WorksPhysics and Astronomy
The Broad-Line Region Size-Luminosity Relation in AGN Constrained by the Viewing Angle with the FRADO Model
Abstract
We investigate the H broad-line region (BLR) size-luminosity (-) relation of active galactic nuclei (AGN) using a Bayesian viewing angle inference framework based on the Failed Radiatively Accelerated Dusty Outflow (FRADO) model. First, we update the FRADO model by introducing a phase factor to account for the anisotropic visibility of the illuminated cloud surface. Second, we construct a Bayesian framework to constrain the viewing angle by matching the FRADO-based to the measured H lags of 151 AGN from Wang & Woo (2024). We successfully reproduce the slope and intrinsic scatter of the observed - relation, finding that the inferred viewing angles are broadly consistent with the estimates from dynamical modeling of velocity-resolved reverberation data, albeit with substantial scatter. As the observed H time lag and line velocity depend on the viewing angle, the black hole mass () calibrated from our work is typically larger than the reverberation-mapping estimates using an average virial factor. Our results provide a physically motivated interpretation of the - relation and the systematic uncertainties in reverberation and single-epoch estimates.
Introduction
Supermassive black holes (SMBHs) at the centers of galaxies are known to play a significant role in galaxy evolution through feedback during the active galactic nucleus (AGN) phases [e.g. 19, 24, 22, 23]. Understanding AGN feedback and black hole physics requires reliable black hole mass () measurements. Reverberation mapping (RM) provides for type-1 AGN based on the time lag between continuum and broad emission-line variability [10, 43], constraining a characteristic size of the broad-line region ().
One of the key results from the RM studies is the empirical relation between and continuum luminosity at 5100Å (), which enables indirect estimates by inferring from using a single-epoch spectrum [e.g. 27, 58, 46, 52, 7]. For example, M. C. Bentz et al. [8] reported a tight – relation with a slope close to 0.5, which is consistent with simple photoionization models. This relation has been widely used for single-epoch estimates [e.g. 50, 45, 4, 48, 39, 47] (Q. Wu & Y. Shen 2022).
However, later studies showed that super-Eddington AGN tend to deviate from the – relation [e.g. 30, 60] (P. Du et al. 2016; P. Du et al. 2018; S. S. Li et al. 2021), although the physical reason is not yet clear (see discussion by J. H. Woo et al. [60]). These results have motivated calibrations accounting for the Eddington ratio effect [see 57] (P. Du et al. 2018; P. Du & J. M. Wang 2019; L. M. Yu et al. 2020). By uniformly analyzing H lags for a large sample of AGN, S. Wang & J. H. Woo [54] recently updated the – relation, reporting that the slope of the relation is with an intrinsic scatter of dex (see also J. H. Woo et al. [60]). J. H. Woo et al. [57] showed that three parameters, , , and gas velocity, define a tight fundamental plane, suggesting the fundamental plane relation is more suitable for single-epoch mass estimates than the two-parameter – relation.
Several theoretical models have been proposed to explain the origin of the BLR (e.g., [20, 18, 32, 12, 6, 5, 25 ) and the $R_{\mathrm{BLR}}$–$L_{5100}$ relation (see also {citep}][38, 32, 12, 5]. Among these models, the Failed Radiatively Accelerated Dusty Outflow (FRADO) model, first introduced by {citet}B. Czerny & K. Hryniewicz (2011), describes the BLR as a population of H$\beta$-emitting clouds lifted above the accretion disk by radiation pressure acting on dust, while gravity from the central black hole eventually causes the clouds to fall back after dust sublimation. The FRADO model predicts the inner and outer boundaries of the BLR, asymmetric emission-line profiles ({citet}M. H. Naddaf [33]), and the dependence of the virial factor on black hole mass ({citet}B. Czerny et al. [13]`).
The 2.5D FRADO model by M. H. Naddaf et al. [35] reproduces an – relation similar to that reported by M. C. Bentz et al. [8] based on the particle-based three-dimensional implementation (see also [36, 26]. However, the 2.5D FRADO model does not explicitly include anisotropic visibility of the illuminated cloud surface to the line-of-sight (LOS), although such anisotropy has been suggested by BLR dynamical modeling studies ([42, 41, 56]). Since the far-side clouds may contribute more than near-side clouds to the transfer function used to define the representative time lag, a phase-dependent weighting scheme is needed to properly calculate the H time lag.
The viewing angle is another key factor in determining the LOS-dependent time lag. In practice, however, the viewing angle remains difficult to determine observationally for individual AGN, while several recent studies have attempted to constrain it using dynamical modeling of velocity-resolved RM data (e.g., [40, 42, 56, 55]) or interferometric phase spectra (e.g., [51, 1, 2, 3]).
With these motivations, we investigate whether the FRADO model, combined with a viewing angle inference framework, can reproduce the observed – relation, by determining the representative time lag from the transfer function after correcting for the anisotropic visibility of each cloud. In our Bayesian framework, we constrain the viewing angle (), which determines the projection effects on the observed time lag, line velocity, and . In § 2, we present the update of the 2.5D FRADO model. The Bayesian framework for the viewing angle calibration is presented in § 3. In § 4, we present the Bayesian inference results, including the inferred viewing angle and AGN properties. Discussion is presented in § 5, followed by summary and conclusions in § 6.
Time Lag constrained by the FRADO Model
In this section, we describe the update of the FRADO model and the calculation of H time lag. First, we describe the implementation of the BLR in the FRADO model (§ 2.1). Second, we present the phase factor to include anisotropic visibility of BLR clouds (§ 2.2). Third, we describe the grid model used to calculate time lag over a continuous parameter space (§ 2.3).
