Lyman Continuum Emission from Active Galactic Nuclei at 2.3 ≲ z ≲ 3.7 in the UVCANDELS Fields

We present the results of our search for Lyman continuum (LyC)-emitting (weak) active galactic nuclei (AGN) at redshifts 2.3 ≲ z ≲ 4.9 from Hubble Space Telescope (HST) Wide Field Camera 3 (WFC3) F275W observations in the Ultraviolet Imaging of the Cosmic Assembly Near-infrared Deep Extragalactic Legacy Survey (UVCANDELS) fields. We also include LyC emission from AGN using HST WFC3 F225W, F275W, and F336W imaging found in Early Release Science (ERS) and Hubble Deep UV Legacy Survey data. We performed exhaustive queries of the Vizier database to locate AGN with high-quality spectroscopic redshifts. In total, we found 51 AGN that met our criteria within the UVCANDELS and ERS footprints. Out of these 51, we find 12 AGN that had ≥4σ detected LyC flux in the WFC3/UVIS images. Using a wide variety of space-based plus ground-based data, ranging from X-ray to radio wavelengths, we fit the multiwavelength photometric data of each AGN to a CIGALE spectral energy distribution (SED) using AGN models and correlate various SED parameters to the LyC flux. Kolmogorov–Smirnov tests of the SED parameter distributions for the LyC-detected and nondetected AGN showed they are likely not distinct samples. However, we find that the X-ray luminosity, star formation onset age, and disk luminosity show strong correlations relative to their emitted LyC flux. We also find strong correlations of the LyC flux to several dust parameters, i.e., polar and toroidal dust emission and 6 μm luminosity, and anticorrelations with metallicity and A FUV. We simulate the LyC escape fraction (f esc) using the CIGALE and intergalactic medium transmission models for the LyC-detected AGN and find an average f esc ≃ 18%, weighted by uncertainties. We stack the LyC fluxes of subsamples of AGN according to the wavelength continuum region in which they are detected and find no significant distinctions in their LyC emission, although our submillimeter-detected F336W sample (3.15 < z < 3.71) shows the brightest stacked LyC flux. These findings indicate that LyC production and escape in AGN are more complicated than the simple assumption of thermal emission and a 100% escape fraction. Further testing of AGN models with larger samples than presented here is needed.

In order to be compatible with the observed rapid reionization scenario of 100 Myr (Bolan et al. 2022), a second population of low-mass, star-forming galaxies with a high LyC f esc must have formed ubiquitously in the Universe from 6  z  8.These galaxies would require a stellar initial mass function such that enough supernovae could clear out neutral hydrogen in the interstellar medium (ISM) via outflows and winds to allow highmass, main-sequence stars to emit the sufficient amount of LyC to reionize the IGM with f esc ; 10%-20% (Finkelstein et al. 2019;Yung et al. 2020aYung et al. , 2020b;;Mutch et al. 2024).These conditions may have been met given the rising star formation histories in some bright z > 7 galaxies observed with JWST (e.g., Finkelstein et al. 2022;Giménez-Arteaga et al. 2023;Robertson et al. 2023;Tacchella et al. 2023).Alternatively, a significant population of low-luminosity AGN at z > 6 could contribute substantially to the completion of hydrogen reionization (Giallongo et al. 2015;Madau & Haardt 2015;Khaire et al. 2016;Grazian et al. 2018Grazian et al. , 2020Grazian et al. , 2022;;Yung et al. 2021;Onoue et al. 2023).This AGN reionization paradigm is reinforced by the discovery of 10 8 -10 9 M e quasars at z > 7 when the Universe was ∼750 Myr old (Mortlock et al. 2011;Bañados et al. 2018;Wang et al. 2018Wang et al. , 2021;;Matsuoka et al. 2019;Yang et al. 2019Yang et al. , 2020)), which indicates high-efficiency accretion onto supermassive black holes (SMBHs) while the IGM was being reionized.Furthermore, recent studies of the z  5 AGN luminosity function and space density suggest that faint (M 1450 ; -23) AGN are ∼3-5 times more abundant at z ∼ 5 than found in previous work (e.g., McGreer et al. 2018;Kim et al. 2020;Niida et al. 2020), and a slower evolution of the space density from 3  z  6 is observed (e.g., Giallongo et al. 2019;Grazian et al. 2022;Fontanot et al. 2023).
It is most likely the case that stars in massive and low-mass galaxies, as well as AGN, contributed to reionization.The time line for which sources dominated the ionizing background, and the characteristics of sources with high f esc remain unclear.In this work, we focus on characterizing the physical properties of AGN with high and low LyC emission using imaging from the Hubble Space Telescope (HST) Ultraviolet Imaging of the Cosmic Assembly Near-infrared Deep Extragalactic Legacy Survey Fields (UVCANDELS) program (Wang et al. 2023), which expands on our work from Smith et al. (2018Smith et al. ( , 2020;;hereafter S18 and S20, respectively) using Wide Field Camera 3 (WFC3)/UVIS Early Release Science (ERS; Windhorst et al. 2011) and Hubble Deep UV Legacy (HDUV) survey (Oesch et al. 2018) imaging.These constraints can be extrapolated to AGN found during the reionization epoch, and provide further insight into the role of AGN as reionizers of the IGM, since AGN spectral energy distribution (SED) shapes are known not to evolve with cosmic time (e.g., Shen et al. 2019;Yang et al. 2021).We present our analysis as follows: in Section 2 we describe the data we use for the LyC measurements.In Section 3 we detail our sample selection of AGN used for our LyC studies.In Section 4 we review our CIGALE SED fitting configuration and the ancillary photometry used for fitting.In Section 5 we outline our LyC photometry and f esc methodology.In Section 6 we present our results, in Section 7 we discuss our results, and in Section 8 we summarize our conclusions.We assume Planck Collaboration (Aghanim et al. 2020) cosmology (flat Lambda cold dark matter (ΛCDM), with H 0 = 67.4km s −1 Mpc −1 and Ω m = 0.315) and use AB magnitudes (Oke & Gunn 1983) throughout.

