al. 2023).2. We run Stage 2 pipeline Image2Pipeline to the countrate images with default parameters.This step includes wcs assignment, flat-fielding, and photometric calibration, and returns fully calibrated individual exposures.We skip the Skymatch step in Stage 3 pipeline and subtract two-dimensional background after masking sources and bad pixels in individual exposures using SExtractorBackground with a box size of 50 pixels7 implemented in photutils (Bradley et al. 2022).Masks are generated with dilated (using a circular footprint with a radius of 5 pixels) segmentation maps of sources with at least 10 connecting pixels and fluxes above 2 × pixel-wise standard deviation in convolved images (Gaussian kernel with full-width-at-half-maximum, FWHM =3 pixels).We note that our adopted box size may oversubtract faint and diffuse emission of very large objects (size ?50 pixels).3. We then subtract the "wisp" features caused by scattered light coming off-axis and bouncing off the top secondary mirror strut.(2023).We obtain the scaling factor multiplied to the template to subtract the wisps by fitting the template to the source emission-masked images using the scipy.optimize.leastsqfunction.To quantify the significance of wisps in each exposure, we select a 5 × 5 pixel window targeted on the brightest wisp feature in each detector and calculate the median ratio within the window between the wisps to be subtracted (template times scaling factor) and the background noise (σ from the error extension of calibrated images).The distributions of the wisp significance are shown in Appendix Figure 16.A3, A4, and B3 imaged with the F115W filter have wisp significance <0.1σ, suggesting negligible effects of wisps on these detectors.Therefore, we only perform wisp subtraction in the four detectors imaged with the F150W filter and B4 imaged with the F115W filter.4. We also find "claw" features in six observations taken in 2023 January (Observation number 043-048).Claws are artifacts that occasionally appear due to scattered light from extremely bright stars located in a specific susceptibility region that is very far from the field-ofview (FoV) of NIRCam.For COSMOS-Web observations, we find claws in the B1 and B2 detectors imaged with the F150W filter (Appendix Figure 17) instead of A1, A2, and B4 detectors reported in JWST documentation10 and with slightly different appearances (seem to be flipped).This is likely due to the differences in the location of the bright star in the susceptibility region and the position angle of the telescope.As claws move from observations to observations (much smaller movement among four dither exposures within a visit), we manually mask the pixels that are affected for each exposure.These features are at the level of ∼0.8 times the background noise σ in individual exposures, comparable to the level of wisps in A3 detector in the F150W filter.Note that both wisps and claws are more prominent with respect to the background in smoothed images used for source detection and may be confused as sources with low surface brightness. 5. Finally, we run Stage 3 pipeline Image3Pipeline to produce a single mosaic for each filter by combining all the calibrated images (all dithers and detectors).We turn on tweakreg step for astrometry correction, outlier_detection step for outlier pixel rejection (bad or cosmic-ray affected pixels), and resample step for image resample and combination.We correct and tie the astrometry of NIRCam images to that of the COSMOS2020 "Farmer" catalog (Weaver et al. 2022), whose astrometry was tied to Gaia (Gaia Collaboration et al. 2016).We use the updated version COSMOS2020_FARMER_R1_v2.2_p3 from CDS VizieR (Weaver & Kauffmann 2022) and adopt coordinates from model fitting (RAmdeg and DEmdeg) for sources with reliable model fitting results and small drift from detection (FModel=0).We adopt the default configurations for outlier_detection except for setting maskpt=0.5,which is the fraction of maximum weight to use as the lower limit for valid data.We find that a significant number of good pixels have weight below 0.7 (default) times the maximum value of the weight map.If not modified, outlier pixels with relatively small weight would not be identified and remain in the final mosaic.For the resample step, we adopt common output image shape (output_shape), reference pixel position (crpix) and coordinate (crval), pixel scale (pixel_scale=0 03 pixel −1 ), and input pixel "shrunk" fraction (pixfrac=0.8) for all four filters associated with the same observation.All of the other parameters are kept as their default values.
Figure 1 shows an example of a pseudo-color mosaic of COSMOS-Web NIRCam imaging of observation 009.We also compare positions of bright sources (signal-to-noise ratio, S/N >50; containing both resolved and unresolved sources) in our NIRCam mosaics with those in the COSMOS2020 catalog.The agreement with the COSMOS2020 catalog is quite good, with median differences and standard deviations of coordinates between all sources in the COSMOS2020 catalog and their counterparts in NIRCam images: ΔR. A. = − 1±6, −3 ± 4, −2 ± 3, −2 ± 3 mas, Δdecl.= 4 ± 4, 3 ± 4, 1 ± 3, and 1 ± 3 mas among 80 visits for the F115W, F150W, F277W, and F444W filters, respectively.The standard deviations of ΔR.A. and Δdecl.are larger for sources in individual visits (30-40 mas).Wavelength-dependent galaxy morphology, such as more frequent presence of off-nucleus star-forming clumps (e.g., Elmegreen et al. 2007;Guo et al. 2015), may contribute to the scatter.
Figure 2 compares our NIRCam F150W mosaic at Observation 047 with those from the 30 mas version of the public data release 0.2 by the COSMOS-Web team.11Our mosaics show overall improved background subtraction, with significant improvement of wisps and claws removal.Average 5σ point source depths for areas covered by four dither exposures (calculated within 0 15 radius apertures without application of aperture corrections) are 27.40, 27.66, 28.34, and 28.25 AB mag in F115W, F150W, F277W, and F444W, respectively.These values are generally consistent with those reported in Casey et al. (2023).