Calculation of Transfer Function
The BLR described by the FRADO model is an ensemble of non-interacting clouds, whose time-independent particle distribution is used to calculate the representative H time lag (). The particle distribution is determined by the black hole mass and luminosity (or equivalently, Eddington ratio ), which set the gravitational potential of the central black hole and radiation pressure from the accretion disk, respectively. Once lifted from the disk surface, the clouds evolve under the combined effects of gravity and radiation pressure. The disk region contributing to the radiation pressure on each cloud is determined using the simplified shielding model introduced in M. H. Naddaf et al. [34] and M. H. Naddaf et al. [35]. When a cloud reaches the dust sublimation temperature, its dust is destroyed, radiation pressure ceases, and the cloud falls back toward the disk. Together with gravitational rotation, this vertical oscillatory motion produces the characteristic three-dimensional distribution of BLR clouds.
For a given BLR geometry specified by black hole mass and , the viewing angle () determines the LOS-dependent , which is calculated in three steps: computing the time lags of individual clouds, constructing the phase-weighted transfer function, and deriving , and extending the grid model to a continuous parameter space.
We start by calculating the time lag of individual clouds. The time lag of each cloud is defined as the light-travel path difference between the reprocessed lineemission and the continuum. Assuming that the continuum propagates directly from the central source to the observer, the time lag of each cloud is given by
where is the cloud position vector relative to the central black hole, is the LOS unit vector, and is the speed of light.
Then, we construct the transfer function by collecting the time lags of individual clouds.
Introduction of a Phase Factor to the Transfer Function
We introduce a phase factor into the transfer function to account for the anisotropic visibility of the illuminated side of each cloud. Following M. H. Naddaf et al. [36], we first assign each cloud an H emissivity determined by the incident photon flux. Then we multiply the emissivity by a dimensionless phase factor that accounts for the anisotropic visibility of the illuminated cloud surface. Because BLR clouds are expected to be optically thick to the central ionizing radiation, only the surface facing the central source is illuminated, and only part of this surface is visible to the observer (see Figure 1).

Figure 1. Illustration of the phase factor (Equation (2)). The position vectors of three arbitrary clouds are denoted with red arrows. For each cloud, the illuminated surface due to the central source is shown in white, whereas the right panel shows the reflected surfaces to the LOS in white.
Assuming parallel incident rays along the direction and spherical clouds, we define the phase factor as the fraction of the projected cloud surface visible to the observer that is illuminated by the central source as:
where is the normalized position vector of a cloud. Although FRADO clouds are point particles, we introduce a spherical shape to define the dimensionless phase
factor. The phase factor is unity when an observer sees the fully illuminated surface of a cloud, and zero when the illuminated surface is invisible along the LOS. As a result, clouds located on the far-side of the BLR from an observer have larger phase factors than those on the near-side, increasing the contribution of the far-side BLR to the transfer function (see also [42]) as illustrated in Figure 2. This effect of the phase factor becomes more pronounced with a larger , as the projected separation of the two sides increases and the median of the transfer function shifts toward longer lags.

Figure 2. Example transfer functions of an AGN with and . The total transfer function (black line) is calculated using the H emissivity with and without the phase factor (dashed and solid lines, respectively). The corresponding far-side and near-side transfer functions are shown in red and blue, respectively.
We define as the centroid of the transfer function region above 70% of the peak. This definition yields time lags broadly consistent with the Gaussian-smoothed far-side peak adopted by M. H. Naddaf et al. [36]. Although differs from the cross-correlation function (CCF) lags commonly measured above 80% of the CCF peak in RM studies, it provides a similar characteristic BLR response time.
We present as a function of viewing angle for various pairs of and in Figure 3, showing that generally increases with viewing angle. In the following analysis, we will use the viewing angle dependent as a representative time lag of the BLR.

Figure 3. Viewing angle dependence of for AGN with various and . is defined as the centroid above the 70% peak threshold (solid) or the far-side peak (dotted lines).
Extension to a Continuous Parameter Space
We extend the discrete FRADO lag grid to a continuous parameter space to evaluate for arbitrary input parameters of mass and Eddington ratio. The cloud distributions are computed on a two-dimensional grid spanning –10 and –0, with evaluated over –, appropriate for type 1 AGN. Rather than interpolating the full four-dimensional hypersurface , we evaluate the grid at 0.5° intervals in and perform interpolation or extrapolation in the space at each fixed . In the low-luminosity regime, some combinations of and fall below the incident ionizing photon-flux threshold adopted in the 2.5D FRADO model (M. H. Naddaf et al. [36]), . Consequently, the populated grid points form a triangular region in the plane rather than a complete rectangle.
For a set of input parameters with sufficient surrounding grid support, we estimate using local interpolation. We first identify the neighboring grid points along the and directions and compute local linear estimates of . The two local estimates are then combined using an inverse-variance-weighted average to obtain the interpolated lag and its uncertainty.
If the input lies outside the populated region or lacks sufficient local grid support, we use plane-based extrapolation. In this case, we fit a plane to all populated grid points in the space and evaluate it at the input values of and . This treatment is justified because the FRADO time-lag surface at fixed is nearly planar in the space, with a typical rms scatter of dex or smaller. This hybrid interpolation and extrapolation scheme provides throughout the continuous parameter space, including regions without sufficient local grid support.
Bayesian Framework for constraining the viewing angle
In this section, we present a Bayesian framework for inferring the viewing angle by matching to the observed H lag (), for a given AGN with the observed RM mass () and Eddington ratio (). We apply this framework to a sample of RM AGN from S. Wang & J. H. Woo [54] in § 4.