Data
Our main imaging data used for studying the LyC emission from AGN were collected as part of the UVCANDELS program (Teplitz 2018).This UV imaging covers the COSMOS, EGS, and completes the GOODS North and South fields in F275W with a three orbit depth and coordinated parallels using the F435W filter with the Advanced Camera for Surveys (ACS).The F435W exposures have some variation in depth due to roll angle constraints and overlap with other exposures.These data can be found on MAST via doi:10.17909/fw24-0b81.UVCAN-DELS also includes ground-based Large Binocular Telescope U-band imaging of GOODS North and COSMOS that probes down to m AB ∼ 28 mag (Ashcraft et al. 2018(Ashcraft et al. , 2023;;Otteson et al. 2021;Redshaw et al. 2022).We also include UV imaging from the HDUV survey (Oesch et al. 2018), the Hubble Ultraviolet Ultra Deep Field (UVUDF; Teplitz et al. 2013;Rafelski et al. 2015), and ERS field (Windhorst et al. 2011) in F225W, F275W, and F336W for our AGN LyC analysis when available.The various depths reached by each survey are captured in the photometric uncertainties (see Section 4.1).Further details on the image quality and drizzling parameters of the UVCANDELS imaging can be found in Wang et al. (2023).
The diversity of our sample is further exemplified in Figure 1.Here we show distributions of several parameters of the best-fitting CIGALE (Boquien et al. 2019) SEDs (see Section 4 for a full description of our fitting analysis) with LyC signal-to-noise ratio (S/N) measurements from their respective WFC3/UVIS images of 4σ and <4σ shown as blue and red bars, respectively.We chose 4σ as the detection threshold since the other 39 AGN below 4σ showed no obvious LyC flux.The M AB in the first panel is the monochromatic 1450 Å luminosity calculated from the best-fitting CIGALE SED by integrating the flux between λ rest = 1449-1551 Å (e.g., Kulkarni et al. 2019).
We take the remaining parameters from the physical property estimation variables defined in the CIGALE configuration file.Accordingly, the adopted A V is taken from the Av_ISM parameter in the attenuation component of the full SED model, SFR now is defined as the sfr in the sfh component, Z is the metallicity parameter in the stellar component, is taken as the sum of the values m_gas_old, m_gas_young, m_star_old, and m_star_young in the stellar component, the X-ray photon index Γ is the gam parameter in the xray component, L stars is equal to the sum of the lum_old and lum_young parameters from the stellar component, L dust abs is the luminosity absorbed by dust taken from the luminosity parameter in the dust component, L disk is the disk_luminosity parameter in the agn component, L torus is the torus_dust_luminosity parameter in the agn component, L dust AGN is the emitted dust luminosity taken from the total_dust_luminosity parameter in the agn component, and L X-ray is the agn_Lx_total parameter in the xray component.K-S tests of each distribution indicate that the two samples are likely from the same larger sample, as the null hypothesis cannot be rejected.Therefore, high-S/N LyC flux emitted from AGN may not be discernible from AGN emitting fainter LyC flux when comparing these parameters.
Our sample consists of a somewhat fainter population of AGN, all fainter than M * (M AB ; -27; see Kulkarni et al. 2019) at the average redshift of z ∼ 3. Their SED fits exhibit dusty, weak AGN characteristics, with the stellar component being the dominant source of their intrinsic luminosity for ∼35% of the sample, while dust dominates the luminosity for ∼45%.Only ∼4% of the sample has the AGN disk component as their brightest feature.Grazian et al. (2018) and Romano et al. (2019) performed a similar study to our present work on AGN with -25.1 M UV -23.3 and -29.0 M UV -26.0, respectively.
Most of their AGN are brighter than ours, which likely explains the ubiquity of detections among their samples.Thus, this work serves as a complement to the study of more luminous quasar LyC detections, which do not include LyC nondetected AGN.Their escape fractions are also much higher than the present sample, even for our LyC-detected AGN (see Section 1), with the lowest f esc = 44% from the Grazian et al. (2018) sample.Compared to the Wang et al. (2023) galaxy sample, 19/51 AGN host galaxies lie in the same range of -22 < M UV < -18, with four of these being LyC detected.Here, M UV is the monochromatic 1500 Å absolute magnitude calculated from the best-fit CIGALE SEDs, excluding the AGN components.Eight AGN host galaxies are brighter than −22 mag, two of which are LyC detected, and 24 are fainter than −18 mag, six of which are LyC detected.This likely indicates that the AGN itself is emitting most of the detected LyC rather than the host galaxyʼs massive stars.Compared to the Steidel et al. (2018)  ).However, we find that 7/51 AGN host galaxy stellar components exceed this ratio, by an order of magnitude in three cases.Two of these were detected in LyC, namely S-CANDELS J123714.30+621208.5 and WISEA J123623.00 +621526.8, with a ratio of ∼9 and ∼20, respectively.WISEA J033209.44-274807.3 had an exceptional Å M AB 1700 ; −25.13 for its non-AGN SED components, though it was one of the four bright nondetections.
The X-ray luminosity outshines all other SED components for only ∼16% of the sample, indicative of mostly obscured AGN, and all X-ray-detected AGN display a soft photon index (Γ < 2).The population also shows significant star formation, with 50% of the sources showing a star formation rate (SFR) between 10 and 100 M e yr −1 , and ∼37% at SFR > 100 M e yr −1 , with only the remaining 10% at SFR < 10 M e yr −1 .The metallicities are more diverse, with ∼32% showing Z < 0.01, and ∼28% having Z 0.02, and the rest of the sample having 0.01 Z 0.02.