PSF Construction
Accurate modeling of the PSF is crucial to robustly measure the structural properties of galaxies and AGNs.Recently, Zhuang & Shen (2023) have demonstrated that NIRCam mosaics have significant spatial PSF variations.The maximum and root mean square (rms) fractional variations of PSF FWHM decrease from ∼20% and 5% in the F070W filter to ∼3% and 0.6% in the F444W filter, with exact values depending on the adopted resample parameters and dither pattern.They recommend the use of PSFEx (Bertin 2011) to model the PSF and its spatial variation given its superior performance over other commonly used methods.Therefore, we follow their procedures and construct PSF models consistently for both JWST NIRCam and HST ACS mosaics.We run SExtractor (Bertin & Arnouts 1996) to select high S/N (SNR_WIN>100), nonblended (FLAGS<2), nonirregular (ELONGATION<1.5),point-like (CLASS_STAR>0.8), and without bad pixels (IMAFLAGS_ISO=0).We further exclude sources with X-ray detection by matching with the Chandra COSMOS-Legacy catalog (Marchesi et al. 2016a) and 2.5σ outliers in half-light radius (FLUX_RADIUS).We adopt the PIXEL_AUTO basis type and an oversampling factor of 2 (PSF model pixel scale is half the input value) with PSF model size of 201 pixels for HST ACS F814W and NIRCam SW filters, and 301 pixels for NIRCam LW filters.
For each NIRCam mosaic, we construct three types of PSF models: (1) a global PSF model using all of the point-like sources across the entire FoV of each dither-combined mosaic; (2) a local PSF model accounting for the spatial variation of PSF within the mosaic; (3) a set of two PSF models (broad and narrow), each of which is constructed using the broader or  narrower half of the point sources divided by their median FWHM.Given the small amount of point sources in individual mosaic with median numbers of ∼34-40 sources across four filters, only a first order polynomial (linear) spatial-dependent model can provide usable local PSF model (no artifacts and adequate S/N).
For HST ACS F814W images, we only construct one PSF model for each target due to the adoption of HST ACS mosaic product.We select all point-like sources located within a radius of 1 5 of the target (Section 4) and construct a global PSF model using PSFEx following the same procedures as for NIRCam mosaics.A radius of 1.5 arcmin12 is chosen to roughly match the FoV of 202 × 202 arcsec 2 for HST ACS exposures to minimize PSF variation, with the caveat that the source may not lie at the center and thus the point-like sources nearby may come from different observations.As a result, a median of 13 point-like sources are available for PSF construction for each AGN target.

PSF Properties
Assuming that the global PSF model represents typical PSF in each mosaic, Figure 3 shows the PSF FWHM of each observation as a function of observation time.The FWHM of a PSF model is measured by fitting its core with a twodimensional elliptical Gaussian function following Zhuang & Shen (2023).For all observations as a whole, the median PSF FWHMs of the F115W, F150W, F277W, and F444W filters are 116 2 2 , and 158 2 1 mas, respectively, with superscript and subscript indicating the difference between 84th and 16th percentiles and the median (Table 1).These PSF FWHMs are much improved (∼7%-15% for F115W and ∼6%-13% for F150W) for SW filters and slightly smaller for LW filters (∼3%-6% for F277W and ∼1%-2% for F444W) compared to those constructed from point-like sources in mosaics with the same pixel scale but larger pixel shrunk fraction (pixfrac=1) (Finkelstein et al. 2023;Zhuang & Shen 2023).The temporal variation of PSF FWHM is dominated by short timescale fluctuations, with fractional rms of ∼2.9%, 2.4%, 1.9%, and 0.8% for the F115W, F150W, F277W, and F444W filters, respectively (Table 1).The temporal variation in terms of FWHM is smaller compared to pixel-level flux fluctuations (on the order of ∼3%-4%) using calibrated individual exposures of two epochs presented in Nardiello et al. (2022).
We estimate the spatial variation across the FoV of an individual observation by measuring the rms of FWHMs of local PSF models at 100 positions randomly distributed across the FoV.The median 3σ-clipped fractional variations over all 80 visits in the F115W, F150W, F277W, and F444W filters are ∼2.8%,2.1%, 2.7%, and 1.0%, respectively.These values are comparable with the temporal variation mentioned above, except for F277W, which is 0.8% larger.
As an independent check of the spatial variations estimated using local PSF models, we directly measure the FWHM of point-like sources used for PSF model construction following the same method as for PSF models.With the caveat that the median values of FWHM may be slightly different due to different pixelscales for PSF model (0 015) and mosaics (0 03) and that the FWHM rms may be affected by the spatial distribution of sources, we assume that the ratio between rms and median of PSF FWHM remains the same.We find that the spatial variations of point-like sources in individual observations are higher than that estimated from randomly placed local PSF models, with median 3σ-clipped fractional variations of ∼7.1%, 5.0%, 2.4%, and 1.5% among 80 visits (Table 1).In other words, the local PSF models underestimate the true spatial variations of the PSF.
To understand the differences between the two approaches, we take a closer look at the distribution of PSF FWHM in the entire FoV of each mosaic.We do not find a clear pattern with location in individual observations or in a combined map that is constructed using all of the sources in 70 visits with the same JWST V3 position angle (107°).After binning the FWHM of sources in the combined map with bin sizes ranging from 50 to 500 pixels, we find a gradual reduction of the PSF FWHM rms in the binned maps, with a reduction of ∼50% for the bin size of 500 pixels.This suggests that the spatial variation in individual observations is dominated by random variations.This may be a result of the adopted dither pattern of COSMOS-Web (Figure 2 in Casey et al. 2023), which results in coverage with different number of exposures and detectors from different Modules.
To summarize, we find that the median PSF FWHMs in our COSMOS-Web NIRCam mosaics are 56.2,61.1, 116, and 158 mas for the F115W, F150W, F277W, and F444W filters, respectively.The PSFs of NIRCam imaging have significant  For HST ACS mosaics, the FWHM of the derived PSF models of our objects ranges from 94 to 104 mas, with a median of 98 mas and a rms of 2 mas.The amount of fractional variation of HST ACS PSF FWHM (∼2%) is between that of F150W and F277W filters of JWST NIRCam.