For each trial value of for an AGN with fixed and , we obtain and as input parameters for the FRADO model and evaluate the likelihood of given . Our Bayesian framework is summarized in Figure 4. Note that and represent the observed mass and Eddington ratio from the RM studies, while the unsubscripted quantities, and , denote the corresponding intrinsic parameters using the inferred .

Figure 4. Bayesian framework used in this work. Given the observed quantities (, , and ), the framework uses the FRADO model and the viewing angle calibration of and to infer . Each trial parameter vector determines the relative probabilities of the trial viewing angles for each AGN. At each trial , the corresponding and are calibrated and used to compute . The likelihood is evaluated by comparing with and marginalizing over . Markov chain Monte Carlo (MCMC) is then used to sample . The posterior samples are propagated to infer the distributions of , , , and .
-Dependence of the Observed AGN Properties
The observed RM is defined as
where is the H lag, is the H line width, and is an average virial factor. Since changes as a function of the viewing angle to the LOS, and corresponds to the LOS velocity, depends on .
The virial factor of individual AGN also depends on the viewing angle since it reflects the BLR geometry and kinematics projected along the LOS. In fact, the viewing angle is likely to be a dominant source of variation in the virial factor [see also 41, 21, 56, 53]. However, an average virial factor is often adopted in practice due to the difficulty of constraining a virial factor for individual AGN [e.g. 59, 54]. This average virial factor may correspond to a representative viewing angle of the adopted AGN sample, although the virial factor also depends on other BLR properties, such as the opening angle and radial extent of the BLR. The viewing angle dependence is also clear in the observed Eddington ratio defined as
The disk luminosity observed at 5100 Å () is often calculated from the measured flux assuming isotropic emission, although the disk emission is anisotropic with respect to the LOS. By adopting a simplified thin Lambertian disk prescription, we find that the observed continuum luminosity scales with due to the projected emitting area. However, the constant bolometric correction factor of 10 commonly applied to is calibrated for an AGN population and therefore implicitly corresponds to . Thus, the bolometric luminosity and Eddington ratio of each AGN need to be calibrated using its individual viewing angle.
Constraining Representative Viewing Angle
We quantify the viewing angle effect on the observed parameters by parameterizing the difference between the inferred and a representative viewing angle under the assumption that only the viewing angle determines the virial factor. We determine the representative viewing angle corresponding to the average virial factor in Equation (3). While is typically adopted from the calibration of the –stellar velocity dispersion relation by J. H. Woo et al. [59], individual estimates of and are not available for the AGN used in the calibration. Instead, we use 37 velocity-resolved RM AGN from L. Villafaña et al. [53], for which is calculated as the ratio of the inferred from dynamical modeling to the observed virial product, and is inferred from the same modeling. For the same sample, was constrained by S. Wang et al. [55]: for the rms line dispersion and for the mean line dispersion, which are almost equivalent to those constrained by J. H. Woo et al. [59]. Thus, adopting this value rather than the value from J. H. Woo et al. [59] does not significantly affect our results. For each AGN in the sample of L. Villafaña et al. [53], we generate paired Monte Carlo samples of and . We then assign each Monte Carlo draw a Gaussian weight according to the deviation of its from the adopted , such that draws closer to receive larger weights. The weighted draws from all AGN are then combined into a single distribution, from which we adopt the median as the representative viewing angle.
Modeling -Dependence
First, we model the viewing angle dependence of the time lag. Since the far-side BLR dominates the transfer function (§ 2.2), we assume the lag increases with the viewing angle as for a thin BLR disk. By accounting for the shape of the transfer function (see Figure 2), we generalize this dependence as , where the dimensionless factor captures variations in the transfer function shape arising from the extended BLR geometry in the FRADO model. Then, the dependence of is expressed as
Second, we adopt the prescription of S. Collin et al. [11] for the line velocity by combining rotational and turbulent components. If the rotation is dominant, the projected LOS velocity scales approximately as and the corresponding mass scales as . In addition, we include a turbulent velocity component related to the BLR thickness (or opening angle). Following S. Collin et al. [11], we parameterize the ratio of the turbulent velocity to the local Keplerian rotational velocity with the BLR aspect ratio. We describe the aspect ratio as , where is the BLR half-opening angle, and introduce as a free parameter that determines the strength of the turbulent contribution to . Therefore, scales as . Because the calibrated mass is inversely proportional to this factor, a nonzero prevents it from diverging when approaches zero. We constrain through Bayesian inference and treat it as a global parameter common to all AGN in our H RM sample. We discuss the implications of this parameter further in Section 5.3.
Third, combining the dependences of and , we describe the -dependence on mass as
Fourth, under a simplified thin Lambertian disk prescription, we assume that the observed continuum luminosity scales with , which represents the projected emitting area. Therefore, the calibration of is given as
Finally, by combining Equations 6 and 7, we obtain the Eddington ratio prescription as
We parameterize the opening angle and shape factor as functions of the intrinsic AGN luminosity derived from the FRADO model as described in Section 3.4. Note that is defined as the altitude range that includes 95% of the H emission.
Conditional Weights for Marginalization
We construct a conditional weight used to marginalize the likelihood over using Equation 8, which relates to through the calibration. While is not expected to correlate with or , the observed values, and may depend on as described in the previous section (see Equations (6) and (8)).
We build an intrinsic distribution of as an asymmetric split-normal distribution centered at , with left and right scatters and , respectively. For a fixed , (8) shows that the viewing angle calibration naturally produces an anti-correlation between and , because the ratio increases monotonically with for the positive values of and relevant to our model. Changing shifts this relation horizontally in the – plane. The parameters , , and therefore define the family of – relations allowed for the population where the central locus is set by .
The resulting parameter vector is
We assign uniform priors over the ranges listed in Table 1.
| Parameter | Prior Range |
Table 1. Uniform prior ranges adopted for the global parameters in the Bayesian inference.