Photometry
The photometry used for the SED fitting was compiled from region-based queries of the NED database.We compiled all available photometry within 1 5 of the centroid of the corresponding χ 2 image mentioned in Section 3.This position offset tolerance was chosen so that our query would capture photometry with larger point-spread functions, e.g., from X-ray and ground-based data.We homogenized the retrieved photometry to units of mJy based on the instrument and band for all available photometric data.We then added the photometric data point with the highest S/N for each band to our photometric catalog if there were more than one measurement for a particular band.We also filtered out photometric bands from our query that did not have available transmission curves to be used for SED fitting.
The resulting photometric catalog included data from the 2.2 m MPG/ESO WFI (B, V, R, and I), Two Micron All Sky Survey (2MASS) J and K s , AKARI S11 and L18W, ATCA CABB and 1.4 GHz bands, AzTEC 1.16 mm, Blanco/DECam (g, r, i, z, and Y), Blanco/MOSAIC-II Hα, CFHT/WIRCam J and K, Chandra, COMBO-17, Gemini/QUIRC H and K, Hale/WIRC J and K s , Herschel (PACS and SPIRE), HST (ACS/WFC F435W, F606W,  Kulkarni et al. 2019).A V is taken from the Av_ISM parameter of the attenuation component, SFR now is from sfr in sfh, Z is from metallicity in stellar, is the sum of m_gas_old, m_gas_young, m_star_old, and m_star_young in stellar, Γ is from gam in xray, L stars is the sum of lum_old and lum_young from stellar, L dustabs is from luminosity in dust, L disk is from disk_luminosity in agn, L torus is from torus_dust_luminosity in agn, L dustAGN is from total_dust_luminosity in agn, and L X-ray is the agn_Lx_total parameter in the xray component.Table 3.