The AGN Sample
The Chandra COSMOS-Legacy survey (Civano et al. 2016) is a 4.6 Ms project that covers 2.15 deg 2 of the COSMOS field with a flux limit of ∼2 × 10 −16 erg s −1 in the 0.5-2 keV band.The identification of optical and NIR counterparts of 4016 X-ray sources, the collection and measurements of their redshifts, X-ray spectral properties (e.g., flux, hardness ratio), and AGN type classification are presented in Marchesi et al. 2016a.In this work, we use the 632 spectroscopicallyconfirmed broad-line AGNs (Clsp =1) with reliable redshifts (q_zspec 1.5) as our parent sample.Among them, 145 fall in the current footprint of COSMOS-Web NIRCam imaging.We exclude two sources (CID-58 and CID-161) that are contaminated by the spikes of nearby bright stars.Redshifts of CID-104 and CID-1305 are incorrect and are manually corrected to z = 0.920 and z = 2.177 by examining their spectra from Hasinger et al. (2018).CID-1305 is also included in Suh et al. (2020) with the same correct redshift as ours.The optical counterpart of CID-1132 is incorrect and is corrected to the source located ∼1 5 to the west, which is much closer to the Chandra X-ray coordinates (∼0 3).Due to small overlapping regions among adjacent observations, nine objects have the majority of their emission covered by two observations, offering independent information for uncertainty estimates.
We calculate the rest-frame 2-10 keV intrinsic (absorptioncorrected) luminosity (L X ) for our sample assuming a powerlaw spectrum with a photon index Γ = 1.8 as: where f E E , 1 2 is the observed flux between energies E 1 and E 2 in keV, D L is the luminosity distance, and C E E , 1 2 is the absorption correction factor given in the soft (0.5-2.0 keV), hard (2.0-10.0keV), and full (0.5-10.0 keV) bands in Marchesi et al. (2016a).To minimize the effect of absorption, we adopt a priority order of hard > full > soft band to derive L X when sources are detected in more than one band.Note that although our sample consists of optical type 1 AGNs, their X-ray emission could still be obscured, and a correction factor estimated from a simple hardness ratio value in Marchesi et al. (2016a) may not be appropriate (Li et al. 2019;Peca et al. 2021).Therefore, we also collect L X and N H from Marchesi et al. (2016b) and Lanzuisi et al. (2018), in which detailed X-ray spectral fitting is performed using an absorbed powerlaw model and the physical MYTorus model (Murphy & Yaqoob 2009), respectively, for a subset of the Marchesi et al.Lanzuisi et al. (2018).The heavy obscuration inferred from its X-ray spectrum is consistent with very low AGN contribution (2%) from our image decomposition (Section 5.1).Moreover, we do not find concrete evidence for the presence of broad lines in CID-45 after visual inspection of its optical spectrum from Hasinger et al. (2018).Even so, we keep this object in our sample as the properties of its host galaxy remain robust.For CID-104 and CID-1305, their L X values are derived from full and hard band flux, respectively, using the updated redshifts without absorption correction because their correction factors are derived using the wrong redshifts in Marchesi et al. (2016a).The presence of a bright nucleus in all five filters of the two objects is consistent with little absorption.
We convert rest-frame 2-10 keV X-ray luminosity to bolometric luminosity (L bol ) adopting luminosity-dependent bolometric correction from Lusso et al. (2012), as parameterized by Yang et al. (2018).The final sample consists of 143 broad-line AGNs spanning redshift range of z = 0.35-3.5 (median z = 1.8) and log L bol range of 43.6-47.2(median 45.3) erg s −1 .Figure 4 compares the distribution of redshift and L bol of our X-ray broad-line AGN sample with the parent X-ray AGN sample from Marchesi et al. (2016a) containing all the sources (including those with photometric redshifts and type 2 AGNs) with L X 10 42 erg s −1 .A gallery of F277W images from our sample is shown in Figure 5.The properties of our sample are summarized in Table 2.

Multiwavelength Simultaneous AGN-host Image Decomposition
We perform multiwavelength simultaneous AGN-host image decomposition using GALFITM (Häußler et al. 2013), which is a multiwavelength version of GALFIT (Peng et al. 2002(Peng et al. , 2010)).GALFITM takes into account the wavelength-dependent galaxy structure that is mainly due to different spatial distributions of stellar populations, metallicity, and dust attenuation.GALFITM is widely adopted to study the structure of galaxies of all redshifts (e.g., Häußler et al. 2022;Gillman et al. 2023;Kartaltepe et al. 2023;Sun et al. 2024) and to decompose the host galaxies of AGNs in the local universe and at cosmic noon (Zhuang & Ho 2022, 2023;Zhuang & Shen 2023) Since our objects span a wide range of redshift (z ≈ 0.35-3.5,corresponding to angular scales of 4.9-8.5 kpc arcsec −1 ), we only adopt a PSF model for the AGN and a single Sérsic model for its host galaxy for consistency in this work.We will study their substructures (e.g., bulges and bars) in a follow-up paper.We use a cutout size of the larger value between 20 times FLUX_RADIUS in the F444W filter and 101 × 101 pixels (∼3″).Blended companions are fitted simultaneously using either a Sérsic model if CLASS_STAR<0.9 in all five filters or a PSF model.All of the other nonblended sources in the cutout are masked.The centers of the AGN and the host are tied together and allowed to vary independently in different filters to account for residual astrometry mismatches (0.3 pixel among NIRCam images and ∼1.3 pixel between HST ACS F814W and NIRCam F277W).We also perform decomposition with AGN and host centers free to vary.We find consistent host parameters from decomposition regardless if we force the AGN and the host models to have the same center.We adopt constant ellipticity and position angle parameters across wavelength, leave magnitudes of the Sérsic component ( ḿSersic ) and the AGN component (m AGN ) free to vary, and allow the Sérsic index n and effective radius R e to vary quadratically with wavelength following Häußler et al. (2013) and Zhuang & Shen (2023).13n is restricted between 0.3 and 7 and R e between 0.5 and 100 pixels, corresponding to 0.07-25 kpc in the redshift range of our sample.
We use measurements from SExtractor as initial guesses in GALFITM, with m AGN and ḿSersic equal to 0.75 mag + total magnitude (assuming equal contribution from AGN and the host galaxy), R e equal to the half-light radius in the F277W filter, and n equal to 1.For the sigma image, we use the error image from the resample step for JWST NIRCam images.Since only a weight image (inverse variance of background) is available for ACS F814W mosaics, we construct sigma images by adding Poisson noise to the background noise from the weight image in quadrature, which preserves patterns from data reduction (such as cosmic-ray flagging, flat-fielding, and resampling).We note that although complicated substructures (e.g., bulge, bar, and spiral arms) are present in our sample (see Section 6.2), a single Sérsic model can recover the total galaxy emission as robustly as more detailed models (e.g., bulge +disk), as suggested in studies of nearby galaxies (e.g., Casura et al. 2022;Häußler et al. 2022).
Since we have three PSF models for each NIRCam filter (global, broad, and narrow), we choose the best one for each filter by performing AGN-host decomposition in that filter.We adopt the global PSF model by default unless the improvement of the reduced chi-squared (c n 2 ) of the fit from broad or narrow PSF model is larger than 10%.As a result, global PSF models perform equally well or the best in all four filters in 134 sets of images.Broad and narrow PSF models perform significantly better in at least one filter in 11 and 7 sets of images, respectively.
An example of our multiwavelength simultaneous AGN-host image decomposition is shown in Figure 6.For all objects except CID-352, the host galaxy can clearly be seen in the data −nucleus panel or surface brightness profiles in LW filters.We thus exclude CID-352 from the following analysis.Making use of results from SED fitting to AGN contamination-subtracted host fluxes (Section 5.3), we estimate the host-to-total flux fraction ( f host ) at rest-frame 5000 Å and 1 μm.The 16th, 50th, and 84th percentiles of f host are ∼26%, 58%, and 91% at rest-frame 5000 Å and ∼42%, 70%, and 91% at restframe 1 μm.
In total, 7/143 (∼5%) of objects have saturated pixels at the center (four in the F277W filter and three in the F444W filter), with a median number of saturated pixels of 18 (∼2.5 pixel radius circle).To quantify the potential bias of the central pixel saturation to the host parameters, we pick three objects with f host at 16th, 50th, and 84th percentiles of the entire sample, respectively, and manually mask the central 5 × 5 pixels in the F277W and F444W filters to mimic saturated central pixels.We then perform the decomposition twice, each with modified mask in one filter.Thanks to the excellent resolution of NIRCam and the multiwavelength simultaneous decomposition capability of galfitm, we find that ḿSersic , R e , and n measured with central pixels masked have only subtle difference with maximum absolute difference (median absolute difference) of 0.09 (<0.01) mag for ḿSersic , 0.05 (<0.01) dex for R e , and 0.07 (<0.01) dex for n, compared to those measured without masking central pixels.The differences of these parameters are all smaller than the typical uncertainties (0.2 mag for ḿSersic , 0.1 dex for R e and n in Section 5.2).Therefore, we keep these objects in our analysis below.