For each sampled , we assign each AGN an object-specific conditional weight over the allowed grid. The likelihood is then marginalized over this grid. For the th AGN with an observed Eddington ratio , we define the intrinsic Eddington ratio at the th viewing angle grid point using (8):
where
Here, represents the combined rotational and turbulent contributions to at , while denotes the corresponding value at . Moreover, using the FRADO cloud distributions, we calculate and over the full grid and obtain linear fits as functions of the luminosity proxy :
For each evaluated at , we assign the th AGN a (unnormalized) conditional weight . This is defined as
The factor represents the isotropic viewing angle prior expected for randomly oriented systems in three-dimensional space [28]. Meanwhile, the Jacobian term converts the population weight defined in space into the corresponding weight in space. We normalize the weights as
so that
Here, denotes the grid spacing in . For consistency with the FRADO model grid, we restrict the viewing angle to .
Marginalized Likelihood and Posterior
Using the conditional viewing angle weights defined in § 3.4, we construct the -marginalized likelihood of and combine it with the priors on to obtain the posterior. For each AGN and viewing angle grid point , we apply the calibration described in § 3.3 to obtain and from and . These quantities, together with , are then used to evaluate and its uncertainty. For each AGN and viewing angle grid point , we define the likelihood by modeling as a Gaussian variable centered on ,
where
Here, and denote the uncertainties of and in log-space, respectively. The likelihood is then marginalized over the grid,
where is the normalized conditional weight for the th AGN at . The conditional likelihood is given by
where
The total log-likelihood of the sample is therefore
Combining the likelihood with the prior on yields the posterior distribution
In addition, we account for the measurement uncertainties in and by drawing 25 Gaussian realizations for each AGN and marginalizing the likelihood over these realizations. For each realization, we recompute the viewing angle calibration, the conditional viewing angle weights, and . The final likelihood is obtained by averaging the -marginalized likelihood over the 25 realizations.
We sample the posterior using the affine-invariant ensemble sampler implemented in the emcee package. The resulting posterior samples constrain the four model parameters and are used to infer of each AGN. We then propagate these posterior samples to obtain the corresponding posterior distributions of , , and posterior-predictive model time lag, .
Results
In this section, we present the main results obtained by applying the Bayesian framework to the H RM sample of S. Wang & J. H. Woo [54]. First, we describe the posterior constraints on the model parameters and the resulting distributions of , , and for the H RM sample (§ 4.1). Second, we investigate the sample selection effect on the inferred – and – relations (§ 4.2). Third, we compare the inferred with the observed H lags and evaluate the reproduced – relation against observations (§ 4.3).
Posterior Constraints
Applying the Bayesian framework to the H RM sample, we obtain the posterior constraints on the model parameters and the viewing angles as shown in Figure 5a. From the joint posterior distributions of the four parameters, we obtain the 1D parameter constraints as , , , and with 1- uncertainties. Note that these posterior constraints depend on the input sample. Thus, these parameters are relevant for the H RM sample. By defining the maximum a posteriori (MAP) parameter vector, , that maximizes the posterior probability in the full four-dimensional parameter space, we obtain , , , and . We adopt as the representative parameter vector for the fixed- analyses in Section 4.3, which assess the sensitivity of the results to uncertainty in .

Figure 5. Left: Joint posterior distributions of the four model parameters. Regions enclosing 68%, 95%, and 99.7% are denoted as contours, while the marginalized 1D posterior distributions are shown in the diagonal panels. The 16th, 50th, and 84th percentiles of the 1D distributions are indicated by the blue dashed vertical lines. Right: 2D posterior probability distribution of the observed Eddington ratio () and the inferred viewing angle (), showing the conditional distribution of at a given . Regions enclosing 16%, 50%, and 84% are denoted as contours.
Using the posterior constraints on , we infer the distribution for each AGN in the H RM sample. We present the resulting family of – relations with a probability density map in Figure 5b. Despite the broad dispersion in at fixed , AGN with higher preferentially have smaller viewing angles.
Then, we construct the joint posterior distributions of the inferred and the calibrated and by drawing 200 samples of from the posterior and generating 200 realizations for each sampled vector. For each realization, we obtain the corresponding and through the viewing angle calibration (see § 3.3). We present the resulting distributions in Figure 6. Note that and are expected to be independent of . However, the posterior contours exhibit residual trends with , which we attribute to the incompleteness of the H RM sample (see § 4.2).

Figure 6. Joint posterior distributions of the viewing angle and the intrinsic (left) and Eddington ratio (right). Regions enclosing 16%, 50%, and 90% are denoted as contours. The centroid trends are indicated by the white lines.
For the H RM sample, we obtain a marginalized distribution with a mean of , which is smaller than the representative viewing angle corresponding to (see Figure 7). It is also substantially smaller than the isotropic viewing angle prior adopted in our inference, . As the BLR is visible at relatively low viewing angles, this result is qualitatively consistent with the typical expected for type 1 AGN. We discuss this result further in Section 5.1.

Figure 7. Marginalized posterior probability distribution of (black line), constructed using the sampling configuration of Figure 6. The 16th–84th percentile range is indicated by the shaded gray region, while the median is marked by the vertical dotted line. For comparison, the distribution expected for an isotropic orientation in three-dimensional space, , is shown by the red line.
Selection Effects in the – and – Relations
Since is expected to be independent of the intrinsic AGN properties, and , the weak correlations shown in Figure 6 require further investigation. We therefore test whether these trends arise from the limited dynamic range of the H RM sample. To test this interpretation, we generate a mock population of AGN such that follows the same Gaussian distribution as the H RM sample, while is uniformly distributed.