CIGALE Configuration
We used the CIGALE (Boquien et al. 2019) SED fitting code to model the AGN emission from X-ray to radio wavelengths.These models enable us to calculate the intrinsically produced LyC of the AGN and host galaxy.We used the model and templates to model the star formation history and stellar populations, templates to simulate emission lines from the ISM, model to simulate dust attenuation, templates to fit the AGN emission, templates to simulate the thermal dust emission, model to simulate the X-ray emission from stars and the AGN, and a module to simulate galaxy synchrotron and AGN emission during fitting.
These components are a combination of mathematical models with free parameters and templates with predefined choices of parameters, i.e., the bc03, nebular, dale2014, and skirtor2016 templates.All of the templates, however, are in turn based on physically motivated, empirical mathematical models (Ferland et al. 1998(Ferland et al. , 2013;;Bruzual & Charlot 2003;Stalevski et al. 2012Stalevski et al. , 2016)).These templates, of course, limit the parameter space by the resolution and range of the parameters in the templates provided by CIGALE.Incorporating additional models not included in the main distribution of CIGALE could improve this.Another limitation we observed comes from how the component scales the LyC to provide energy to emission lines.When the observed LyC flux exceeds the modeled LyC flux (which includes a wavelength-dependent IGM attenuation factor at redshift z), the LyC scaling may produce erroneous results.This case can occur for high-f esc galaxies and AGN, since higher f esc values have likely been observed through lines of sight with higher than average IGM transmission.When including photometry below the rest-frame Lyman limit, the modeled LyC gets scaled to meet this observed LyC flux, creating an unphysical "jump" or discontinuity at 912 Å.To circumvent this, we did not fit any SEDs to photometry measured at rest-frame wavelengths shorter than 1216 Å.
We also modified the code to extrapolate the power-law fit of the X-ray model to 1000 Å specifically for our AGN since the AGN continuum power law shows a break at around this wavelength (Telfer et al. 2002), and also to fill the discontinuity in the model between 220  Å  400 that exists without this modification.This will also account for the hard-UV tail of the AGN warm corona emission (e.g., Lusso et al. 2015;Petrucci et al. 2020).

Lyman Continuum Photometry and Escape Fractions
We performed all LyC photometry as we have in our previous LyC work from S18 and S20.In summary, we treat each pixel as a random variable with normally distributed flux, where the mean is set to the pixel value from the drizzled science image, and the dispersion is set to the rms value from the corresponding pixel in the weight map, plus the local sky dispersion measured in the 151 × 151 pixels (4 53 × 4 54) surrounding the target, normalized to each individual pixel.Neighboring object pixels were also masked in this process.We then generated 10,000 randomly drawn 151 × 151 pixel images using these statistics, and measured the flux within a Kron-like elliptical aperture detected in the image for each realized image (created using all available HST data; see Szalay et al. 1999) using SExtractor (Bertin & Arnouts 1996).The resulting flux distribution from these measurements was used to determine the LyC flux and its uncertainties given in Table 1.When performing photometric measurements on our LyC stacks (see Section 5.1 for details on our stacking procedure), we treated each pixel in each subimage of a stack as a random variable which creates a 151 × 151 × 10,000 data cube for that subimage, then stacked each data cube using a weighted sum.We then normalized the stack of data cubes by a stack of all weight map cutouts.The flux distribution was then generated in the same way as the individually detected AGN, taking slices from the data cube and measuring the central isophotal flux, using the image for detection.The resulting stacked photometry and its uncertainties are presented in more detail Section 5.1 and Table 2.
To calculate the LyC f esc , we used the same method as in S18 and S20.In our method, we simply take the ratio of the observed LyC flux to an intrinsic LyC flux modeled using the unattenuated SED (see Figure 2), the WFC3/UVIS filter curve corresponding to the observed LyC flux filter, and IGM attenuation curves modeled at the redshift of the target AGN.In short, we modeled the intrinsic LyC flux from the best-fitting CIGALE SED using only the following components: stellar.old,stellar.young,nebular.lines_young,nebular.lines_old,nebular.continuum_young,nebular.continuum_old,agn.SKIRTOR2016_disk, agn.SKIRTOR2016_torus, agn.SKIRTOR-2016_polar_dust, xray.galaxy,xray.agn,radio.sf_nonthermal,and radio.agn.We summed these components to acquire our intrinsic SED, then we calculated the inner product of this SED with the IGM attenuation curves and filter transmission curves to obtain our final intrinsic LyC flux model.We performed this inner product 10,000 times for the variety of sight lines through the IGM at the redshift of the AGN using the model from Inoue et al. (2014).We used the flux distribution described in the previous paragraph as the observed LyC flux, took the ratio of these two distributions, and used this ratio to calculate the f esc distribution and its statistics, which are presented in Table 1.It is worth noting that the modeled LyC from the SED, as well as the apertures shown in Figure 3, are intended to capture both the LyC emission from the central AGN as well as any stellar LyC leaking from the host galaxy.
We find a wide range of f esc values from 1% to 100% or greater, which represent the mode of the f esc distribution.The corresponding 1σ uncertainties, along with the f esc values, are shown in Table 1.An inverse-variance weighted-average value including all values in Table 1 (with a ceiling of 100%) equates to approximately f esc ; 18%.f esc distributions with modes above 100% result from a higher observed LyC compared to the modeled, intrinsic LyC obtained from the best-fit CIGALE SED.Although these f esc values are unphysical, and at a ∼2.7σ significance, we show the f esc in these cases to highlight the need for AGN models that can accommodate high LyC emission.