Uncertainty in AGN-host Parameters
As discussed in previous works, the nominal error reported by the fitting code always underestimates the true uncertainty (e.g., Häussler et al. 2007;van der Wel et al. 2014).Moreover, PSF mismatch, which is unavoidable in practice, could also lead to systematic biases to the recovered host galaxy parameters (Zhuang & Shen 2023).In this paper, we derive more reliable uncertainties for host parameters using duplicate observations of nine AGNs in our sample, assuming that their properties are representative among the full sample of 143 objects.These nine objects span a wide range of f host from ∼0 to 1, ḿSersic from 18.5 to 24.3 mag, R e from 0 02 to 1 1, and n from 0.3 to 7, covering the parameter space of the majority of our full sample.Figure 7 presents the absolute differences between two duplicate observations of host parameters ( ḿSersic , R e , and n) versus f host in five filters.We find that host parameters derived from duplicate observations are in good agreement with each other across most of the f host range except for f host  5%, which is mainly driven by strong AGN contamination.The difference of host parameters between two independent observations includes the effect of imperfect astrometry, pixel spatial sampling, and PSF modeling, and thus provides a better estimate of the true uncertainty.Therefore, with the assumption that the results from these objects are representative of the parent sample, we adopt a fiducial systematic uncertainty of 0.2 mag for ḿSersic , 0.1 dex for R e and n in all five filters to be added in quadrature to the uncertainty reported by GALFITM.Our adopted uncertainty of 0.2 mag for the host magnitude is generally consistent with that derived from mock AGN images by comparing the input and fitted parameters with similar f host and host magnitude as our sample (Zhuang & Shen 2023).We note that assuming a fixed uncertainty for the entire sample may lead to overestimated or underestimated error for individual objects.

Stellar Masses of AGN-host Galaxies
We correct foreground Galactic extinction using the extinction curve from Cardelli et al. (1989) with R V = 3.1 and the dust map from Schlegel et al. (1998).We then adopt the SED fitting code CIGALE v2022.1 (Boquien et al. 2019;  Yang et al. 2022) to estimate the stellar masses of AGN-host galaxies using fluxes after removing AGN contamination.
To investigate the reliability of stellar mass estimates from SED fitting, we compare the stellar mass estimates from the best-fit template (M * best ) and those from the probability density function distribution of the likelihood of all the templates (M * bayes ).All objects have , indicating that our decomposition is reliable and robust for the vast majority of the sample.These six objects have excess host fluxes in the F814W or F115W filter likely due to an overestimation of host emission in objects with very low f host (Zhuang & Shen 2023).For CID-1065 with c = n 2.5 2 , it has host flux excess compared to the best-fit SED template in the F444W filter likely due to underestimated AGN torus emission due to a compact host structure.We then fit host SEDs of these objects after excluding the filter with excess flux.The resulting stellar masses are consistent (1 × uncertainty) with the original values.For nine objects with duplicate observations, their stellar masses are all consistent.Therefore, we adopt the mean of the two estimates as their final stellar masses.Note that we do not include uncertainties from model assumptions, such as stellar initial mass function, stellar population model and star formation history, which can introduce ∼0.3 dex uncertainty to our stellar mass estimates (Conroy 2013).
The stellar masses of our sample span over 10 10.1-11.5 M e with a median of 10 10.8 M e .We find a significant positive correlation between M * and L bol (Spearman correlation strength ρ = 0.3, p-value=3 × 10 −4 ), with no correlation between M * and redshift.This is consistent with observations that the host galaxies of AGNs have abundant molecular gas reservoir, such that more massive hosts have more abundant gas content, capable of fueling more luminous AGN activities (e.g., Beelen et al. 2004;Bischetti et al. 2021).