We compare the and distributions of the mock AGN sample and the H RM sample in Figure 8a. The H RM sample lacks low-Eddington-ratio AGN at high luminosities and high-Eddington-ratio AGN at low luminosities, whereas the mock sample provides denser coverage of both regimes. With this more complete coverage in the distribution, we now confirm no dependence of on or , as shown in Figure 8b and Figure 8c. The residual correlations in Figure 6 are likely caused by the incomplete coverage of the H RM sample in the plane. Note that the mock sample also shows relatively small since we use the posterior samples of obtained from the H RM sample.

Figure 8. Left: Joint distribution of and for the H RM sample, color-coded by the median inferred . A mock population of 5000 AGN is shown as green points for comparison. Middle: Joint posterior distribution of and the inferred . Right: Joint posterior distribution of and the inferred . In the middle and right panels, the 16%, 50%, and 90% posterior-probability regions are shown as contours, and the centroid trends are indicated by the overplotted lines.
– Relation
Using the sampling configuration described in § 4.1, we draw Monte Carlo samples of for each AGN in the H RM sample. We apply the viewing angle calibration to each sample to obtain the corresponding and , and then use to calculate with the FRADO model (see § 2). The resulting samples define the posterior-predictive distribution of for each AGN.
As a consistency check, we compare with (Figure 9a), finding consistency between them with a negligible offset of 0.037 dex and an rms scatter of 0.22 dex, indicating that we properly reproduced the observed time lag. To test the effect of the uncertainty in , we fix the parameter vector at and generate 1000 viewing angle realizations. We find that the is consistent with with an offset of 0.042 dex and an rms scatter of 0.21 dex, indicating that uncertainty in has little effect on the result. We also repeat the analysis without the measurement uncertainties in and , finding a comparable result.

Figure 9. Left: Comparison of with . The error bar of indicates the 68% credible intervals of the posterior-predictive distributions. The median inferred of each AGN is color-coded. Right: The reproduced ‘observed’ – relation at fixed with the best-fit (red line) and the 1- uncertainty range obtained from 1000 draws (red shaded region). The values of and are indicated by color and point size, respectively. The best-fit relation for the H RM AGN from S. Wang & J. H. Woo [54] (black line) and the M. C. Bentz et al. [8] relation (green dash-dotted line) are also presented for comparison.
Then, we test whether our model can reproduce the observed – relation. We first fix the parameter vector at and generate 1000 Monte Carlo realizations of for the H RM sample. For each realization, we measure the slope and intrinsic scatter and construct their probability distributions. We obtain an – slope of and an intrinsic scatter of dex using the 1000 realizations, where the uncertainties represent the 1- ranges, finding that the model properly reproduces the observed – relation reported by S. Wang & J. H. Woo [54]. We repeat the analysis by drawing from its full posterior distribution using the same sampling configuration from § 4.1. We obtain an – slope of and an intrinsic scatter of dex, which are in better agreement with those measured by S. Wang & J. H. Woo [54]. In Figure 9b, we illustrate the observed – relation generated from a single draw at fixed .
AGN with higher tend to lie below the mean – relation, qualitatively consistent with the trend with observed Eddington ratio. However, the calibrated distribution of the H RM sample is systematically shifted toward lower values relative to the distribution of S. Wang & J. H. Woo [54]. We discuss this further in § 5.4.
We successfully reproduce the shallower slope of the observed R-L relation and the residual trend toward smaller for AGN with higher [54]. Note that this shallower slope is mainly caused by the viewing angle effect of the H RM sample, which has a specific Eddington ratio distribution. Although there is a self-correlation between and since , the offset from the intrinsic R-L relation is not expected to directly correlate with the intrinsic (see Figure 19 of J. H. Woo et al. [60]).
Discussion
In this section, we discuss the inferred viewing angle distribution of the H RM sample (§ 5.1), the dependence of our results on the adopted representative viewing angle (§ 5.2), the treatment of the BLR opening angle (§ 5.3), and the differences between the observed AGN properties and their calibrated counterparts (§ 5.4).
Viewing Angles of the H RM Sample
Our Bayesian inference applied to the H RM sample favors viewing angles smaller than the representative viewing angle , which was derived from the virial factors and viewing angles of individual AGN (see § 3.2). As shown in Figure 7, the viewing angle distribution of the sample is skewed toward low viewing angles compared to the isotropic expectation. This trend is also evident in comparison with constrained by dynamical modeling. Overall, the inferred viewing angles of the H RM sample are lower than those typically expected for type 1 AGN.
Several studies have suggested that narrow-line Seyfert 1 galaxies (NLS1s), which typically show high Eddington ratios, may preferentially be viewed at low viewing angles [31, 37, 9, 14, 44] (T. Sarrabata et al. 2018). Although the majority of AGN in the H RM sample have H FWHM values greater than 2000 km s and therefore are not classified as NLS1s, the proposed face-on interpretation of high-Eddington-ratio NLS1s is qualitatively analogous to the anticorrelation between and (Figure 5b).
Compared to the viewing angles from dynamical modeling reported by L. Villafaña et al. [53], the inferred is lower by for the 31 overlapping AGN. The two estimates are broadly consistent within their uncertainties. However, several objects have inferred viewing angles substantially smaller than the corresponding dynamical modeling estimates.
These differences arise because the inference adjusts to bring into agreement with . When , the viewing angle calibration increases and decreases relative to and , respectively (see § 3.3). Because depends strongly on (see Figure 3), the increase in produces a longer . Thus, objects whose initial based on is much shorter than favor smaller than . Because this behavior depends on the adopted viewing angle calibration, the current calibration model may require further refinement. In particular, the viewing angle dependence of the line velocity should be modeled more realistically using the FRADO cloud distribution. We defer further calibration to future work.