Lyman Continuum Stacking
Stacking is a technique generally used to increase the S/N of signals from sources of interest, or to find the average signal of a sample.In this work, we stack LyC flux from galaxies hosting AGN to compare the total LyC signal brightness of various subsets within our sample based on observed spectral features.
To distinguish these subsets, we used the archival photometry retrieved from the various catalogs hosted on Vizier and categorized the detected photometric bands into regions of the electromagnetic spectrum, namely X-ray, optical, IR, MIR, FIR, submillimeter, microwave, and radio.We did not include a UV-detected sample since U band may overlap with the restframe LyC for some higher-z AGN.Each of these subsamples is defined by having a detection in their respective region of the electromagnetic spectrum.The bandpass filters used for determining the inclusion in each subsample can be found in Section 4.1.
We show the resulting stacks of our AGN subsamples in Figures 4 and 5, and the photometry performed for each stack is presented in Table 2.We find the stack with the brightest LyC emission is the submillimeter-detected AGN in the F336W stack, with redshifts that range from 3.15 < z < 3.71.There are only three objects in this stack, so it is likely there is a detection bias since detection in submillimeter often implies an overall brighter luminosity across the SED.The brightest LyC stack of all AGN detected in a particular region of the electromagnetic spectrum was the IR-and MIR-detected AGN LyC stacks, which include 50 and 49 out of the 51 AGN, respectively.The AGN missing from these stacks simply did not have any IR/MIR band observations.A zero-pointnormalized stack (the "All LyC" row in the All AGN set of stacks in Table 2) was brighter than these two stacks, and includes all 51 AGN.We also stack the LyC nondetections in each of our three redshift bins, as well as a stack of all nondetections.None of these stacks resulted in detected LyC.
These results demonstrate that stacks based on detected regions in the AGN electromagnetic continuum do not correlate well with their LyC emission.There are many factors that can affect the LyC flux that is detected, e.g., sample size statistics, IGM line-of-sight attenuation, AGN inclination, observed filter and its proximity to the Lyman limit, or other intrinsic physical factors of the host galaxy or AGN.A correlation analysis on an individual basis of each AGNʼs LyC emission may reveal more clues regarding which regions of the continuum could indicate LyC leakage.(2) redshift range of the galaxies included in LyC/UV continuum (UVC) stacks.Overlapping ranges for the low-and high-z "All" samples are due to splitting the F225W, F275W, and F336W samples by their median redshift before combining the three subsamples into either the high-or low-z "All" sample; (3) average redshift of all galaxies in each stack; (4) number of galaxies with reliable spectroscopic redshifts included in each stack; (5) observed total AB magnitude of LyC emission from the stack Source Extractor MAG_AUTO) aperture matched to UVC, indicated by the blue ellipses in Figures 3-5 We also created stacks based on the filter cutoffʼs proximity to the AGNʼs observed-frame Lyman limit.We stacked the LyC cutouts in two redshift bins (low z and high z) separated by their median redshift for four samples in the "All AGN" set of stacks, i.e., the F225W, F275W, F336W, and "All" samples.In all cases, we see that the low-z substacks, which would have their Lyman limit closer to the filter cutoff, show brigher LyC emission than the high-z substacks.This result is not surprising since a given filter would observe shorter-wavelength flux for higher-redshift galaxies, and the central AGN and stellar sources produce less LyC flux further from the Lyman limit.From this, we conclude that LyC observations should attempt to select targets with Lyman limits as close as possible to their filter cutoffs, while minimizing red leak in the filter beyond the Lyman limit below some fiducially low percentage.
Of the stacks based on redshift ranges which correspond to the observed WFC3/UVIS F225W, F275W, and F336W filters, we find that the stack of all galaxies in the F336W sample (3.08 < z < 4.88, 〈z〉 ∼ 3.5924) containing 16 AGN was the brightest substack with m AB = 27.66 ± 0.56.This result is interesting, since the higher-redshift AGN should experience more attenuation from the IGM.Furthermore, AGN activity peaks at z ∼ 2-3 in space density (Kulkarni et al. 2019), and the AGN in this stack are expected to be intrinsically fainter.Since AGN do not exhibit evolution in their SED with redshift (e.g., Barth et al. 2003;Fan et al. 2004;Iwamuro et al. 2004;Jiang et al. 2007), the LyC emission from AGN may not be redshift dependent either.Thus, it is more likely the case that LyC emission from AGN is solely dependent on the physical parameters of the AGN and host galaxy, or inhibited to some degree by the surrounding IGM.