Stellar Mass Comparison with Previous Works
We compare our stellar masses with those reported in earlier works that are derived using two approaches.First, multiwavelength AGN+host SED fitting (i.e., total AGN+host fluxes without performing image decomposition) implementing AGN and galaxy templates (Zou et al. 2019;Suh et al. 2020). 14econd, mass-to-light ratio from host color from AGN+host image decomposition to HST images from Ding et al. (2020).After crossmatch, we obtain 95 objects from Zou et al. (2019), 26 objects from Suh et al. (2020), and five objects from Ding et al. (2020) that are also covered by our AGN sample.We have corrected their stellar masses to match our adopted cosmology.
We first compare our stellar masses with those from fits to AGN+host SEDs using AGN and stellar templates (Zou et al. 2019;Suh et al. 2020).We find that our stellar masses are systematically larger than those from Zou et al. (2019) with a median difference of -+ 0.16 0.24 0.35 dex (upperand lower-script indicating differences between 84th and 16th percentiles from the median) and systematically lower than those from Suh et al. (2020) with a median difference of - -+ 0.34 0.30 0.20 dex.Four    systematics, and have a small chance of catastrophic failures (deviation from the truth by more than 1 dex).
For the five M *  10 10.7 M e AGNs overlapping with Ding et al. (2020), three objects imaged with HST F140W filter have M * differences of ∼0.3 dex (ours are higher) and the other two imaged with HST F125W filter have differences of ∼0 and −0.1 dex.Many factors can contribute to the differences.At z ≈ 1.5, the F140W filter only probes rest-frame ∼5600 Å, missing the peak emissions from older stellar populations that contribute the majority of the stellar mass in these massive systems.Since only two HST filters are available, Ding et al.
(2020) estimated stellar masses using mass-to-light ratios from stellar populations with fixed age (1 Gyr for z < 1.44 and 0.625 Gyr for z > 1.44), adding extra uncertainty to the measurements.As a comparison, the stellar age from our SED fitting ranges from 1.2 to 1.6 Gyr.The three objects with ∼0.3 dex stellar mass differences have z > 1.44, which indicates that the difference is partly due to their adopted fixed stellar population template.Moreover, the much-improved spatial resolution and sensitivity of JWST NIRCam imaging compared to HST makes our decomposition more reliable.By comparing f host in HST F140W with NIRCam F150W and that in HST F125W with NIRCam F115W, we find an overestimation of ∼10% ± 6% in their decomposition to HST images.

Results and Discussion
6.1.Color, Morphology, and Structure of AGN-host Galaxies 6.1.1.AGN Hosts Are Mainly Star-forming Galaxies We classify the host galaxies of AGNs into star-forming and quiescent galaxies using the UVJ diagram from Williams et al. 2009.Rest-frame U − V and V − J colors are obtained from CIGALE.We find that the hosts of 86 ± 4 (61% ± 3%; considering the uncertainty of color) AGNs are classified as starforming galaxies.A large fraction of AGN hosts are star-forming galaxies is consistent with previous studies of X-ray AGNs at similar redshifts (e.g., Rosario et al. 2013;Sun et al. 2015;Mountrichas et al. 2021;Coleman et al. 2022) and those of luminous AGNs or quasars in the nearby Universe (e.g., Jarvis et al. 2020;Li et al. 2021bLi et al. , 2024a;;Xie et al. 2021;Zhuang & Ho 2022, 2023).Our results support the mutual dependence of BH accretion and host star formation on molecular gas reservoir (e.g., Shangguan et al. 2020;Koss et al. 2021;Zhuang et al. 2021) and suggest no instantaneous negative AGN feedback (e.g., Costa et al. 2014;Harrison et al. 2018).

Sérsic Index Distribution
We estimate the Sérsic index n at rest-frame 1 μm by fitting a second-order polynomial to wavelength and n of five filters, following our assumption of wavelength-dependence in the decomposition (Section 5.1). Figure 9 shows the histogram of Sérsic index n at rest-frame 1 μm.We find that for X-ray broadline AGNs, n peaks around 1-2, with a relatively flat distribution toward larger n.If we classify objects with n 2 as late-type diskdominated galaxies and n > 2 as early-type bulge-dominated galaxies (Zhuang & Ho 2023), then 54 AGNs (∼38%) live in latetype disk-dominated galaxies.We do not find a clear dependence of n on stellar mass, with ∼40% and 37% of late-type galaxies with M * < 10 10.8 M e and M * 10 10.8 M e , respectively.The fraction remains robust if we switch to n at rest-frame 5000 Å.This fraction is in broad agreement with AGNs in the nearby Universe and at 1 < z < 3 (e.g., Fan et al. 2014;Rosario et al. 2015;Ding et al. 2020;J. Li et al. 2021b;Bennert et al. 2021;Kim et al. 2021;Li et al. 2023;Zhuang & Ho 2023), suggesting no significant redshift evolution of Sérsic index of AGNhost galaxies.
We note that while low n values (∼1) indicate that the galaxy is dominated by stellar disk, large n values (n  2) do not necessarily rule out the presence of a disk.For example, a combination of a high-surface-brightness small-scale (pseudo-)bulge and a lowsurface-brightness large-scale disk can also mimic the surface brightness profile with a large n.In a follow-up paper, we will perform detailed bulge(+bar)+disk decomposition to the AGNhost galaxies to study the properties of bulge and disk, and compare with those in a non-AGN control sample.

Stellar Mass-Size Relation
The relation of the stellar mass of a galaxy and the size distribution of its mass, stellar mass-size (M * -R e ), encodes information about the galaxy's evolution and assembly history.To investigate the position of X-ray broad-line AGNs on the M * -R e plane, we estimate R e at rest-frame 0.5 μm, following the same procedures as n mentioned above.R e estimated using this method is in excellent agreement (--+ 0.01 0.08 0.04 dex) with that estimated using color gradient in van der Wel et al. (2014).
For the M * -R e relation, we use more accurate host color classification.We only consider the classification of an object accurate if it is classified as the same type (star-forming or quiescent) in at least 75% of the cases after perturbing its U − V and V − J colors using their uncertainties in 100 trials.As a result, we have 66 star-forming hosts, 29 quiescent hosts, and 47 hosts without reliable classification.
Figure 10 shows the M * -R e relation of X-ray broad-line AGNs in different redshift bins, compared to the non-AGN galaxies in van der Wel et al. (2014).The classification of star-forming and quiescent galaxies in van der Wel et al. ( 2014) is also based on the UVJ diagram from Williams et al. (2009).We find that AGN hosts generally follow the M * -R e relation for non-AGN galaxies in the same redshift range (van der Wel et al. 2014).However, starforming AGN hosts lie slightly lower compared to their non-AGN counterparts at 1 < z 2.5, where we have more than 10 starforming AGN hosts.This result, albeit with small statistics, echoes previous studies at similar redshifts (e.g., Barro et al. 2014;Silverman et al. 2019) and supports the idea that a significant fraction of AGNs tends to be hosted by compact star-forming galaxies (e.g., Li et al. 2021b), which are associated with concentrated molecular gas and dust distribution, as suggested by recent submillimeter observations of AGNs at cosmic noon (e.g., Stacey et al. 2021;Jones et al. 2023).