The discrepancy between the inferred and the dynamical modeling estimates may also arise from the degeneracy between and other BLR parameters, particularly the opening angle . In our Bayesian framework, a large weakens the viewing angle dependence of the model time lag as well as the line velocity (see § 3.3), thereby broadening the probability distribution of . The preference for small viewing angles is presumably due to the relatively small predicted by the FRADO model: for the H RM sample, rarely exceeds 30°, even for the most luminous AGN (see § 5.3). The degeneracy among the inferred , , and has been reported in BLR dynamical modeling (see also A. Pancoast et al. [41]; C. J. Grier et al. [21]). In dynamical modeling, this degeneracy produces a positive correlation between and in the joint posterior, allowing both parameters to shift toward either smaller or larger values. Consistent with this interpretation, AGN for which the inferred viewing angles are substantially smaller than the dynamical modeling estimates gener-
ally have large opening angles inferred from dynamical modeling.
In dynamical modeling, is treated as a free parameter. In contrast, we calculate from the FRADO cloud distribution specified by and (see § 3.4). This prescription yields a narrower range of with typically smaller values than those inferred from dynamical modeling. Thus, it is necessary to further study the degeneracy between and to confirm the relatively low viewing angles inferred for the H RM sample. We further discuss the opening angle implementation in § 5.3. Although observationally challenging, spectropolarimetric constraints on the viewing angle, once available for individual AGN in a larger RM sample, will provide an important independent test of the inferred viewing angle distribution.
Representative Viewing Angle
The inferred of individual AGN slightly depends on the choice of , for which we adopt based on the 37 AGN with dynamical modeling results (see § 3.2). As a consistency check, we derive an alternative assuming an isotropic distribution (see also [28]). For randomly oriented AGN, the viewing angle probability density, , follows , where is the probability density per unit viewing angle. Assuming that type 1 AGN are observed within , the mean of this distribution is . Using this distribution, we define the average log virial factor as
and determine from
where the viewing angle dependence of the virial factor is expressed as
This expression follows from the mass calibration in Equation (6). From Equations (12) and (13), we obtain approximate ranges of – and – over the typical range of – in the H RM sample. Furthermore, our Bayesian inference constrains to – across –. For the adopted ranges of , , and , the corresponding representative viewing angle varies only from to , indicating weak sensitivity to these inputs. We therefore adopt for the alternative configuration.
Using this alternative , we repeat the Bayesian inference and compare the posterior-predictive with . We find consistency between and , with an offset of dex and an intrinsic scatter of dex.
Next, we compare the inferred for with the dynamical modeling estimates, as in Section 5.1. We obtain an offset of , showing better agreement than for . As future H RM samples become larger and less affected by selection effects, the representative viewing angle adopted in our Bayesian framework may approach .
Lastly, we revisit the inferred viewing angle distribution of the H RM sample for the alternative choice of . The marginalized distribution shifts toward larger viewing angles. Nonetheless, the resulting distribution still shows a substantially larger fraction at small viewing angles than the isotropic distribution.
Opening Angle
In our Bayesian framework, modulates the viewing angle dependence of through the turbulent contribution , where the free parameter controls its amplitude (see § 3.3). The inferred is ( for ), implying that the opening angle contributes to the turbulent-to-rotational velocity ratio with a coefficient of –. Over the typical ranges of and for the H RM sample, the rotational component remains dominant, while the turbulent component associated with is subdominant but non-negligible, weakening the viewing angle dependence of the line velocity. This result is qualitatively consistent with C. J. Grier et al. [21], who reported that the viewing angle has a greater effect on the line profile shape than the opening angle.
As discussed in Section 5.1, the degeneracy between and is an important source of uncertainty in the inferred . In this work, we calculate from the angular extent of the FRADO cloud distribution (see § 3.3). The resulting opening angles rarely exceed , even for the most luminous AGN in the H RM sample. In contrast, the opening angles inferred from dynamical modeling span approximately – and are typically larger than those predicted by the FRADO model. In both our inference and BLR dynamical modeling, the degeneracy between and produces a positive correlation, such that a larger is associated with a larger (see Section 5.1). The limited range of in our model may therefore contribute both to the overall preference for small viewing angles in the H RM sample and to the substantially smaller values inferred for some AGN relative to the dynamical modeling estimates. A more realistic treatment of cloud irradiation in future versions of the FRADO model may shift the predicted distribution toward larger values and broaden its range.
Systematic Uncertainties of the Observed AGN Properties
The observed RM () and Eddington ratio () deviate from the calibrated mass () and Eddington ratio () owing to viewing angle effects. By comparing the observed and intrinsic parameters, we discuss the systematic uncertainties of the observed mass and Eddington ratio. Using the same sampling configuration as in § 4.1, we construct joint probability density maps of (, ) and (, ) from paired realizations per AGN. To account for measurement uncertainties, we generate 25 Monte Carlo draws of and , select one random draw for each realization, and apply the viewing angle calibration using the sampled to obtain and .
As presented in Figure 10, the observed is smaller than by dex. In contrast, is larger than by dex. The upward mass shift suggests that the RM of the H RM sample may be systematically underestimated, implying that the fraction of super-Eddington AGN may be lower than that estimated using the RM . In fact, only of the calibrated draws exceed unity, whereas 13 of the 151 AGN () have .

Figure 10. Left: Comparison between the RM () and the intrinsic mass (). Right: Comparison between the observed Eddington ratio () and the intrinsic . White contours denote the 16%, 50%, and 90% probability regions and the centroids of the distributions are indicated by crosses.