Results
As seen in Section 3, our AGN with LyC flux detected at >4σ and those with LyC measured below 4σ are most likely not distinct samples when considering the SED parameters we used for characterization.In order to ascertain some parameter dependence on the LyC flux, we computed the Pearson correlation coefficient between the LyC flux, including upper limits, and several SED parameters, namely M AB , stellar mass, the ISM A V , metallicity, the instantaneous SFR, the ionization parameter U, the X-ray photon index, the AGN inclination, the luminosities of the young, old, and young + old stellar populations, the AGN disk, the AGN dust torus, AGN polar dust, the disk + torus + polar dust, the ISM dust, and the X-ray emission.Since our LyC photometric data are composed of mostly upper limits, the LyC flux is most appropriate to correlate parameters to, since scaling the UV sky background to high-redshift luminosity would only increase the uncertainty  1.The observed photometry is shown as light blue error bars, and the LyC detection is indicated as a purple error bar.The fitted SED is shown as a black curve, and the blue curve, used for estimating the escape fraction, includes the same SED components as the black curve, excluding all absorption components.The stellar component is shown as the red curve, the AGN component is shown as the green curve, the X-ray emission is shown as the magenta curve, and the radio emission is shown as the orange curve. in the data.Furthermore, flux correlations can inform future searches for AGN LyC leakage detections in imaging.
In Figure 6, we show LyC flux versus these parameters along with a Bayesian fit to the data.We used normal distributions to model the LyC flux probability for data with S/N > 1σ, and treat the LyC flux with S/N < 1σ as censored data, modeling these data using arbitrary probability distributions while constraining the log-cumulative distribution function to the 1σ uncertainty of the observed LyC flux.Since the points with highest S/N have the most weight in these fits, we also fit a line to the data excluding the point with the highest flux (plotted as a dashed red curve).As shown, this does not have a significant impact on the general direction of these lines, though the curves for metallicity and A FUV show the most change.Of the eight curves, five are negligibly different, and the curves for -L X 2 10keV and τ SFR are within the uncertainty ranges of the data.The p-values are generally higher without this bright point, which indicates that these correlations would significantly improve with more LyC-detected AGN included in the analysis.

Discussion
In general, our results indicate that higher LyC emission can be traced to intrinsically higher LyC production via the AGN accretion disk and host galaxy stellar sources.We compared subsamples of our AGN split into redshift bins and detected regions of the AGN continuum regions and find no significant distinctions.Our correlation tests show that high LyC-emitting AGN can be found from various markers that might indicate a high ionizing photon production, i.e., X-ray and IR/dust emission brightness.X-ray emission from AGN is created as a result of Comptonization of optical/UV photons produced by the accretion disk, which up-scatter off of a corona of hot electrons in close proximity to the SMBH into the X-ray regime (Haardt & Maraschi 1991).A significant fraction of these UV photons will be ionizing; therefore, the LyC that is not processed into X-rays can escape the host galaxy and ionize the IGM.Thus it is not surprising that X-ray luminosity correlates significantly with observed LyC flux.
Some of the X-ray emission is also reprocessed by the dust torus, the broad-line region, and the accretion disk.The toroidal and polar dust will become heated by the UV/optical light produced by the accretion disk, giving rise to extinction and luminous IR features in the SED.It is therefore reasonable to predict that brighter LyC production would result in brighter dust emission features, which is what we observed from correlating these parameters.The LyC that is not absorbed or reprocessed by dust will also have some probability of escaping the host galaxy, and more flux will be observed if more LyC is produced by the disk.The detected LyC emission may rely on the line of sight or angle at which the AGN is observed.
Stellar sources in the host galaxy will also contribute to the observed LyC since the accretion disk would mostly saturate LyC absorbers in the ISM and near the SMBH.This is true if the stellar sources can produce LyC in significant amounts, which are typically A-, B-, and O-type stars.The most massive O-type stars will produce the most LyC, though they typically only live for 10 Myr before going supernova, 200 Myr for B-types, and several hundred megayears for A types.Therefore, one would expect that recent star formation episodes producing newly formed massive stars would increase the observed LyC emitted by galaxies, which is corroborated by our correlation results.Furthermore, metal-poor stars should produce more LyC than more metal-rich stars since their photospheres would absorb less of the high-energy ionizing light, which we also observe in our correlation analysis.
We therefore infer that AGN with detectable LyC flux would be those with bright X-ray or thermal dust emission, with additional flux coming from the host galaxy stellar population produced within a few megayears of a star formation episode that produced lower-metallicity stars.Given the generally fainter luminosities of our AGN sample, a natural follow-up study would be to include brighter AGN with more robust LyC detections outside the UVCANDELS fields, such as those in Grazian et al. (2018) and Romano et al. (2019), which could potentially result in more accurate correlation fits.