The Ubiquitous Presence of Nonaxisymmetric Features
Gas-rich major mergers have long been considered as an important triggering mechanism of luminous AGNs, despite the longstanding debate of whether it is the primary triggering mechanism (e.g., Alexander & Hickox 2012;Treister et al. 2012;Ellison et al. 2019).Leveraging the sensitivity and restframe near-IR coverage of JWST NIRCam, we study the close environment of AGNs at cosmic noon.We select close major merger candidates by applying the following two criteria: (1) sources have clear evidences of interacting/tidal features in residual maps in any of five filters (25 candidates); and (2)  sources have companions with magnitude difference (compared to the magnitude of the AGN host) smaller than 1.5 mag (i.e., flux ratio of 1:4) in the F277W filter and within a projected distance of 20 kpc (16 candidates).With 5σ detection limit of 27.7-28.3mag in the F277W filter (Casey et al. 2023) and the faintest host of 23.3 mag, we can detect all potential companions if present.In total, 31 sources satisfy either criterion and 10 sources satisfy both criteria.We caution that candidates selected from the first criterion may contain interacting features caused by minor mergers, while those selected from the second criterion may not have physically associated companions because no reliable spectroscopic redshifts are available for the companions.Therefore, 31/143 (∼22%) should be considered as a strict upper limit of major merger fraction, while 10/143 (7%) can be considered as a conservative lower limit.Figure 11 shows the major merger rate as a function of redshift and AGN strength.Given the limited sample size of this work, we do not find clear dependence of major merger rate on redshift and AGN strength.The overall major merger rate in AGNs is consistent with massive inactive galaxies at similar redshifts (Bundy et al. 2009;Duncan et al. 2019;Ferreira et al. 2020;Whitney et al. 2021).
Minor mergers and secular processes can also drive gas inflow and fuel BH accretion (e.g., Kormendy & Kennicutt 2004;Kaviraj 2014;Kim & Kim 2014;Shu 2016).We visually check the normalized residual (data−PSF−Sérsic) images (see Figure 12 for an example in the F277W filter) in all five filters and identify clear presence of stellar bar or spiral arms in 40 (28%) objects.Besides the aforementioned strong nonaxisymmetric features, nearly all of the objects show various degrees of disturbed morphology or faint companion(s) (flux ratio =1: 4), which may be interpreted as evidence of minor mergers.The ubiquitous presence of nonaxisymmetric features combined with a small major merger fraction (7%-22%) highlights the importance of gas fueling by secular processes and minor mergers, with the caveat of the small sample size investigated here.A quantitative comparison to a redshift-, stellar mass-, and galaxy structure-matched non-AGN sample and deep spectroscopic survey of companions is required to fully understand the role of different mechanisms in triggering AGN activities and if AGNs are special with respect to the general galaxy population.

AGN-host Centroid Offset
Off-nucleus AGNs may indicate the presence of close dual supermassive BHs in the inspiraling phase during galaxy merger or recoiled BHs from gravitational kickout of the binary coalesces (e.g., Baker et al. 2006;Campanelli et al. 2007;Barth et al. 2008;Comerford & Greene 2014).We investigate the AGN-host offset in our X-ray broad-line sample, utilizing the results from image decomposition results without tying the centers of the PSF and Sérsic components.Here, the PSF-Sérsic offset refers to that between the AGN and its associated stellar core, rather than between the AGN and a close companion in a double-core system, such as CID-42 (Li et al. 2024b).Figure 13 shows the normalized probability of AGN-host offset in five filters used in this work.We find that the median offsets in all five filters are all quite small (0.4 kpc), suggesting good alignment of the AGN and the galaxy center.
However, the median offset and the spread increase systematically toward shorter wavelengths.This can be driven by physical reasons such as nonaxisymmetric spatial distribution of extinction and stellar populations so that their presence may depend on wavelength.An example is shown at the upper inset of Figure 13.Galaxies that are more asymmetric and bright star-forming clumps are common at z  1 (e.g., Elmegreen et al. 2007;Guo et al. 2015),  which may be due to fragmentation of massive gas disks driven by gravitational instabilities (e.g., Noguchi 1999;Dekel et al. 2009).Wuyts et al. (2012) found that off-center clumps can contribute up to ∼20% to the integrated SFR, but only 7% or less to the total mass in massive star-forming galaxies at z ≈ 1-2, consistent with decreasing AGN-host offset toward longer wavelength.Meanwhile, AGN-host offset can be entirely due to intrinsic faintness of the host associated with low f host in rest-frame UV, as illustrated at the lower inset of Figure 13.Zhuang & Shen (2023) have demonstrated that artificial offsets are commonly found in hosts with low host surface brightness using mock AGNs.
The observed residual AGN-host offset in the F277W and F444W is small (∼0.2 kpc, <1 pixel).These residual offsets could be due to the presence of nuclear substructure or PSF mismatch.The long tail of the distribution is due to a combination of faint nucleus and asymmetric light distribution from interactions (Section 6.2).Nonetheless, the high resolution of JWST NIRCam provides more stringent constraints on AGN-host offset compared to previous results based on HST images and Gaia astrometry (e.g., Shen et al. 2019;Li et al. 2023).Therefore, our results suggest that there is no significant AGN-host offset in the majority of X-ray broad-line AGNs.The genuine detection of an off-nucleus AGNs must require a well-defined host centroid and a significant AGN-host centroid offset beyond systematic uncertainties.This topic will be addressed in a follow-up paper.et al. (2020) to M BH for consistency.In total, we have 30 objects at 1.0 < z < 2.5 (median redshift 1.8).CID-452 is covered by both samples with reasonable agreement in M BH between two measurements (0.28 dex difference).We adopt a typical uncertainty of 0.4 dex for M BH estimated from single-epoch method.
Figure 14 shows M BH versus M * relation for our 1.0 < z < 2.5 X-ray broad-line AGNs.Taking at face value, 1.0 < z < 2.5 X-ray broad-line AGNs follow the relation defined by local inactive early-type galaxies but lie systematically above local inactive latetype galaxies (Greene et al. 2020).Comparing with nearby (z 0.35) optical broad-line AGNs from Zhuang & Ho (2023), 1.0 < z < 2.5 AGNs are preferentially lying near the upper envelope of the distribution of low-z AGNs.
However, we caution that the position of AGNs on the M BH -M * plane can be significantly affected by various factors.These factors include the selection bias induced by the flux limit of the spectroscopic observations, the AGN duty-cycle dependent observational bias, and substantial measurement uncertainties in the mass terms (Lauer et al. 2007;Shen & Kelly 2010;Schulze & Wisotzki 2011;Li et al. 2021a).As a result, both high-z and local AGNs may be biased tracers of the intrinsic mass relation of the underlying SMBH population.Specifically, at high redshifts, the detection limit of the BH mass is limited by the depth of spectroscopic follow-up observations.Therefore, the observed mass relation is biased toward the most luminous and massive BHs.Meanwhile, in the nearby universe, the AGN duty cycle tends to drop significantly toward massive BHs (as opposed to the roughly M BH -independent duty cycle at z > 1; e.g., Schulze et al. 2015), which means that more massive BHs have a smaller probability to be observed as AGNs.Consequently, nearby AGNs may preferentially populate the lower envelope of the M BH -M * plane (e.g., Schulze & Wisotzki 2011;Li et al. 2021a).It is also possible that the mass relations of high-z and local AGNs are intrinsically different due to evolution.For example, Zhuang & Ho (2023) showed that the position of AGNs on M BH -M * plane is connected with the properties of BH accretion and star formation.Considering more intense star formation and BH activities toward higher redshift, the M BH -M * relation may evolve with redshift.
We have not independently measured the broad-line AGN spectra to derive single-epoch BH masses using the latest recipes (Shen et al. 2023b).The significant uncertainties associated with the single-epoch M BH estimates (∼0.35 dex for Hβ and Mg II and ∼0.5 dex for C IV; Shen et al. 2023b), along with a potential systematic luminosity-dependent bias that results from uncorrelated variations between AGN luminosity and broad-line width, could significantly modify the intrinsic M BH -M * distribution by flattening its slope and increasing its scatter (Shen 2013;Li et al. 2021a).While conducting an indepth analysis of these biases is beyond the scope of this paper, we plan to uniformly analyze all available spectroscopic data covering the COSMOS-Web footprint to measure broad lines and derive more reliable BH masses in a future work.We will also perform a detailed modeling of selection biases and investigate the intrinsic M BH -M * relation and its cosmic evolution at 0.5  z  3.5 using the most accurate BH mass and stellar mass measurements.Wisp significance is defined as the median ratio within a small window targeted at the brightest wisp feature between the wisps to be subtracted and the background noise (σ).Vertical dashed black line indicates our criterion (>0.1σ) for wisp subtraction: B4 of F115W and A3, A4, B3, and B4 of F150W.