The systematically larger intrinsic appears to be in tension with S. Wang et al. [55], who reported that dynamical modeling masses are consistent on average with the RM . By comparing the dynamical-model-based reported by L. Villafaña et al. [53] with the inferred intrinsic for the 31 overlapping AGN, we find an offset of dex. The trend toward larger mass in our Bayesian framework relative to the dynamical modeling estimates is more prominent for high-mass AGN. Since the mass discrepancy is driven by the preference for small inferred viewing angles in the H RM sample, the degeneracy between and may provide an underlying explanation (see § 5.1). In particular, the limited range of predicted by our model (see also § 5.3) may restrict the corresponding range of , leading to smaller inferred viewing angles and hence higher calibrated masses than those obtained from dynamical modeling.
Summary and Conclusions
In this work, we present the updated FRADO model and the Bayesian framework to investigate the viewing angle dependence of the – relation. We summarize our main results as follows:
We updated the FRADO model by including a phase factor for anisotropic visibility of the BLR cloud surface to the line-of-sight. We find that the far-side BLR has a larger weight than the near-side BLR in the transfer function, increasing the viewing angle dependence of the BLR time lag predicted by the model.
Using a Bayesian framework, we constrained the viewing angle of each AGN in the H RM sample,
finding that the inferred viewing angles are relatively small and broadly consistent with the viewing angle estimates based on dynamical modeling of velocity-resolved lags [53]. Several objects showed significant offsets, presumably due to the large opening angle inferred from dynamical modeling, reflecting the degeneracy between the viewing angle and the BLR opening angle.
The updated FRADO combined with a Bayesian framework reproduces the observed – relation. As the viewing angle affects the H time lag and the anisotropic disk luminosity, the slope and scatter of the observed – relation can be reproduced with the inferred viewing angles for the H RM sample.
Since the RM is determined using an average virial factor corresponding to the representative viewing angle of the sample, the FRADO-based is larger than the RM-based by dex because the viewing angles inferred for individual AGN are smaller than the representative viewing angle. This systematic mass offset requires further investigation, particularly regarding the limited range of opening angles predicted by the FRADO model.
Overall, the updated FRADO model successfully reproduces the observed – relation. The dependence of , , and on viewing angle should be investigated further using both observational and theoretical approaches.
Acknowledgments
This work has been supported by the Basic Science Research Program through the National Research Foundation of the Korean Government (RS-2026-25487266).
References
- [1]Amorim, A., Bauböck, M., Brandner, W., et al. 2020, A&A, 643, A154, doi: 10.1051/0004-6361/202039067DOI
- [2]Amorim, A., Bauböck, M., Brandner, W., et al. 2021, A&A, 648, A117, doi: 10.1051/0004-6361/202040061DOI
- [3]Amorim, A., Bourdarot, G., Brandner, W., et al. 2024, A&A, 684, A167, doi: 10.1051/0004-6361/202348167DOI
- [4]Baron, D., & Ménard, B. 2019, MNRAS, 487, 3404, doi: 10.1093/mnras/stz1546DOI
- [5]Baskin, A., & Laor, A. 2018, Monthly Notices of the Royal Astronomical Society, 474, 1970, doi: 10.1093/MNRAS/STX2850DOI
- [6]Begelman, M. C., & Silk, J. 2017, Monthly Notices of the Royal Astronomical Society, 464, 2311, doi: 10.1093/MNRAS/STW2533DOI
- [7]Bentz, M. C., Walsh, J. L., Barth, A. J., et al. 2009, ApJ, 705, 199, doi: 10.1088/0004-637X/705/1/199DOI
- [8]Bentz, M. C., Denney, K. D., Grier, C. J., et al. 2013, ApJ, 767, 149, doi: 10.1088/0004-637X/767/2/149DOI
- [9]Bian, W., & Zhao, Y. 2004, MNRAS, 352, 823, doi: 10.1111/j.1365-2966.2004.07967.xDOI
- [10]Blandford, R. D., & McKee, C. F. 1982, ApJ, 255, 419, doi: 10.1086/159843DOI
- [11]Collin, S., Kawaguchi, T., Peterson, B. M., & Vestergaard, M. 2006, A&A, 456, 75, doi: 10.1051/0004-6361:20064878DOI
- [12]Czerny, B., & Hryniewicz, K. 2011, A&A, 525, L8, doi: 10.1051/0004-6361/201016025DOI
- [13]Czerny, B., Li, Y.-R., Hryniewicz, K., et al. 2017, ApJ, 846, 154, doi: 10.3847/1538-4357/AA8810DOI