Conclusions
We measured the escaping LyC flux from 51 AGN with highquality spectroscopic redshifts found in the UVCANDELS, ERS, HDUV, and UVUDF data after an exhaustive search of Vizier AGN databases.We found 12 AGN with LyC emission where the detection reached a S/N > 4σ.We used best-fitting CIGALE SEDs to model all of our AGN intrinsic flux and physical parameters using the highest-S/N photometric data retrieved from NED database queries, ranging from X-ray to radio.We used these models, along with the IGM Monte Carlo code of Inoue et al. (2014), to calculate the f esc of the 12 AGN with LyC detections.We correlated the SED parameters with the LyC flux of our AGN and fit lines to these correlations using a Bayesian regression, treating nondetections as censored data.We also stacked LyC cutout images of subsamples of AGN based on detected regions in the electromagnetic spectrum using the NED photometry, as well as subsamples of AGN based on the proximity of their LyC filter cutoff to their observed-frame Lyman limit.Our results are as follows: 1.The CIGALE SED fits of our 51 AGN indicate a weak, dusty, and mildly star-forming AGN population.From K-S tests, the 12 LyC-detected AGN are not distinguishable from the non-LyC-detected AGN in SED parameter histograms.The LyC-detected AGN f esc values show a wide range, from 1%  f esc  100%, or more (indicative of observed LyC flux outshining model predictions), with an inverse-variance weighted-average f esc ; 18%. 2. We find no significant correlations of the LyC emission for stacked subsamples based on their detections in the various regions of the AGN continuum.The brightest substack in LyC was the submillimeter stack in F336W (3.15 < z < 3.71), although this stack had three AGN and likely has a detection bias.The full stack of all 51 AGN was the brightest stack compared to stacks of all AGN detected in some region of the AGN continuum.The stack of all galaxies in the F336W sample (3.08 < z < 4.88, 〈z〉 ∼ 3.5924), containing 16 AGN, was the brightest substack with m AB = 27.66 ± 0.56, outshining the two lower-redshift bins at 2.28 < z < 2.30 and 2.38 < z < 3.08.After splitting each of these three samples by their median redshift into higher-and lower-redshift bins, then stacking these subsamples, we find that the lower-z substacks show brighter LyC flux.3. On an individual basis, we find significant correlation between the observed LyC flux and the 2-10 keV X-ray luminosity, the main stellar populationʼs star formation onset age, AGN disk luminosity, AGN torus luminosity, the AGN monochromatic 6 μm luminosity, and the AGN 1.4 GHz power.We also find significant anticorrelations of the measured LyC flux with stellar metallicity and A FUV .These findings indicate a connection to the emitted LyC flux and X-ray and thermal dust emission, and indicate that more recent star formation episodes produce higher LyC emission, where LyC emission also favors metal-poor galaxies.The strongest correlating parameter is the X-ray luminosity from 2 to 10 keV, indicating a strong X-ray to LyC emission connection for AGN.The SFH τ onset age parameter shows a strong correlation, which implies higher LyC emission after more recent star formation episodes.Several dust parameters (torus dust luminosity, polar dust luminosity, 6 μm luminosity, and A FUV ) show strong correlations with LyC flux, as well as 1.4 GHZ radio emission.Stellar metallicity shows a strong anticorrelation to LyC flux, which may signify a preference for higher LyC emission from metalpoor galaxies.
Although these correlations may not be exactly linear in nature, they do suggest a monotonically increasing relationship with observed LyC flux.The result of only finding 12 detections out of 51 AGN, with no evidence for redshift dependence, shows that the physical environments of the AGN and host galaxies must play an intimate role in the escape of LyC.Our sample of AGN is still small, and only covers a redshift range from z ∼ 2.3-4.9.However, we are beginning to ascertain the physical origins for LyC escape in AGN with this small sample.This analysis should be expanded upon, both with more HST WFC3/UVIS data and to lower redshift using data from higher-energy observatories.Expanding the sample to higher redshifts may be challenging given the increasing opacity of the IGM.Since AGN show essentially no evolution in their SED over cosmic time, the correlation of their physical parameters to their LyC emission could reveal the true impact of AGN on reionization.Their inferred LyC flux at the most relevant redshifts at z 6 could ideally be obtained simply by finding where these high-z AGN lie on some correlated curves of LyC flux versus a measurable physical parameter.
sample, the monochromatic 1700 Å luminosity of the SEDs in our sample without an AGN component mostly fall within the same range of * L L UV UV < 3 ( * L UV = -21 for z ; 3; Reddy & Steidel 2009