Figure 1 .
Figure 1.Pseudo-color image of the COSMOS-Web observation number 009 constructed from the F444W (red), F150W (green), and F115W (blue) filter images.The scale bar indicates one arcminute.Up direction indicates north and left direction indicates east.

Figure 2 .
Figure 2. Comparison of an F150W mosaic at Observation 047 between the public data release 0.2 by the COSMOS-Web team (top row) and our reduction (bottom row).Mosaics have been smoothed using a Gaussian kernel with a FWHM of 3 pixels.The left-hand column highlights differences in background subtraction and the right-hand column illustrates the removal of artifact features, i.e., wisps (near the bright star at the top of the image) and claws (near the right edge of the image).

Figure 3 .
Figure 3. PSF FWHM (global model) vs. observation time for four JWST NIRCam filters.Blue, green, orange, and red colors indicate F115W, F150W, F277W, and F444W, respectively.Horizontal dashed lines and shaded areas represent median PSF FWHM and 16th-84th percentiles, with the corresponding statistics shown at the upper right-hand corner.The right-hand panels present histograms of PSF FWHM.

Figure 4 .
Figure 4. Bolometric luminosity (L bol ) vs. redshift for the 143 X-ray broad-line AGNs (red circle) within the current COSMOS-Web footprint, and the parent X-ray AGN sample from Marchesi et al. (2016a) containing all AGN sources with either photometric or spectroscopic redshifts and L X 10 42 erg s −1 in the entire COSMOS field.

Figure 5 .
Figure 5.A gallery of F277W images of our AGN sample in the order of ascending redshift.Object name and redshift are shown at the top and the scale bar of 5 kpc is shown at the lower left-hand corner.Note that some nuclei have saturated central pixels, e.g., CID-47, CID-166, and CID-352.
objects inZou et al. (2019) and one object inSuh et al. (2020) have absolute stellar mass difference larger than 1 dex from our measurements.We have verified that the host emissions from these five objects are robustly detected in our image decomposition, and that the fits to their host SEDs are reliable and physical.As demonstrated in Zhuang & Ho (2023) using mock AGNs generated from real galaxy SEDs, fitting composite AGN+host SEDs without image decomposition can first easily lead to overestimated or underestimated stellar masses as a result of the degeneracy between AGN templates and stellar templates.The systematic overestimation of M * increases with greater AGN contribution and bluer host color, and vice versa for the underestimation of M * .Meanwhile, inconsistent fluxes measured from surveys with different resolutions, sensitivities, and apertures can lead to large scatter.Different choices of parameters of AGN and stellar templates can also lead to systematic differences, as illustrated by the large discrepancy betweenZou et al. (2019) andSuh et al. (2020).Our method, which independently decomposes AGN and host using highresolution imaging, does not suffer from extra complications from adopting different AGN templates.Therefore, our comparison suggests that stellar masses estimated from fitting combined AGN+host SEDs are less reliable, with complicated

Figure 6 .
Figure 6.Multiwavelength simultaneous AGN+host image decomposition to five-band ACS+NIRCam images of the X-ray broad-line AGN CID-307 at z = 2.051 using GALFITM.Images of data, model (AGN+host), data−nucleus, and residual (data−model)/error are shown from left-to right-hand.The best-fit parameters for the host galaxy and host-to-total flux fraction f host are shown at the lower left-hand corner in the model panel.The right-hand column shows the surface brightness (μ) profiles of data and models along the major axis, with the bottom panel showing the difference between data and model (Δμ = μ data − μ model ).

Figure 7 .
Figure7.Absolute difference of host parameters ( ḿSersic : upper row, R e : middle row, and n: bottom row) in duplicate observations of same objects vs. f host .Purple, blue, green, orange, and red dots and errorbars represent measurements and their uncertainties reported by GALFITM in the F814W, F115W, F150W, F277W, and F444W filters, respectively.Black dashed horizontal lines indicate the typical uncertainties added in quadrature to the nominal error reported by GALFITM.See detailed descriptions in the text.