- [14]Decarli, R., Dotti, M., Fontana, M., & Haardt, F. 2008, MNRAS, 386, L15, doi: 10.1111/j.1745-3933.2008.00451.xDOI
- [18]Emmering, R. T., Blandford, R. D., Shlosman, I., et al. 1992, ApJ, 385, 460, doi: 10.1086/170955DOI
- [19]Fabian, A. C. 2012, ARA&A, 50, 455, doi: 10.1146/ANNUREV-ASTRO-081811-125521DOI
- [20]Ferland, G. J., & Persson, S. E. 1989, The Astrophysical Journal, 347, 656
- [21]Grier, C. J., Pancoast, A., Barth, A. J., et al. 2017, ApJ, 849, 146, doi: 10.3847/1538-4357/aa901bDOI
- [22]Hardcastle, M. J., & Croston, J. H. 2020, New Astronomy Reviews, 88, 101539, doi: 10.1016/j.newar.2020.101539DOI
- [23]Harrison, C. M., & Almeida, C. R. 2024, Galaxies, 12, 17, doi: 10.3390/galaxies12020017DOI
- [24]Heckman, T. M., & Best, P. N. 2014, ARA&A, 52, 589, doi: 10.1146/annurev-astro-081913-035722DOI
- [25]Hopkins, P. F. 2025, Open Journal of Astrophysics, 8, 56, doi: 10.33232/001c.137969DOI
- [26]Jaiswal, V. K., Mandal, A. K., Prince, R., et al. 2025, A&A, 702, A92, doi: 10.1051/0004-6361/202452497DOI
- [27]Kaspi, S., Smith, P. S., Netzer, H., et al. 2000, ApJ, 533, 631, doi: 10.1086/308704DOI
- [28]Lawrence, A., & Elvis, M. 2010, ApJ, 714, 561, doi: 10.1088/0004-637X/714/1/561DOI
- [30]Martínez-Aldama, M. L., Zajaček, M., Czerny, B., & Panda, S. 2020, ApJ, 903, 86, doi: 10.3847/1538-4357/abb6f8
- [31]Murayama, T., Taniguchi, Y., & Nagao, T. 2001, in Astronomical Society of the Pacific Conference Series, Vol. 222, Narrow-Line Seyfert 1 Galaxies and Their Place in the Universe, 375
- [32]Murray, N., Chiang, J., Grossman, S. A., & Voit, G. M. 1995, ApJ, 451, 498, doi: 10.1086/176238DOI
- [33]Naddaf, M. H. 2024, Universe, 11, 8, doi: 10.3390/universe11010008DOI
- [34]Naddaf, M. H., Czerny, B., & Szczerba, R. 2020, Front. Astron. Space Sci., 7, 15, doi: 10.3389/fspas.2020.00015DOI
- [35]Naddaf, M. H., Czerny, B., & Szczerba, R. 2021, ApJ, 920, 30, doi: 10.3847/1538-4357/ac139dDOI
- [36]Naddaf, M. H., Martinez-Aldama, M. L., Hutsemekers, D., Savic, D., & Czerny, B. 2025, A&A, 702, A46, doi: 10.1051/0004-6361/202555767DOI
- [37]Nagao, T., Taniguchi, Y., & Murayama, T. 2001, in Astronomical Society of the Pacific Conference Series, Vol. 222, Narrow-Line Seyfert 1 Galaxies and Their Place in the Universe, 379
- [38]Netzer, H., & Laor, A. 1993, ApJL, 404, L51, doi: 10.1086/186741
- [39]Pan, Z., Jiang, L., Guo, W.-J., et al. 2025, ApJ, 987, 48, doi: 10.3847/1538-4357/add7ddDOI
- [40]Pancoast, A., Brewer, B. J., & Treu, T. 2011, ApJ, 730, 139, doi: 10.1088/0004-637X/730/2/139DOI
- [41]Pancoast, A., Brewer, B. J., Treu, T., et al. 2014, Monthly Notices of the Royal Astronomical Society, 445, 3073, doi: 10.1093/MNRAS/STU1419DOI
- [42]Pancoast, A., Brewer, B. J., Treu, T., et al. 2012, ApJ, 754, 49, doi: 10.1088/0004-637X/754/1/49DOI
- [43]Peterson, B. M. 1993, PASP, 105, 247, doi: 10.1086/133140DOI
- [44]Peterson, B. M. 2011, https://arxiv.org/pdf/1109.4181
- [45]Peterson, B. M. 2014, Space Sci. Rev., 183, 253, doi: 10.1007/s11214-013-9987-4DOI
- [46]Peterson, B. M., Ferrarese, L., Gilbert, K. M., et al. 2004, ApJ, 613, 682, doi: 10.1086/423269DOI
- [47]Pucha, R., Juneau, S., Mezcua, M., et al. 2026, A New Record Census of Dwarf AGN and a Bimodal M_BH–M_* Scaling Relation with DESI DR1, https://arxiv.org/abs/2606.02699
- [48]Rakshit, S., Stalin, C. S., & Kotilainen, J. 2020, ApJS, 249, 17, doi: 10.3847/1538-4357/ab99c5
- [50]Shen, Y., Richards, G. T., Strauss, M. A., et al. 2011, ApJS, 194, 45, doi: 10.1088/0067-0049/194/2/45DOI
- [51]Sturm, E., Dexter, J., Pfuhl, O., et al. 2018, Nature, 563, 657, doi: 10.1038/s41586-018-0731-9DOI
- [52]Vestergaard, M., & Peterson, B. M. 2006, ApJ, 641, 689, doi: 10.1086/500572DOI
- [53]Villafaña, L., Treu, T., Wang, S., et al. 2026, ApJ, 1004, 151, doi: 10.3847/1538-4357/ae69cfDOI
- [54]Wang, S., & Woo, J. H. 2024, ApJS, 275, 13, doi: 10.3847/1538-4365/AD74F2DOI
- [55]Wang, S., Woo, J. H., Villafaña, L., Treu, T., & Gallo, E. 2026, ApJ, 1000, 180, doi: 10.3847/1538-4357/AE4699DOI
- [56]Williams, P. R., Pancoast, A., Treu, T., et al. 2018, ApJ, 866, 75, doi: 10.3847/1538-4357/AE0866
- [57]Woo, J. H., Kim, J. C., Cho, H., & Wang, S. 2026, ApJ, 1003, 123, doi: 10.3847/1538-4357/ae6109DOI
- [58]Woo, J. H., & Urry, C. M. 2002, ApJL, 581, L5
- [59]Woo, J. H., Yoon, Y., Park, S., Park, D., & Kim, S. C. 2015, The Astrophysical Journal, 801, 38, doi: 10.1088/0004-637X/801/1/38DOI
- [60]Woo, J. H., Wang, S., Rakshit, S., et al. 2024, ApJ, 962, 67, doi: 10.3847/1538-4357/AD132FDOI