Figure 1 .
Figure 1.SED parameter histograms of the 51 AGN in our sample.Blue bars indicate AGN with a 4σ LyC S/N measurement, and red bars represent <4σ.A Kolmogorov-Smirnov (K-S) test of the two histograms cannot reject the null hypothesis in any histogram.The M AB is calculated from the best-fitting CIGALE SED at λ rest = 1450 Å (e.g., Kulkarni et al. 2019).A V is taken from the Av_ISM parameter of the attenuation component, SFR now is from sfr in sfh, Z is from metallicity in stellar,is the sum of m_gas_old, m_gas_young, m_star_old, and m_star_young in stellar, Γ is from gam in xray, L stars is the sum of lum_old and lum_young from stellar, L dustabs is from luminosity in dust, L disk is from disk_luminosity in agn, L torus is from torus_dust_luminosity in agn, L dustAGN is from total_dust_luminosity in agn, and L X-ray is the agn_Lx_total parameter in the xray component.

Figure 2 .
Figure2.CIGALE SED fitting results for the LyC-detected AGN in Table1.The observed photometry is shown as light blue error bars, and the LyC detection is indicated as a purple error bar.The fitted SED is shown as a black curve, and the blue curve, used for estimating the escape fraction, includes the same SED components as the black curve, excluding all absorption components.The stellar component is shown as the red curve, the AGN component is shown as the green curve, the X-ray emission is shown as the magenta curve, and the radio emission is shown as the orange curve.

Figure 3 .
Figure 3. 4 53 × 4 53 cutouts of our LyC-detected AGN showing their LyC image (LyC, left) and ∼1500 Å rest-frame UVC (right).The NED object name and corresponding filters are also shown.The image detected apertures are shown as blue ellipses.

Figure 4 .
Figure 4. LyC and UVC stacks of all AGN in the total sample.Blue ellipses show the SExtractor detection apertures from the UVC stack.The number of AGN included in each stack is indicated on the top-left corner of each individual panel.The left two columns show the full sample of AGN where LyC can be measured in either the filter indicated on the bottom right of the panel, or the total AGN ("All") sample in the bottom row.The middle and right pair of columns represent half of the AGN in their row, split into low-and high-redshift samples using the average redshift of the subsample as a separator.Top row: AGN with LyC measurable in F225W.Top-middle row: same but for F275W.Top-bottom row: same but for F336W.Bottom row: same, but for all AGN in the total sample.

Figure 5 .
Figure 5. LyC and UVC stacks of the AGN subsamples listed in Table 2. Top row: subsample of AGN that were detected by Chandra and/or Swift.Second row: AGN detected in an IR band (J, H, or K ).Third row: AGN detected in a MIR band (IRAC, WISE 3.4-12 μm, or AKARI S11).Fourth row: AGN detected in a FIR band (WISE 22 μm, PACS, IRS, MIPS, or AKARI L18W).Fifth row: AGN detected in a submillimeter band (SPIRE, SHARC2 350 μm, or LABOCA 870 μm).Sixth row: AGN detected in the microwave region (IRAM or AzTEC).Seventh row: AGN detected in the radio (ATCA or VLA).Bottom row: LyC nondetected AGN, indicated by LyC.The IR-detected AGN show the brightest flux among the "All" subsamples, with 39 included in the stack.The brightest subsample of all panels is the F336W IR-detected subsample.

Figure 6 .
Figure 6.LyC flux vs. various best-fitting SED parameters with strong correlations for our AGN sample.Light blue points indicate detected LyC while orange bars are 1σ upper limits to the LyC flux.The dark blue line is the Bayesian fit to the data with slope and intercept shown (m and b, respectively), treating upper limits as censored, and the red dashed line is the same fit, excluding the brightest point with the largest weight.Despite this exclusion, the general trends still survive.All correlation coefficients (R) shown reject the null hypothesis (p < 0.05).The strongest correlating parameter is the X-ray luminosity from 2 to 10 keV, indicating a strong X-ray to LyC emission connection for AGN.The SFH τ onset age parameter shows a strong correlation, which implies higher LyC emission after more recent star formation episodes.Several dust parameters (torus dust luminosity, polar dust luminosity, 6 μm luminosity, and A FUV ) show strong correlations with LyC flux, as well as 1.4 GHZ radio emission.Stellar metallicity shows a strong anticorrelation to LyC flux, which may signify a preference for higher LyC emission from metalpoor galaxies.
Table columns: (1) NED object ID; (2) AGN spectroscopic redshift; (3) LyC AB magnitude measured in the corresponding WFC3/UVIS filter; (4) S/N of the measured LyC flux; (5) the f esc of the AGN and host galaxy and its uncertainties.Values over 100% are included to show the limitations of the SED models.(6) The nonionizing UV AB magnitude measured from the ACS/WFC F606W filter.(This table is available in its entirety in machine-readable form.)