Figure 8 .
Figure 8. Four examples of SED fitting to the AGN-subtracted host photometry using CIGALE across the redshift range of 0.35  z  3.5.The observed fluxes are shown in purple open circles.The best-fit total model spectrum is shown in the black solid curve, with its components of stellar attenuated emission and nebular emission shown in the yellow and green solid curves.Red filled circles represent the model fluxes and the dashed blue curve represents the stellar emission before attenuation.Reduced χ squared (c n 2 ) of the fit is shown at the lower left-hand corner.Relative residuals (Data−Model)/Data are shown in the lower panel.

Figure 9 .
Figure 9. Histogram of Sérsic index n at rest-frame 1 μm.Red vertical dashed lines indicate 15th, 50th, and 84th percentiles of the sample with statistics shown at the upper right-hand corner.

Figure 10 .
Figure 10.Host stellar mass-size relation for X-ray-selected broad-line AGNs.Size is R e interpolated to rest-frame 5000 Å. Blue and red shaded areas represent the 16th-84th percentile distributions of star-forming and quiescent non-AGN galaxies from van der Wel et al. (2014), respectively.Blue and red dots represent starforming and quiescent AGN hosts classified using the UVJ diagram, while black dots represent AGNs without robust classification given the uncertainty of colors and proximity to the boundary.Open symbols indicate objects with R e < 1 pixel.Median sizes as a function of stellar mass are shown with solid curves for objects at 1 < z 2.5.Error bar indicates typical uncertainty.

Figure 11 .
Figure11.Major merger fraction vs. redshift (top panel) and AGN bolometric luminosity (bottom panel).Blue and red colors indicate upper and lower limit of major merger rate, respectively.The numbers of all objects (black), and the upper limit (blue) and lower limit (red) of major merger candidates in each bin (separated by black vertical dashed lines) are shown at the top.

Figure 12 .
Figure 12.A gallery of normalized residual images (residual/error) of our sample in the F277W filter after subtracting best-fit PSF and Sérsic models in the order of ascending redshift.Object name and redshift are shown at the top and the scale bar of 5 kpc is shown at the lower left-hand corner.The colorbar is shown in the bottom right-hand panel.Nonaxisymmetric features are ubiquitous in X-ray-selected broad-line AGN hosts.
6.4.The BH Mass-Host Stellar Mass Relation Suh et al. (2020) and Ding et al. (2020) presented M BH measurements for their samples using single-epoch spectra.Since no line width and luminosity are provided in Suh et al. (2020), we adopt their M BH calibration and convert line measurements in Ding

Figure 13 .
Figure 13.Normalized probability density from Gaussian kernel-density estimate of AGN-host offset in the F814W (purple), F115W (blue), F150W (green), F277W (orange), and F444W (red) filters.Medians and the differences between 84th and 16th percentiles are shown at top right-hand corner.Inset figures show F814W data, F814W data−nucleus, F277W data, and F277W data−nucleus of CID-277 (top) and CID-731 (bottom) from left-to right-hand.Blue cross the location of AGN and black plus represents the location of host galaxy.Figure 14.The relation between M BH and M * relation for 1.0 z 2.5 X-ray broad-line AGNs (red triangles: z < 1.8; red dots: z 1.8).Gray contours and dots indicate 10%, 30%, 60%, 80%, and 95% of the entire distribution and individual objects outside the 95% contour of z 0.35 broad-line AGNs from Zhuang & Ho (2023), respectively.Solid and dashed blue lines represent bestfit relations for local early-type inactive galaxies and late-type inactive galaxies from Greene et al. (2020), respectively.The errorbar indicates typical uncertainty.

Figure 14 .
Figure 13.Normalized probability density from Gaussian kernel-density estimate of AGN-host offset in the F814W (purple), F115W (blue), F150W (green), F277W (orange), and F444W (red) filters.Medians and the differences between 84th and 16th percentiles are shown at top right-hand corner.Inset figures show F814W data, F814W data−nucleus, F277W data, and F277W data−nucleus of CID-277 (top) and CID-731 (bottom) from left-to right-hand.Blue cross the location of AGN and black plus represents the location of host galaxy.Figure 14.The relation between M BH and M * relation for 1.0 z 2.5 X-ray broad-line AGNs (red triangles: z < 1.8; red dots: z 1.8).Gray contours and dots indicate 10%, 30%, 60%, 80%, and 95% of the entire distribution and individual objects outside the 95% contour of z 0.35 broad-line AGNs from Zhuang & Ho (2023), respectively.Solid and dashed blue lines represent bestfit relations for local early-type inactive galaxies and late-type inactive galaxies from Greene et al. (2020), respectively.The errorbar indicates typical uncertainty.

Figure 15 .
Figure 15.Wisp templates in the F150W filter for COSMOS-Web observations created by median-stacking the source emission-masked images for four affected SW detectors: A4 (top left-hand panel), A3 (bottom left-hand panel), B3 (top right-hand panel), and B4 (bottom right-hand panel).

Figure 16 .
Figure16.Wisp significance in different detectors (A3: red, A4: blue, B3: orange, and B4: green) and filters (F115W: open step and F150W: filled step).Wisp significance is defined as the median ratio within a small window targeted at the brightest wisp feature between the wisps to be subtracted and the background noise (σ).Vertical dashed black line indicates our criterion (>0.1σ) for wisp subtraction: B4 of F115W and A3, A4, B3, and B4 of F150W.

Table 1
Properties of NIRCam PSF PSF FWHMs and temporal variations are based on the global PSF model of individual observations.Spatial variations are based on point-like sources in the mosaics.temporaland spatial variations in terms of FWHM.The temporal variation of PSF FWHM decreases from ∼2.8% (F115W) to 0.8% (F444W) and is dominated by short timescale fluctuation.The spatial variation of PSF FWHM in individual observations are much larger compared to the temporal variation and dominated by random variation, with 5% for SW filters and ∼2% for LW filters.The linear spatialdependent model (i.e., the local PSF model) is not accurate enough to model the random spatial PSF variations in individual observations and thus is not recommended for further analysis.Besides variations due to instrumental effects, the stochastic aliasing of the PSF in the F115W and F150W filters, the number of sources used for PSF construction, the dither pattern, and the accuracy of astrometry alignment for individual sources in different exposures may contribute to the observed temporal and spatial variations.We recommend the use of global, broad, and narrow PSF models for COSMOS-Web NIRCam mosaics.The PSF models along with reduced mosaics of each observation in each filter will be made publicly available to the community in Dataverse via doi:10.7910/DVN/1GDKDY.Users can adopt the one that best models the profile of the target of interest.